> ## Content Index
> Fetch the complete content index at: https://theneedlenews.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Erin Reed Has Been Hiding AI-Generated Text in Her Reporting For Years
- URL: https://theneedlenews.com/2026/09/erin-reeds-ai-journalism-career/
- Published: 2026-09-30T13:00:46.000Z
- Updated: 2026-09-30T16:30:29.000Z
- Description: Much of what she has written for Erin in the Morning since Spring 2023 has been written and/or influenced by AI. The resulting articles are full of unacknowledged and unretracted errors.
- Author: Jane Migliara Brigham
- Tags: Investigations, #DontNormalizeQuotes

*In the interest of full disclosure, we want to address our personal connection to this story.*

*First, Erin Reed’s outlet, Erin in the Morning, is a competitor of The Needle. Second, one of the reporters on this story,* [*Jane Migliara Brigham*](https://theneedlenews.com/author/jane/)*, was previously interviewed by Erin Reed for a reporting position at Erin in the Morning. Reed ultimately hired S. Baum, who is mentioned in this story.* 

*We accept that some readers may feel that this suggests bias. Nonetheless, we stand by this investigation, and we believe our findings are too important not to release.*

## **Summary**

Erin Reed, owner of the news outlet *Erin in the Morning*, has been hiding AI-generated text in her reporting for years, and passing it off as her own writing. There is clear and convincing evidence that Reed has been copying from large language models for the vast majority of her independent journalism career, and that this has fundamentally altered how she presents her stories to her audience.

*The Needle* reached this conclusion by analyzing every article she published to her outlet, *Erin in the Morning,* from its founding in July 2022 through Aug. 2026\. This includes all 754 articles on her solo byline, totaling 713,163 words. 

We then used machine learning and stylometrics to analyze the growing influence of large language models (LLMs) on her work. Our analysis shows that Reed’s writing changes at the same time (and in the ways) that AI models changed. LLMs tend to be more long-winded than humans, and Reed’s writing has gotten more long-winded since the release of ChatGPT. Over a three year span, her writing went from being at grade level for a high school junior to being as difficult to read as an academic paper.

The differences are visible line by line. The paragraphs which we can prove Reed pasted from somewhere outside her normal editing process are longer, more complex, and far harder to read. The stuff she is pasting without proper citation is stylistically different from her other paragraphs. Those style differences largely match the known tells of LLM writing. 

Since Feb. 2026, at least 60% of her sentences show signs of AI influence. Erin Reed’s writing style now more closely resembles the writing style of an LLM than it resembles her own writing from before the release of ChatGPT less than four years ago.

The rise of LLM-written text also saw her use em dashes for the first time. While she claims that she has “*always*” used a lot of em-dashes, the mark was not written in her articles a single time before the release of ChatGPT. Her recent writing uses them 3-4 times as often as the average human author.

Her articles are also full of errors, many of which are directly attributable to LLM hallucinations. In one case, this resulted in her referring to a trans man journalist by the women's equivalent of his real name. This and dozens of other unacknowledged errors are still visible in her reporting.

She once denounced another news site for feeding her own article into an LLM and publishing it as their own. According to all available evidence, she makes articles in a similar way.

Reed has never told her readers that she uses LLMs to generate any portion of her articles. She is hiding her reliance on LLM writing from her readers, and presenting the machine’s work as her own.

- *\-To see how AI is affecting journalism as a whole, read Sections 1 and 2.*
- *\-To see how Erin Reed uses AI, read Sections 3, 5, and 6.*
- *\-To see how prominent Erin Reed is in the trans news ecosystem, read Sections 4.*
- *\-To see what Reed’s AI use has done to her journalism, read Section 8.*
- *\-The analysis of why this matters is in Sections 8 and 9.*
- *\-The methodology is at the bottom of this article.*

#### Get sharp, independent reporting in your inbox.

[Sign Up Here](javascript:) 

## **1: The Journalist, the Reader, and the Robot**

Using LLM in journalism, whether for writing, research, and/or analysis, is not harmless. It fundamentally alters the connection between the author and the reader for the worse. 

News readers rely on journalists to take complex information about the world and summarize it in a way that can be understood by non-specialists. Every part of a journalist’s job, from research, to talking with sources, to formatting the information, to fact checking, to writing in an accessible way, is in service of this goal.

Journalists using chatbots to research and write their work gets in the way of all this.

### **1.1: Undisclosed AI Use**

The point of journalism is to tell human stories. Humans have a way of understanding the world and their role within it which AI does not comprehend. Telling human stories with AI words might get something on a screen, but the true author will never care about the people involved. 

When a journalist brings an AI into their work, they are bringing on a storyteller that doesn't care about the stories in question. This undermines the reasons people ought to care about journalism. If the storyteller doesn't care about their own stories, why should anyone else?

That’s why people have a right to know who is telling the stories they read, and to know why that storyteller cares about the subject. 

By putting your own name and face onto something, you are saying that you care about the story. If you care about something, it is a sign to others that they should care as well.

That’s why journalists are proud to put their name on their work. That’s also why those who let an LLM put its words into their own mouth usually try to hide it. Using non-human words in this fundamentally human pursuit separates the author from the very human connections that make their stories meaningful.

People have a [right to know](https://humanstatement.org/?ref=theneedlenews.com) whether they are in [human hands or robotic hands](https://danielmiessler.com/blog/ai-influence-level-ail?ref=theneedlenews.com). 

*On this topic, we should disclose that the display charts in this story were converted into their HTML form with the assistance of automated tools. No other tools were available that could do this while still keeping them legible.* 

*The software used to understand the data were scripts coded in R using various R language packages, with final data processing and rendering done in Positron using the Quarto engine. Every word inside the charts, and every design decision, was made by humans. All decisions and interpretations of what the data means were made by humans.* 

### **1.2: AI in Reporting**

This is as true in journalism as it is everywhere else.

Earlier this year, the queer news site *PinkNews* announced it was moving into a ‘[reporter-free newsroom](https://www.wearequeeraf.com/pinknews-to-move-to-a-reporter-free-newsroom-with-its-journalists-facing-redundancy/?ref=theneedlenews.com)’ after laying off several reporters. In the time since, one former employee [stated](https://www.linkedin.com/posts/sophieperryportfolio%5Fpinknews-is-making-more-staff-redundant-share-7496202936885260289-nk3b/?ref=theneedlenews.com) that it has been “*heading down the route of AI slop*”. 

By cutting journalists, and opening themselves up to accusations of AI writing, *PinkNews* took what little reputation they had and set it on fire. 

Anyone using AI to write for them opens themselves up to all the biases and bigotries included in the LLM’s training data. That training data is filled with far more writing from bigots than from queer people. By using an LLM’s statistical odds for what the [next word](https://www.jair.org/index.php/jair/article/view/16665/27172?ref=theneedlenews.com) should be, writers writing for a queer audience expose themselves to the very biases that their audience wants to avoid.

If readers come to an outlet to see their own perspective reflected on the page, and are instead met by the ramblings of a robot, they have no reason to stick around.

Cutting their own journalists in favor of LLMs has not saved this dying news outlet. According to reports, they are set for [another round](https://archive.is/20260821045107/https://pressgazette.co.uk/news/pink-news-to-cut-up-to-11-staff-and-shut-office-amid-snapchat-changes/?ref=theneedlenews.com) of layoffs; the second one this year.

If *PinkNews* had been open about what a reporter-free newsroom means in practice, they could have at least addressed the controversy around AI use, rather than having it define their public image.

It also would have shown that they respect their readers. Their readers could have weighed in, and they could have taken their concerns into account. 

Maybe even more of their audience would have rejected their work. Others might have been fine with it.

Either way, it would have given the reader the chance to choose what they want. News outlets cannot survive without the trust of their readers. 

## **2: Identifying AI Writing**

When an author steals their words and/or ideas from a single source, [we call this plagiarism](https://www.youtube.com/watch?v=yDp3cB5fHXQ&ref=theneedlenews.com). When an author steals from the whole world, we call this [AI-generated writing](https://archive.is/20260320160733/https://www.theatlantic.com/technology/2026/03/hypocrisy-ai-industry/686477/?ref=theneedlenews.com).

This isn’t only a social consensus, or an editorial opinion held by *The Needle.* According to a [court transcript](https://storage.courtlistener.com/recap/gov.uscourts.nysd.612697/gov.uscourts.nysd.612697.1587.1.pdf?ref=theneedlenews.com) of the case against Microsoft, OpenAI, and other AI companies for copyright infringement against news publishers:

> Defendants repeatedly copied millions of Plaintiffs’ copyrighted articles in their entirety without permission to produce substitutive commercial AI products. OpenAI’s Head of ChatGPT wrote that “\[p\]ublishers” face an “existential threat” from those products, SF1466, which, he said, “are largely substitutive, period” and “will get more and more substitutive as they get better.” 

A [Microsoft executive said](https://arstechnica.com/tech-policy/2026/09/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history/?ref=theneedlenews.com) much the same thing by describing the training process for chatbots as the “*largest theft of labor in human history*”.

Despite the enormity of the problem, proving that a given piece of text was generated by an LLM is not an exact science. This is especially true when human and AI writing is mixed into a single text, as is the case for most of the articles we look at.

To identify plagiarism in a text, it can be as easy as finding the stolen text, and comparing it to the work accused of plagiarism.

By comparison, identifying when a text is written by AI is much harder. LLM-generated writing is created by taking an incomprehensibly large amount of writing, learning facts and styles from them, then mixing this into a slurry so complex that not even its programmers are capable of fully understanding it. That slurry can then be asked to write a text based on what it was trained on. The resulting slop draws from an unknowable number of sources, [stealing from](https://www.theguardian.com/books/2026/apr/12/is-ai-the-greatest-art-heist-in-history?ref=theneedlenews.com) all of them, while obscuring the true sources of information.

One [article from](https://lawreview.uchicago.edu/online-archive/plagiarism-copyright-and-ai?ref=theneedlenews.com) the *University of Chicago Law Review* corroborates this.

> “...generative AI is fundamentally different from research tools that link to original sources. A large language model (LLM) like ChatGPT produces text by predicting likely word sequences based on patterns in its training data, rather than by retrieving and crediting specific prior works. The result is a form of probabilistic generation of new content that makes it difficult or impossible to figure out why any given output is generated, much less which training material contributed to it or whether the output’s core “idea” is similar to the idea in a training source. AI output thus may implicitly contain facts or ideas from training data, but without attribution. In short, it breaks traditional citation chains.”

Even if the AI were only stealing from a single source, finding that source (and thus finding the proof of theft) would be like finding a needle in a haystack. 

That's why finding LLM-generated text looks nothing like finding plagiarism. Tests for LLM-generated text look at the style and patterns of the text itself, since these are usually very different between humans and AI.

When one text shows signs of AI influence, its author comes under suspicion. When hundreds of articles from the same author can be shown to have signs of AI influence, ‘suspicion’ is no longer the right word.

Even so, proving that AI was involved is not an exact science. Even if the odds of AI use go from 99% to 99.99999999999…%, it is nearly impossible to turn that into a 100% chance through data analysis alone. To get that level of evidence, one would have to hack into the author’s computer, film them in the act, or get the author to admit it. 

From the University of Berkeley’s “*Understanding Science 101*” [web page](https://undsci.berkeley.edu/understanding-science-101/the-core-of-science-relating-evidence-and-ideas/?ref=theneedlenews.com):

> In this case, the term argument refers not to a disagreement between two people, but to an evidence-based line of reasoning — so scientific arguments are more like the closing argument in a court case (a logical description of what we think and why we think it) than they are like the fights you may have had with siblings. Scientific arguments involve three components: the idea (a hypothesis or theory), the expectations generated by that idea (frequently called predictions), and the actual observations relevant to those expectations (the evidence).

From another [resource](https://thelogicofscience.com/2016/04/19/science-doesnt-prove-anything-and-thats-a-good-thing/?ref=theneedlenews.com):

> The reality is that the scientific method deals in probabilities, not proofs.

In short, the scientific standard for ‘proving’ something’ is impossible by design. Because of that design, humans have a [long established method](https://www.worldhistory.org/Scientific%5FMethod/?ref=theneedlenews.com) of understanding when the evidence we do have is good enough to act on something and make evidence-based conclusions.

The same [researcher wrote](https://thelogicofscience.com/about/?ref=theneedlenews.com):

> In short, we can never be 100% that our perception of reality is accurate, and scientific experiments are virtually impossible to totally and completely control. Further, science often uses inductive logic, and it relies on probabilities to draw conclusions. All of this prevents science from ever proving anything with absolute certainty. That does not, however, mean that science is untrustworthy, or that you can reject it whenever you like. Science tells us what is most likely true given the current evidence, but it is a skeptical process that always acknowledges the possibility of being wrong. Thus, it is technically possible that gravity isn’t true, the sun moves around the earth, vaccines don’t work, etc., but that doesn’t mean that you should assume that those ideas are correct. Similarly, it is also technically possible that I could win the lottery, but that doesn’t make buying a lottery ticket a good investment. Even so, it is technically possible that scientific conclusions are wrong, but it is ludicrous to believe that \[they are wrong\] unless you have strong scientific evidence that discredits them.

When people following the scientific method say that they have ‘proven’ something, they mean they have proven it to the standard above, not that they can prove it to a mathematical certainty. It also doesn't mean they can prove it to a legal standard.

Still, there are established standards by which a certain level of suspicion, when combined with enough evidence, is treated as actionable. When there is clear and convincing evidence of hidden AI use across hundreds of articles, and when there is no longer any reasonable explanation for this besides the use of AI, something has to be done about it.

*The Needle* is not alone in this opinion. 

The University of Cambridge [directly compares](https://www.hps.cam.ac.uk/students/academic-misconduct?ref=theneedlenews.com) AI use to plagiarism.

Its academic misconduct page makes clear that the university “*will not distinguish between the input of AI and the input of another human being in drawing the line between permissible help and plagiarism. A level of assistance that would be inappropriate if received from a person will also be inappropriate if received from AI.*” In practice, this prohibits “*presenting text, ideas and other material generated by AI tools as their own work*”. 

One [article](https://ojdla.com/articles/artificial-intelligence-and-plagiarism?ref=theneedlenews.com) from the *Online Journal of the Distance Learning Administration* makes it clear that AI use is not only plagiarism, but that allowing it degrades learning and academia as a whole. 

> Suppose one university does not take plagiarism seriously, whether AI or something else; it affects all institutions of higher learning. We have a responsibility to uphold the integrity of academia.

Now, swap university for ‘journalism’, swap institutions of higher learning with ‘newsrooms’, and swap integrity of academia with ‘integrity of the fourth estate’. If you do this, you can see the problem with undisclosed *AI use*, whether in journalism, or any other field where citing evidence and sources matters. 

## **3: A Litany of Errors**

Journalists get things wrong sometimes, despite our best efforts. We are human after all (*or at least we should be*). When we find errors after publication, we are expected to do something about it. 

However, if journalists keep getting facts wrong, and especially if they get things wrong in a consistent way, they are expected to change how they obtain those facts. 

AI cannot be trusted to obtain and verify information. Therefore, AI cannot be trusted for anything where evidence is important, including journalism.

Not even forcing an LLM to correctly cite its own sources is enough to correct factual errors. [According to](https://lawreview.uchicago.edu/online-archive/plagiarism-copyright-and-ai?ref=theneedlenews.com) the *University of Chicago Law Review*:

> If you ask an LLM to give you citations, it will, but quite often it will make those up too. More than 150 [lawyers](https://natlawreview.com/article/lawyers-continue-get-hot-water-citing-ai-hallucinated-cases?ref=theneedlenews.com) (and even a couple of [judges](https://perma.cc/W6TP-H6D3?ref=theneedlenews.com)) who have been caught [citing cases that don’t exist](https://scholar.google.com/scholar%5Fcase?case=7204698938630410967&q=Shahid+v.+Esaam+&hl=en&as%5Fsdt=400006&ref=theneedlenews.com) or introducing other AI errors can testify to that. And while new techniques like “[retrieval-augmented generation](https://web.stanford.edu/~jurafsky/slp3/ed3book%5FJan25.pdf?ref=theneedlenews.com)” and “[deep research](https://arxiv.org/pdf/2506.18096?ref=theneedlenews.com)” target some of the problems of entirely invented citations by having the LLM focus on particular sources from the real world, LLMs using those methods often [get things wrong](https://dho.stanford.edu/wp-content/uploads/Legal%5FRAG%5FHallucinations.pdf?ref=theneedlenews.com) even if the sources they cite really exist—sometimes as much as [90%](https://www.nature.com/articles/s41467-025-58551-6.pdf?ref=theneedlenews.com) of the time. And some of those [hallucinations](https://arxiv.org/pdf/2308.04635?ref=theneedlenews.com) can be life threatening, as when Google’s Bard answered a question asking what to do in case of a seizure by taking text from a reputable medical site describing what not to do. 

Erin Reed's writing shows all the hallmarks of these kinds of errors. 

This is especially true when her writing uses pasted in quotes from real people. When she does this, she opens herself up to massive risks. *For the technical details behind this, please read section 6.2.*

If her quotes and information about sources are entering her writing via an LLM, this runs the risk that the quotes and information may be AI-generated hallucinations, or that the LLM is altering quotes and information about real people. 

The more we looked into her writing, the more we came to believe this is key to why she so often gets facts wrong, even when the correct answer can be found in her own cited sources. 

The certainty of AI in its own errors would also explain why Reed almost never acknowledges her mistakes. 

When *The Needle* [fact-checked](https://theneedlenews.com/2025/11/fact-check-erin-reed-is-lying-about-the-risk-of-trans-healthcare-being-cut-at-the-federal-level/) one of her articles last year, rather than responding to us or adjusting her coverage, she simply blocked us for pointing it out. Only a few months prior, Reed had interviewed the author of that article (Jane Migliara Brigham) for a paid reporting position at EITM.

The only correction *The Needle* could find evidence of Reed having ever published was a [2023 correction](https://web.archive.org/web/20230129205255/https://erininthemorn.substack.com/p/this-year-get-girl-scout-cookies) to her list for ordering Girl Scout cookies from trans Girl Scouts. That correction has since been deleted, and quietly rolled into the original post. The only evidence that this correction [ever existed](https://web.archive.org/web/20230116220829/https://erininthemorn.substack.com/p/correction-on-previous-erin-in-the) is buried in the Wayback Machine. 

Erin Reed takes errors on her list about children’s cookies more seriously than she does quoting the right politician when using their quotes in her articles.

This is how Reed wrote [an entire article](https://web.archive.org/web/20260915230527/https://www.erininthemorning.com/p/maine-governor-janet-mills-comes) on anti-trans politicians in Maine, while claiming Graham Platner was defending the rights of trans kids. In reality, the supportive quote belonged to Jordan Wood, a different politician from a different race. Her source [cites the quote correctly](https://washingtonsun.com/2026-election/maine-debate-trans-student-athletes-campaign?redirected=notus%5Forg&ref=theneedlenews.com), but her own article gets it wrong.

The source she cited for this does talk about Platner and Wood. The quote from Wood is in the paragraph before the cited article talks about Platner, but it does not directly quote any of Platner’s words.

The Platner misattribution is not the only time she has ever mis-attributed a statement or named the wrong person with a quote. This is exactly the kind of error you would expect to see when someone takes real quotes from real people, and runs them through an LLM. The AI can hallucinate, get details mixed up, and misattribute information.

Misattributing a politician's statement is not the only error Reed has made this year. 

Her litany of errors in 2026 includes deadnaming and misgendering a competing trans journalist who broke a story for one of her competitors! This was [seven months](https://web.archive.org/web/20260309113942/https://www.erininthemorning.com/p/the-lavender-ledger-your-strategic-871) ago, and the deadname is[ still up](https://web.archive.org/web/20260914004519/https://www.erininthemorning.com/p/the-lavender-ledger-your-strategic-871) as of writing. 

[This wasn’t just any story](https://assignedmedia.org/breaking-news/kansas-revokes-license-no-gender-change/?ref=theneedlenews.com). It was Nate Zuke’s first article that we can find. Many of you have already read it: ‘[She’d Never Changed Her Gender Marker. Kansas Invalidated Her License Anyway.](https://assignedmedia.org/breaking-news/kansas-revokes-license-no-gender-change/?ref=theneedlenews.com)’

It was excellent reporting; arguably one of the most important trans news stories written by a trans journalist this year. 

And it was Erin Reed who deadnamed him to her audience of hundreds of thousands.

It is unclear how Erin Reed would have found Nate Zuke’s deadname without AI. As of Sep. 2026, we could find no bylines under his deadname, either at the outlet he published this story in, or anywhere else. We found nothing in his public presence that would connect him to his deadname. We also found no public presence under his deadname whatsoever.

Archived versions of *Assigned Media’s* coverage from the day it went live, as well as posts by other journalists tagging Nate Zuke who broke the story, show that this story was never published under any name other than ‘Nate Zuke’.

We checked across the internet, and used three different search engines. The only documented proof of the existence of anyone with that name is from public court records in a different state from where he is known to live, and where we have no reason to believe he has ever lived in. 

One explanation is that this name was somewhere in the AI training database, and that it (accurately or not) attributed that name to Nate Zuke.

Another explanation is that an LLM hallucinated his gender, and made up a woman's name for him. This is a [documented problem](https://www.reddit.com/r/claudexplorers/comments/1twtws1/claude%5Fthinks%5Fim%5Fa%5Fgirl/?solution=2d6ea9c8f5e2c10c2d6ea9c8f5e2c10c&js%5Fchallenge=1&jsc%5Ftoken=2824be10929bdc604753c70a67a1c33125fe4bf5b4cc7d63971848ef72511af0&jsc%5Forig%5Fr=&ref=theneedlenews.com) with some LLMs, with this error even appearing from cis people's names. 

Either way, Reed published a trans man's deadname, real or imagined.

Even if this name Reed published isn't Nate’s actual deadname, the impacts are effectively the same. 

After this, [Nate Zuke](https://bsky.app/profile/natezuke.bsky.social?ref=theneedlenews.com) went silent online for nearly 4 months, and has barely posted since. In the seven months since Reed deadnamed him, he has not posted any more articles.

*The Needle was unable to reach Nate Zuke.*

Either Erin Reed deliberately deadnamed a competing trans journalist who, in his very first article, broke one of the biggest trans news stories this year, or she didn't care enough to double-check the output of her LLM, and has been unwilling to correct the record for the last seven months.

This level of open disrespect for trans people goes beyond even what is typical of transphobic papers like the New York Times or the BBC. To find outlets where behavior like this is the norm, you would have to go to openly bigoted right-wing rags.

Even fan-based wikis hold themselves to a [higher standard](https://deltarune.wiki/w/Kris?ref=theneedlenews.com) on referring to trans people than the biggest trans journalist in the world does.

Given Reed’s near-monopoly on our community’s information regarding our civil rights, fuck-ups like this are unacceptable. If Reed gets basic information wrong, it spreads farther than any other trans person’s errors, bar none. 

Reading through Reed’s 106 articles under her solo byline from January to August of this year, and checking the citations and claims therein, *The Needle* was able to find at least 52 factual errors.

To our knowledge, not one has received a correction, retraction, or editor’s note.

*We published a companion article detailing every one of Reed’s unretracted errors that we found from Jan. to Aug. 2026\. There were far too many to include here.*

[READ THE FACT CHECK ARTICLE HERE](https://theneedlenews.com/erin-reed-unretracted-errors)

## 4: Erin Reed's Reach

In the early 2020s, mainstream media burned every iota of trust trans readers might’ve had in them. This was due to [editorial decisions](https://theonion.com/it-is-journalism-s-sacred-duty-to-endanger-the-lives-of-1850126997/?ref=theneedlenews.com) made to not cover us in a fair and accurate manner, most [famously](https://www.advocate.com/politics/national/new-york-times-transgender-controversy?ref=theneedlenews.com) at the [New York Times](https://www.nytimes.lgbt/?ref=theneedlenews.com).

Trans journalism (meaning news sources catering to a trans audience) developed as a response to these failures.

This led to [*Erin in the Morning’s* (EITM) rise](https://www.intomore.com/impact/politics/erin-reed-in-the-morning-interview-trans-bills-2024/?ref=theneedlenews.com) as one of the first major independent trans newsletters. In 2026, it is now the largest, most famous, and most influential news source that markets itself to a trans audience which has ever existed.

Starting her outlet in 2022, Erin Reed was the first person to build success selling trans news to a trans audience. This has made her into a household name among trans people in the United States.

Her position in the field of trans journalism is dominant. In Oct. 2024, in the lead up to the 2024 election, EITM received 97% of all page views on trans news sites. While trans people panicked about what a Trump return to power would mean for us, Reed's writing was how many of us came to understand that potential future. *That includes the authors of this article.*

Her writing set the tone of the trans community at large. Back then, she was arguably the single most influential trans person among trans people in the country.

No single trans person since has been able to match that.

One reason for this is that Reed has been fast and prolific. In early 2023, she began publishing notably longer articles, with more complex and long-winded sentences, while producing them at a faster pace.

According to a dataset of six trans news outlets (including *The Needle*), EITM has the majority of views in trans journalism in most months for which we have data.

#### Share of Page Views, Trans Journalism

Percent of monthly traffic across six trans news outlets, August 2024 to August 2026. Erin in the Morning is the most read trans news outlet in almost every month we have data for. It peaked at 97% of views in Oct. 2024, and held a near-monopoly through May 2025. 0% 25% 50% 75% 100% Erin in the Morning, August 2024: 188,978 (78.0%) Erin in the Morning, September 2024: 144,180 (87.9%) Erin in the Morning, October 2024: 285,751 (96.9%) Erin in the Morning, November 2024: 405,395 (92.8%) Erin in the Morning, December 2024: 274,003 (83.6%) Erin in the Morning, January 2025: 536,868 (90.0%) Erin in the Morning, February 2025: 521,575 (82.6%) Erin in the Morning, March 2025: 737,427 (79.6%) Erin in the Morning, April 2025: 325,304 (90.8%) Erin in the Morning, May 2025: 287,479 (92.8%) Erin in the Morning, June 2025: 231,185 (65.9%) Erin in the Morning, July 2025: 279,803 (72.0%) Erin in the Morning, August 2025: 297,852 (72.6%) Erin in the Morning, September 2025: 752,619 (88.9%) Erin in the Morning, October 2025: 357,450 (63.9%) Erin in the Morning, November 2025: 175,668 (63.3%) Erin in the Morning, December 2025: 156,976 (65.8%) Erin in the Morning, January 2026: 368,120 (68.2%) Erin in the Morning, February 2026: 390,594 (65.0%) Erin in the Morning, March 2026: 244,711 (36.7%) Erin in the Morning, April 2026: 221,149 (51.1%) Erin in the Morning, May 2026: 107,000 (43.7%) Erin in the Morning, June 2026: 174,513 (63.8%) Erin in the Morning, July 2026: 190,343 (44.0%) Erin in the Morning, August 2026: 216,493 (69.5%) Assigned Media, August 2024: 51,051 (21.1%) Assigned Media, September 2024: 14,704 (9.0%) Assigned Media, October 2024: 5,432 (1.8%) Assigned Media, November 2024: 27,782 (6.4%) Assigned Media, December 2024: 46,014 (14.0%) Assigned Media, January 2025: 39,132 (6.6%) Assigned Media, February 2025: 75,480 (12.0%) Assigned Media, March 2025: 139,504 (15.1%) Assigned Media, April 2025: 26,521 (7.4%) Assigned Media, May 2025: 10,450 (3.4%) Assigned Media, June 2025: 61,844 (17.6%) Assigned Media, July 2025: 40,789 (10.5%) Assigned Media, August 2025: 22,518 (5.5%) Assigned Media, September 2025: 37,590 (4.4%) Assigned Media, October 2025: 26,373 (4.7%) Assigned Media, November 2025: 5,509 (2.0%) Assigned Media, December 2025: 33,055 (13.9%) Assigned Media, January 2026: 21,329 (4.0%) Assigned Media, February 2026: 46,804 (7.8%) Assigned Media, March 2026: 95,124 (14.3%) Assigned Media, April 2026: 30,862 (7.1%) Assigned Media, May 2026: 13,029 (5.3%) Assigned Media, June 2026: 14,619 (5.3%) Assigned Media, July 2026: 136,021 (31.5%) Assigned Media, August 2026: 27,238 (8.7%) TransLash, August 2024: 2,167 (0.9%) TransLash, September 2024: 5,195 (3.2%) TransLash, October 2024: 3,772 (1.3%) TransLash, November 2024: 3,548 (0.8%) TransLash, December 2024: 7,891 (2.4%) TransLash, January 2025: 20,421 (3.4%) TransLash, February 2025: 34,509 (5.5%) TransLash, March 2025: 42,429 (4.6%) TransLash, April 2025: 5,741 (1.6%) TransLash, May 2025: 11,011 (3.6%) TransLash, June 2025: 8,372 (2.4%) TransLash, July 2025: 7,740 (2.0%) TransLash, August 2025: 9,965 (2.4%) TransLash, September 2025: 26,311 (3.1%) TransLash, October 2025: 7,415 (1.3%) TransLash, November 2025: 10,620 (3.8%) TransLash, December 2025: 24,362 (10.2%) TransLash, January 2026: 27,597 (5.1%) TransLash, February 2026: 14,005 (2.3%) TransLash, March 2026: 8,393 (1.3%) TransLash, April 2026: 27,831 (6.4%) TransLash, May 2026: 20,911 (8.5%) TransLash, June 2026: 9,079 (3.3%) TransLash, July 2026: 8,176 (1.9%) TransLash, August 2026: 17,040 (5.5%) Trans News Network, July 2025: 2,134 (0.5%) Trans News Network, August 2025: 3,549 (0.9%) Trans News Network, September 2025: 338 (0.0%) Trans News Network, October 2025: 48,865 (8.7%) Trans News Network, November 2025: 859 (0.3%) Trans News Network, December 2025: 1,029 (0.4%) Trans News Network, January 2026: 55,263 (10.2%) Trans News Network, February 2026: 16,280 (2.7%) Trans News Network, March 2026: 1,995 (0.3%) Trans News Network, April 2026: 22,603 (5.2%) Trans News Network, May 2026: 1,952 (0.8%) Trans News Network, June 2026: 291 (0.1%) Trans News Network, July 2026: 1,703 (0.4%) Trans News Network, August 2026: 1,115 (0.4%) Transitics, March 2025: 7,043 (0.8%) Transitics, April 2025: 843 (0.2%) Transitics, May 2025: 947 (0.3%) Transitics, June 2025: 49,362 (14.1%) Transitics, July 2025: 58,338 (15.0%) Transitics, August 2025: 76,156 (18.6%) Transitics, September 2025: 29,670 (3.5%) Transitics, October 2025: 112,828 (20.2%) Transitics, November 2025: 79,270 (28.5%) Transitics, December 2025: 17,442 (7.3%) Transitics, January 2026: 26,405 (4.9%) Transitics, February 2026: 105,830 (17.6%) Transitics, March 2026: 257,651 (38.7%) Transitics, April 2026: 114,825 (26.6%) Transitics, May 2026: 79,104 (32.3%) Transitics, June 2026: 64,853 (23.7%) Transitics, July 2026: 64,455 (14.9%) Transitics, August 2026: 39,792 (12.8%) The Needle, October 2025: 6,560 (1.2%) The Needle, November 2025: 5,760 (2.1%) The Needle, December 2025: 5,560 (2.3%) The Needle, January 2026: 41,140 (7.6%) The Needle, February 2026: 27,050 (4.5%) The Needle, March 2026: 58,120 (8.7%) The Needle, April 2026: 15,190 (3.5%) The Needle, May 2026: 22,950 (9.4%) The Needle, June 2026: 10,220 (3.7%) The Needle, July 2026: 31,730 (7.3%) The Needle, August 2026: 9,900 (3.2%) Aug 2024 Jan 2025 Jul 2025 Jan 2026 Jul 2026 March 2026 sees a growth of competing outlets. 78% 88% 97% 93% 84% 90% 83% 80% 91% 93% 66% 72% 73% 89% 64% 63% 66% 68% 65% 37% 51% 44% 64% 44% 70% Percent of viewsErin in the MorningAssigned MediaTransLashTrans News NetworkTransiticsThe Needle News 

Source: Semrush data-pull of Google Analytics and other data providers, plus The Needle’s internal view count data.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle News

March 2026 was the first month on record where EITM did not get a majority of page views. Her audience share has rebounded to 69.5% of page views in August 2026, the most recent month for which we have complete data.

#### Erin in the Morning vs. the Competition

Total website visits per outlet.

Monthly visits to six trans-news outlets, Aug. 2024 to Aug. 2026. Erin in the Morning monopolizes visits almost every month. 0 200k 400k 600k 800k Aug 2024 Jan 2025 Jul 2025 Jan 2026 Aug 2026 Nov 2024405k Mar 2025737k Sep 2025753k Feb 2026391k 2026 monthly median: Jan. 2026 to Aug. 2026 Erin in the Morning 218,821 visits/mo. (median) Transitics 71,978 visits/mo. (median) Assigned Media 29,050 visits/mo. (median) The Needle News 24,999 visits/mo. (median) TransLash 15,522 visits/mo. (median) Trans News Network 1,974 visits/mo. (median) 

4 of the 6 publishers here had at least one outlier spike in 2026\. Medians were chosen to better show the normal monthly view count. Source: Semrush data-pull of Google Analytics and other data providers, plus The Needle’s internal view count data.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle News

*Erin in the Morning* has a large impact on the trans community, and its writing sets the tone that many trans people use to understand the world. By pure numbers alone, *Erin in the Morning’s* writing matters more than any other source made for a trans audience.

## **5: Erin Reed’s AI Stance**

Erin in the Morning rose in popularity at the same time that Artificial Intelligence (AI) [rose in popularity](https://nymag.com/intelligencer/article/openai-chatgpt-ai-cheating-education-college-students-school.html?ref=theneedlenews.com). As AI has become more ubiquitous, it has become more infamous.

Erin Reed’s public position on AI has been mostly critical. She has accused other outlets of rewriting her work with an LLM, she publicly protested (after community pushback) when her writing was paired with an AI thumbnail, and she cited the rise of AI as the reason why search engines such as Google are no longer usable for journalists. 

She is especially harsh when the signs of AI can be seen in journalism.

The clearest example of this was when she [accused the news site](https://bsky.app/profile/erininthemorning.com/post/3ls2b5yij6s24?ref=theneedlenews.com) *TransVitae* of rewriting her article and releasing it as their own. She also identified other articles on the site that she believed to be AI written.

*The Needle* believes that Reed is correct about her article being stolen.

She avoided saying whether she was 100% certain that the articles were written with AI. She [previously said](https://bsky.app/profile/erininthemorning.com/post/3lolu2k73xs2n?ref=theneedlenews.com) that “*\[i\]t’s impossible to prove*” whether a text is AI-written, citing [the unreliability](https://www.npr.org/2025/12/16/nx-s1-5492397/ai-schools-teachers-students?ref=theneedlenews.com) of AI detection tools.

[She also objected](https://bsky.app/profile/erininthemorning.com/post/3lmskcaf6st2v?ref=theneedlenews.com) to her writing being paired with an AI thumbnail, demanding the outlet in question replace it immediately, which they did. 

As for how AI affects the work of journalists, her thoughts are mixed. She says that AI search results are the reason that Google and other search engines are no longer useful for journalists,[ saying](https://bsky.app/profile/erininthemorning.com/post/3lbq44df6im2v?ref=theneedlenews.com) that “*\[t\]he signal to noise ratio plummeted*” after the rise of AI.

The one positive sentiment she has expressed regarding AI is in its usefulness as a [research tool](https://bsky.app/profile/erininthemorning.com/post/3loluswg5bs2n?ref=theneedlenews.com).

> The sad thing is, I think there are use cases for AI. I couldn’t tell you the number of times I’ve been able to find obscure studies and articles that Google just no longer is good for finding.

Even here, she only admits to using AI in the way she once used Google. 

She clearly understands that the reading public is increasingly opposed to AI in writing. Her public comments on the subject of AI are in line with this sentiment. 

### **5.1: AI-Assisted Language Translation**

Erin Reed has already used AI for translating primary sources from foreign languages, and used the translation as the basis of an article. In other words, we know that AI has been the basis of one of her articles, and we can prove it using only her own words.

On May 7, 2025, [she said](https://bsky.app/profile/erininthemorning.com/post/3loluswg5bs2n?ref=theneedlenews.com) that:

> The sad thing is, I think there are use cases for AI. I couldn’t tell you the number of times I’ve been able to find obscure studies and articles that Google just no longer is good for finding. It’s also light years better at language translation than google translate.  
>  
> But that’s not how it’s used. 

In the same Bluesky thread, she related this form of AI use to one of her past articles, [saying that](https://bsky.app/profile/erininthemorning.com/post/3lolxk6dk3k2n?ref=theneedlenews.com) “*\[l\]iterally the entire reason the German trans care guidelines reached the English speaking world instantly is because of an AI translation of the 400 page document.*”

The story in question was ‘[New German, Swiss, And Austrian Guidelines Recommend Trans Youth Care, Slam Cass Review](https://www.erininthemorning.com/p/new-german-swiss-and-austria-guidelines?ref=theneedlenews.com)’. This was about a German language medical association releasing guidelines defending the necessity of trans kids having access to trans healthcare. 

Reed’s article was released on Sunday, March 9, 2025\. 

The article falsely states that the guidelines were “*released Friday in German*”. That would mean that the German release date was Friday, March 7, 2025\. If true, that would mean that the 400 page document was translated from a foreign language, read, analyzed, and then reported on in 2 days. 

In reality, the original study was [released](https://register.awmf.org/de/leitlinien/detail/028-014?ref=theneedlenews.com) in German on Sep. 30, 2024\. The official English language translation appears to have been released on June 24, 2025, three months after Reed’s article, and a month after she made the statements above.

![image.png](https://storage.ghost.io/c/07/58/0758f73b-8a0d-4fa1-a055-4d836e2c23c5/content/images/2026/09/01b3e3f6-b125-4264-ac15-a81ca8fcaffb.png)

The German language page where the guidelines are hosted lists the release date of the original German language document as Sep. 9, 2024\. This is seven months earlier than the release date cited in Reed’s article.

The original release date is listed as the document’s ‘Stand’, a word which in German roughly translates to ‘status’ in English, and which can be used to refer to a release date. However, it both looks and sounds like the English word ‘stand’, which has a very different meaning. This would make a machine (or someone unfamiliar with the German language) confused, and unlikely to correctly identify this as the original release date.

However, the date below it is labeled as “*Gültig bis:*”, translating into ‘valid until:’. Even if a translator misses where the release date is, a human can easily find it in context by looking at the date above this. That human would not need to speak German to do so; the translators built into many web browsers and the surrounding context clues would be enough.

She would not even have needed to read German to learn this. The actual release date is the first date displayed on the page. That date is shown in an international format (as are the other dates on the page). 

Even Google Translate (which Reed has[ stated](https://bsky.app/profile/erininthemorning.com/post/3loluswg5bs2n?ref=theneedlenews.com) is not as reliable as AI for translations) can translate the words into English with enough accuracy to verify this, provided a human could look at the surrounding context. When *The Needle* checked, Google Translate could more accurately understand this German language page (when read in context) than whichever AI translated that page. 

*The Needle* also has a team member that can read and speak basic German, Jane Migliara Brigham, who confirmed this. Two native German speaking sources confirmed Brigham’s translation as well.

Most importantly, the date ‘March 7, 2025’, which Reed cites as the actual release date, does not appear anywhere on this page for any purpose.

There are 2 dates on the page listed in the year 2025\. However, both are dated after Reed’s article on the subject, meaning she could not have referenced them.[ ](https://web.archive.org/web/20250310181351/https://register.awmf.org/de/leitlinien/detail/028-014)

[An archive](https://web.archive.org/web/20250310181351/https://register.awmf.org/de/leitlinien/detail/028-014) of this page from March 10, 2025, the day after the article’s release, shows that the only date in the year 2025 up to that point was marked as March 10, meaning that it could not have been referenced in the March 9 article.

This is corroborated by the [last available archive](https://register.awmf.org/de/leitlinien/detail/028-014?ref=theneedlenews.com) from before Reed’s article went live, which confirms that the date which Reed claimed as the release date cannot be found on this page.

If Reed had used any of these options, or had simply spoken to a German speaker, this error would not have happened.

We do not know where the alleged release date of March 7, 2025 came from. Someone (or something) involved in this article made up a release date, and Reed did not verify it before she released the story. 

Seeing as Reed credited an AI translation with making the source available, it is reasonable to assume that this release date is an AI hallucination.

If AI really was “[*light years better*](https://bsky.app/profile/erininthemorning.com/post/3loluswg5bs2n?ref=theneedlenews.com)” at translating foreign language documents than Google Translate, this error probably would not have happened. 

Reed’s story on this German study had a reach beyond her own outlet. Her work was adapted for a story in [*The Advocate*](https://www.advocate.com/health/europeans-recommend-gender-affirming-care?ref=theneedlenews.com), under her own byline. That story once again cites the same incorrect release date. 

Her links for the release date in both articles redirect to the very German web page that *The Needl*e used to fact-check her statements.

As of the release of our article, Reed’s incorrect date can still be found in both her original reporting, and in her article at *The Advocate*.

---

If this was the only evidence we had that Erin Reed was using AI in her work, we would not have made this article. The scale of the problem is *so much worse*.

## **6: Erin Reed Uses AI to Write for Her**

There is [clear and convincing evidence](https://www.law.cornell.edu/wex/clear%5Fand%5Fconvincing%5Fevidence?ref=theneedlenews.com) that Erin Reed has been using LLM-generated text in her articles for years, and that the evidence can be found in hundreds of her articles. 

Our claim is not just that she generated a few of her articles with an LLM, while presenting them as her own words. Saying this would undersell how bad the problem is. 

There are hundreds of articles under her byline at EITM that have sections pasted from an external source that we can identify as being an LLM with scientific precision. In every single case, Reed is lying to her audience by failing to acknowledge that she is not the sole author of her articles.

Erin Reed, owner of *Erin in the Morning*, and one the most influential people on the topic of how the trans public sees the world, has been using AI-generated writing in her reporting for the vast majority of her career as a public figure. The authors of this article both believe this to be true, and that there is no other reasonable explanation for the suspicious patterns found in her writing.

### **6.1: Reed Before ChatGPT**

To understand how she writes today, we should look back to how she first wrote. The difference between the two is stark.

These early pieces are of a noticeably different style than much of what she would release in the months that followed. While she was releasing some full-length articles at this time, much of her output was short and covered a variety of formats. 

This pre-ChatGPT period is full of news roundups, links to Reed’s other social media, and the earliest version of her [legislative risk map](https://www.erininthemorning.com/p/new-map-anti-transgender-legislative?ref=theneedlenews.com).

The news roundups are notable because they are summaries of several other news stories, often drawing from other outlets. Each story gets a few paragraphs before going on to the next story. These read less like a normal news article, and more like a Twitter thread.

While her legislative risk maps are still being produced, her news roundups have been virtually absent since the end of 2022\. 

While she still releases legislative risk maps, a feature she’s become [well-known for](https://theneedlenews.com/2026/07/safe-states-dont-make-safer-people/), full length articles mostly replaced the news roundups after 2022.

This period is also the only time in our investigation when she was not working at EITM full-time. [Her LinkedIn](https://www.linkedin.com/in/erin-reed-6b874348/?ref=theneedlenews.com) lists her leaving her day job in Dec. 2022\. In addition, Dec. 2022 was the first full month after her then-girlfriend Zooey Zephyr won her election to the Montana State House.

Her writing from before the release of ChatGPT was 100% human written. It serves as a point of comparison for what comes next.

### **6.2: Pasted Marks**

ChatGPT was released to the public on Nov. 30, 2022\. Soon after, Reed’s writing began to include more Unicode marks that were pasted from an external source. That external source has all the features of contemporary LLMs.

This marked writing contains artifacts which are not found by default in standard word processors. Specifically, it contains apostrophes and quotation marks in the form most commonly found in the writing of LLMs.

### Monthly Rate of Pasted Paragraphs

Monthly Rate of Pasted Paragraphs \[DESC: one sentence on what the chart shows, for screen readers\] 0% 25% 50% 75% 100% Public launch of ChatGPT Claude Opus 4.6 launch 3 month average % pasted paragraphs Monthly % pasted paragraphs Dec 2022 Jun 2023 Dec 2023 Jun 2024 Dec 2024 Jun 2025 Dec 2025 Jun 2026 

The presence of a straight mark inside a paragraph is used as evidence it was pasted. 2025 is invisible to this test. Source: Erin Reed's Substack archive. In house analysis by The Needle.

It’s notable that the sentences containing these artifacts are stylistically different from those without them. If she were simply writing in two separate word processing programs, and copying and pasting text from one to the other, the style of writing likely wouldn’t change. 

Instead, the sentences with those marks are longer and more complex than those without them. This can be measured statistically over her career at EITM. As such, the difference between her pasted and not-pasted writing can be tracked and studied. 

Some LLMs output straight quotation marks and apostrophes. Claude [models](https://github.com/anthropics/anthropic-sdk-typescript/issues/561?ref=theneedlenews.com#issuecomment-2442675237) have done so since at least October 2024\. ChatGPT [used](https://community.openai.com/t/smart-quotes-converted-to-plain-quotes/220797?ref=theneedlenews.com) to output [straight](https://community.openai.com/t/unwanted-replacement-of-curly-symbols-to-straight-during-grammar-checking/531484?ref=theneedlenews.com) marks, but users reported a flip to curly marks in newer models between late 2023 and mid-2024\. From what evidence we can gather, straight marks have been a [default](https://the-decoder.com/heres-how-to-spot-ai-writing-according-to-wikipedia-editors/?ref=theneedlenews.com) across every major LLM’s chat interface at one point or another. However,[ documentation](https://forum.cursor.com/t/inability-to-work-with-files-with-smart-curly-quotes-apostrophes/46213?ref=theneedlenews.com) on this front is unreliable, and is mostly found within user forums. *More on this below.* 

By contrast, most word processors use curly marks, or “[*smart quotes*](https://typographyforlawyers.com/straight-and-curly-quotes.html?ref=theneedlenews.com)”, by default. In the case of quotation marks, they curve into the thing they are quoting, similar to hand-written text.

This is the default setting in the word processor used in [Substack](https://scribblewright.substack.com/p/substacks-favorite-error), the platform Erin Reed uses to publish her articles, as well as most major word processors. Hardly anyone bothers to change this. 

Straight quotation marks and apostrophes are written with [different Unicode characters](https://practicaltypography.com/straight-and-curly-quotes.html?ref=theneedlenews.com) than their curly equivalents. While casual readers might not notice the difference, that difference is obvious to a computer reading text through Unicode.

The difference is visible in the font which EITM articles are posted in. The difference between the two is faint, and hard to see if you aren't looking for it. However, if you know what to look for, you can see it with the human eye.

We were unable to determine what word processor Reed uses when writing her articles. However, unless she was writing the first draft of articles in a notepad app or in a raw .txt file editor, we cannot account for how often these marks appear in her work. And even if she were writing her first drafts in this way, this would not account for the measurable differences between the pasted and typed text.

#### Straight and Curly Marks Look Different Across Most Major Fonts

Once you see the differences, it’s hard to unsee them.

The six quotation mark and apostrophe characters, each drawn in six typefaces. THE NEEDLE’S TYPEFACES COMMON ELSEWHERE Old Standard TT headlines Noto Serif body text Arial chart labels Georgia Needle fallback Times New Roman browser default Courier New typewriter mono QUOTATION MARK U+0022 (1 byte) Basic Latin · straight LEFT DOUBLE QUOTATION MARK U+201C (3 bytes) General Punctuation · curly RIGHT DOUBLE QUOTATION MARK U+201D (3 bytes) General Punctuation · curly APOSTROPHE U+0027 (1 byte) Basic Latin · straight LEFT SINGLE QUOTATION MARK U+2018 (3 bytes) General Punctuation · curly RIGHT SINGLE QUOTATION MARK U+2019 (3 bytes) General Punctuation · curly 

Source: Glyph outlines traced against the fonts in question. Old Standard TT, Noto Serif, and Arial are used by this website. Georgia, Times New Roman, and Courier New are used in many Microsoft products. Character names, code points, and byte counts are all taken from the Unicode Character Database.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

*All numbers below use the 'presence rule' to determine the likelihood that a section had been influenced by an LLM. This means that the presence of a straight mark inside a paragraph is used as evidence it was pasted, unless otherwise specified. However, we ran other tests, including a mixed test that needed at least a majority of straight marks to count, and got largely the same results.* 

### **6.2.1: Why We Conclude That Straight Marks are Pasted From An LLM**

Computers generally do what we tell them to do. However, sometimes computer software gets really complicated, and it doesn’t always work right. These issues are called computer ‘bugs’. People who fix these bugs and build the software love to write down how they did it for other people to see.

Because of this, you can look up how the computers that run the website Erin Reed publishes on (Substack) were told to display certain symbols. What you’ll find is that those computers are running a few different layers of instructions, like a tiered wedding cake. When you look up how each layer works, you can see what the flavor of the cake is.

With the servers that run Reed’s website, the flavor of the cake is one that doesn’t have straight apostrophes or quotation marks. You have to mix in someone else’s cake to get those. It’s like her cake went from chocolate to vanilla, when her writing only has the ingredients for chocolate. 

In this analogy, chocolate means curly marks that were made in a regular word processor, or typed directly into a Substack post. Vanilla means straight marks that were pasted in from something that only puts out straight marks, like a chatbot, .txt editor, or Discord chat. 

Sometimes ingredients get mixed up, but not at the rate we see in her writing. 

When writing on a computer, people can paste from anywhere. We looked at the differences in what is pasted compared to what is not. 

When we look at the text that was pasted in, on the surface, we only know that someone copied it from somewhere. However, when you use math to analyze the differences from what she copied and pasted to what she didn’t, we see more red flags. 

These red flags mean that something is suspicious, not right, or not what it seems. Journalists tell us the news. Their writing is not supposed to be suspicious. We were investigating whether she uses a robot to write for her. So we looked at how robots write. 

When she pastes, her writing looks more like a machine’s writing. When she types, her writing looks less like a machine's writing. 

We can see this when we look at her pasted paragraphs being 3 sentences longer on average; like a personal essay rather than a news article. Her sentences get longer too when they are pasted. 

You know how your fingers leave fingerprints? The way someone writes also leaves fingerprints. These fingerprints can only be made obvious with special training or materials. This is how we analyze her writing style. 

When people do something, it usually leaves evidence. If they’re trying to hide it, they often get sloppy.

### **6.2.2: How AI Writing Looks** 

The difference in Reed’s writing between the typed and pasted paragraphs is clear, even within the same articles. To show this, we will show two paragraphs from the same article, all written since Feb..

Vibes alone are not a reliable AI detector, but we trust that most people familiar with AI writing will be able to tell the difference.

*Every paragraph in the tests below had either 100% pasted marks or 100% non-pasted marks. We did not use any paragraphs with a mix of marks in this test. This section is designed to train the reader to identify the signs in Reed’s own published work.* 

Here are two paragraphs from the article ‘[Trans Youth Suicides Skyrocketed In UK After Care Drawdown](https://www.erininthemorning.com/p/trans-youth-suicides-skyrocketed?ref=theneedlenews.com)’. One was obviously pasted; The other wasn’t.

**Paragraph 1:**

> The gap between the Appleby Report, which was used to absolve the government of any culpability in trans youth deaths, and the Good Law Project's freedom of information request is no accident, but rather, a deliberate framing choice. The Appleby Report only examined patients of GIDS. But in the aftermath of the Bell v Tavistock ruling, wait times for GIDS appointments skyrocketed, now sitting at an estimated average of 25 years. Youth who were suffering no longer saw a pathway to gender-affirming care. Referrals stalled. Diagnoses became increasingly impossible to obtain. As restrictions piled on, these young people saw their future pathways to obtaining gender-affirming care shuttered in front of them. These youth were ignored in the Appleby report.

**Paragraph 2:**

> The UK government fought hard to keep transgender suicide data from the public, and now, two years after its attempt to bury the deaths of children, that data is out. Even what has been released is limited: the NCMD’s own response indicates that its methodology may not capture all transgender youth suicides in the country, meaning the true toll is likely higher. But what is now known is the grisly impact that anti-trans restrictions have had on the youth of England. And in the United States, where a similar effort to restrict gender-affirming care is underway across the nation, it is likely that similar devastation is unfolding, unreported and silent.

The answer is that Paragraph 1 was pasted. Despite it covering the same topic as Paragraph 2, the language is dryer and less emotional. Paragraph 2 also reads as being more directed, with one sentence neatly flowing into the next.

Here are two examples from the article ‘[Colorado Supreme Court May Force Children's Hospital To Resume Trans Youth Care](https://www.erininthemorning.com/p/colorado-supreme-court-may-force?ref=theneedlenews.com)’.

**Paragraph 1:**

> On Tuesday, the Colorado Supreme Court heard oral arguments over whether Children's Hospital Colorado can be forced to resume gender-affirming care for transgender youth. The hospital was one of roughly 40 across the country that capitulated to Trump administration threats and shuttered their trans youth care programs. However, the hospital's position has grown increasingly untenable, as hospitals in states like Minnesota and California have begun reversing course and as the Trump administration has suffered mounting losses in federal courts—including an Oregon ruling that vacated the very declaration the hospital cited as justification for halting care. Hearing arguments on Tuesday, several justices appeared skeptical of the hospital's rationale, questioning whether Colorado's civil rights protections for transgender people—among the strongest in the nation—can simply be overridden by federal threats that do not constitute law.

**Paragraph 2:**

> Several justices appeared to share that concern. Justice William W. Hood III questioned whether the hospital’s catastrophic predictions were even realistic, wondering aloud whether the “catastrophic nature of it makes it politically untenable”—suggesting the federal government would never actually follow through on shutting down a major children’s hospital. It is worth noting that no hospital that has continued to provide gender-affirming care has had its Medicaid funding revoked.

Despite the use of an em dash in both samples, Paragraph 1 is pasted. It contains two run-on sentences in a row, both made longer by the inclusion of an em dash. A human editor would tell the author to split them up.

Lastly, look at two paragraphs from the article ‘[Its Time For Mayor Mamdani To Live Up To His Campaign Promises On Transgender Youth](https://www.erininthemorning.com/p/its-time-for-mayor-mamdani-to-live?ref=theneedlenews.com)’.

**Paragraph 1:**

> In January, the Mamdani administration stood up two new youth clinics at H+H Woodhull and Queens with $4 million from MetroPlusHealth, H+H’s own insurance plan. He has shown he can move fast when he wants to. The question is whether he will do the same for trans kids being told they have 30 days of medication left and nowhere to go.

**Paragraph 2:**

> Mamdani built his campaign on the backs of transgender people who worked hard to elect him. And to his credit, he ran one of the most refreshing campaigns towards transgender people I have ever seen from a major Democratic politician. We were one of the first to report on his support early on, when he appeared at a protest against NYU Langone over the very same hospital's pause on trans youth care. Throughout his campaign, he never backed down from his promise that he had transgender people's backs, even releasing a two-minute ad about the history of transgender people in New York City, set to the music of transgender artist SOPHIE, celebrating that history in a way that made clear he cared.

Despite the lack of any of the more obvious signs of AI writing, Paragraph 2 was the pasted one.

The key to every single test was finding the straight apostrophes and quote marks. You can find examples of the difference in the graphic above this test. If you find the straight marks in the text, you can tell what is pasted without any knowledge of AI writing style.

Reed’s writing is full of text with all the signs of AI writing, but mixed in with text with a human touch. Her style is not entirely that of AI, but one where the human and the LLM text sit alongside each other, pretending to be a singular voice.

### **6.2.3: Synthetic Corpus**

When testing which LLM model writes with which marks, *The Needle* performed two levels of analysis. First, we prompted the web interfaces for Gemini, Copilot, Grok, ChatGPT, Claude, and Perplexity and Answer, giving each LLM the exact same prompt. *(This was done in Sep. 2026\. Results may vary in the future.)*

In the test, the current-era versions of both Gemini and Claude produced straight marks 100% of the time.

#### Every AI Stuck With One Punctuation Style

A test of which LLM outputs which style of punctuation. 

Every AI Stuck With One Punctuation StyleSix rows, one per LLM, with one unit drawn for every quote mark the system set. Two outputs set only straight marks, 15 and 5 of them; four set only curly marks, 7, 7, 6, 2 of them. Across 42 marks in 1,744 words. no LLM used both styles.15 straightClaude15 of the set’s 42 marks5 straightGemini7 curlyGrok7 curlyCopilotthe only output withnon-breaking hyphens6 curlyChatGPT2 curlyPerplexityverdict rests on two marksStraight marks: U+0022, U+0027Curly marks: U+2018, U+2019, U+201C, U+201D 

The same instructions were given to 6 contemporary AI families. They were told to create an article with a headline, an opening paragraph, three additional paragraphs, and a quoted source.

This test shows what they each do now, but not what previous models did. Given one sample each, this test is not itself evidence. For defensible evidence, we had to test the models at scale.

To also test older models, and to do it at scale, we ran a [synthetic corpus](https://github.com/firasd/vibesbench/blob/main/docs/2026/A/the-curious-case-claudes-quotes.md?ref=theneedlenews.com). This corpus used outputs from versions of ChatGPT, Claude, and Gemini across several years. We had each model write a single news article using the same prompt as in the first test. However, we made one change: we supplied the headline from a randomized selection of Erin Reed’s own headlines, matched to that model’s year. We ran the same list of 50 headlines for each model relevant to that year.

Each of the following model versions were told to generate 1000 documents in this way.

- \-ChatGPT from 2023, 2024, 2025, and 2026
- \-Claude from 2024, 2025, and 2026
- \-Gemini from 2024, 2025, 2026

There are no 2023 models of Claude and Gemini that can still be run as of Sep. 2026\. 

#### Only ChatGPT can still be tested on its 2023 clone

__Every model that this method was tested on.__
| Family  | Vintage | Model                         | Documents | Tier                   |
| ------- | ------- | ----------------------------- | --------- | ---------------------- |
| ChatGPT | 2023    | openai/gpt-3.5-turbo          | 1,000     | Flagship               |
| ChatGPT | 2024    | openai/gpt-4o                 | 1,000     | Flagship               |
| ChatGPT | 2024    | openai/gpt-4o-mini            | 1,000     | Small tier, not pooled |
| ChatGPT | 2025    | openai/gpt-5                  | 1,000     | Flagship               |
| ChatGPT | 2026    | openai/gpt-5.6-terra          | 1,000     | Flagship               |
| Claude  | 2023    |                               |           | Not served             |
| Claude  | 2024    | anthropic/claude-3-haiku      | 1,000     | Small tier             |
| Claude  | 2025    | anthropic/claude-opus-4.5     | 860       | Flagship               |
| Claude  | 2026    | anthropic/claude-opus-5       | 1,000     | Flagship               |
| Gemini  | 2023    |                               |           | Not served             |
| Gemini  | 2024    | google/gemma-2-27b-it         | 998       | Open weights           |
| Gemini  | 2025    | google/gemini-2.5-pro         | 1,000     | Flagship               |
| Gemini  | 2026    | google/gemini-3.1-pro-preview | 1,000     | Flagship               |

Source: Generation run on Sep. 8, 2026, routed through OpenRouter. The ten models produced a total of 9,858 documents. In house analysis by The Needle.[\[1\]](#src-S1)

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

The test results showed Gemini using the straight punctuation less than Reed does. This largely rules out Gemini as the LLM Reed is using.

The same cannot be said for ChatGPT and Claude. As measured by our test, OpenAI flipped the punctuation produced by ChatGPT from straight to curly between the 2024 and 2025 models. Pasted writing from 2025 ChatGPT models wouldn’t show up on this test. 2025 is the same period where Reed’s paste marks largely disappear from her published work. 

Then in Feb. 2026, her paste marks turn back on like a light switch, and the 2026 test shows that Claude and Gemini both still output straight punctuation. However, Claude outputs straight punctuation 100% of the time, whereas Gemini outputs it about 10% less than Reed’s own published work. 

We can pin down this light-switch moment to a four day period between Feb. 5 and 9, 2026\. Feb. 5 is the day Claude Opus 4.6 launched, and Feb. 9 has her first article after that launch. This is when Reed began writing with Claude.

From then on, her typical output drastically changed. 

### **6.2.4: Paragraph Lengths**

When Reed doesn’t paste her text in, she writes paragraphs in the typical journalistic style: her paragraphs are typically 2 sentences long. This is standard for news coverage, and is part of the [AP Style](https://owl.purdue.edu/owl/subject%5Fspecific%5Fwriting/journalism%5Fand%5Fjournalistic%5Fwriting/ap%5Fstyle.html?ref=theneedlenews.com) taught to anyone learning journalism.

When she pastes her text in, her paragraphs jump to a median of 5 sentences long.

One of the first things you learn in journalism school is to keep your news paragraphs at 2 sentences or less, unless there is a pressing need to do otherwise. 

#### Reed’s Pasted Paragraphs are Longer than her Typed Paragraphs

Median sentences per paragraph under Erin Reed’s byline. Jul. 9, 2022 to Aug. 31, 2026.

Median sentences per paragraph, with and without paste markers Paragraphs carrying paste markers run a median of 5 sentences. Paragraphs without them run a median of 2, a gap of 3 sentences, across 11,474 paragraphs in 754 articles. 0 1 2 3 4 5 Sentences per paragraph Pasted Non-pasted 5 2 +3 sentences p = 1.4 × 10⁻³¹³ · Cliff’s delta = 0.54 (large) 

Paragraphs with paste markers (n = 1,824; middle half 3 to 6) Paragraphs without them (n = 9,650; middle half 1 to 4) 

Source: Erin Reed’s Substack archive, 754 articles. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

When Reed doesn’t paste, she observes this convention. When she pastes, she writes like an essayist.

It seems like her LLM ghostwriter doesn’t know what AP style is, or the conventions of writing a news story.

### **6.2.5: 2025 is an Outlier**

This method of measuring what text is pasted from an external source does not work for the year 2025\. This is not because she began writing the way she did before the release of ChatGPT, but because contemporary models of ChatGPT stopped using straight marks in their outputs. Therefore, 2025 is the one year where the absence of straight marks does not indicate an absence of pasted text.

We believe she was still using contemporary ChatGPT models during this time. *More on this in Section 6.4.*

### **6.2.6: 2026 is Off the Charts**

Since Feb. 9, 2026, the date of her first article after Claude Opus 4.6 released, and when her writing began resembling that of Claude, 65.7% of the words she published on Substack were in pasted paragraphs. 

#### The Rate of Straight Marks Spikes After Release of Claude Opus 4.6

This chart shows every day Reed published from January through August 2026\. Each number is the percent of the day’s published words that sit inside a pasted paragraph. January 2026 0.3% of 19,516 words · 13 days M T W T F S S 1 2 3 4 5 0% 6 0% 7 0% 8 0% 9 10 11 12 0% 13 14 15 0% 16 17 18 19 0% 20 0% 21 0% 22 0% 23 24 25 26 8.7% 27 28 0% 29 30 0% 31 February 2026 51.4% of 19,653 words · 15 days M T W T F S S 1 0% 2 3 4 5 0% 6 7 8 9 58.4% 10 11 88.9% 12 79.4% 13 99.2% 14 15 16 0% 17 74.9% 18 19 64.4% 20 0% 21 22 23 66.3% 24 43.9% 25 33.5% 26 83.6% 27 77.7% 28 March 2026 71.6% of 20,295 words · 18 days M T W T F S S 1 2 72.5% 3 74.2% 4 86.5% 5 19.7% 6 66.3% 7 8 0% 9 85.5% 10 28.9% 11 61.4% 12 99.1% 13 49.2% 14 15 16 17 85.1% 18 19 50.7% 20 21 22 23 79.1% 24 90.5% 25 26 27 99.2% 28 29 30 83.2% 31 81.2% April 2026 72.6% of 18,115 words · 15 days M T W T F S S 1 89.6% 2 90.4% 3 99.7% 4 5 6 7 36.4% 8 83.2% 9 98.5% 10 11 12 13 86.3% 14 15 78.7% 16 80.9% 17 18 19 20 52.7% 21 22 64.8% 23 99.8% 24 0% 25 26 27 28 29 99.6% 30 81.5% May 2026 70.4% of 12,858 words · 12 days M T W T F S S 1 67.5% 2 3 4 63.2% 5 6 70.5% 7 43.5% 8 93.2% 9 10 11 12 13 14 82.2% 15 16 17 18 0% 19 95.9% 20 21 89.1% 22 23 24 25 26 62.7% 27 57.0% 28 99.0% 29 30 31 June 2026 71.9% of 13,962 words · 14 days M T W T F S S 1 88.9% 2 64.8% 3 32.2% 4 89.3% 5 6 7 8 85.6% 9 37.5% 10 81.8% 11 12 13 14 15 98.7% 16 55.4% 17 18 84.2% 19 20 21 22 80.0% 23 24 85.3% 25 26 41.0% 27 28 29 30 89.4% July 2026 34.7% of 8,948 words · 8 days M T W T F S S 1 2 3 4 5 6 7 8 9 10 11 12 13 0% 14 15 16 91.0% 17 18 19 83.7% 20 21 74.6% 22 23 0% 24 25 26 27 0% 28 29 30 24.9% 31 0% August 2026 77.7% of 3,344 words · 4 days M T W T F S S 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 62.7% 20 21 22 23 24 25 68.5% 26 27 81.4% 28 29 30 31 98.6% Share of that day’s words inside of a pasted paragraph 0% 20% 40% 60% 80% 100% 

Day without an articleFeb. 9, 2026: the first day pasting is (again) obviousFeb. 5, 2026: Claude Opus 4.6 released

A paragraph counts as pasted when it carries any straight quotation marks or apostrophes, whatever else is in it. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

By contrast, from the beginning of the year to the release of Claude Opus 4.6, of the 642 paragraphs she published, only one is obviously pasted. 

Since the release of Claude Opus 4.6, ten separate articles had at least 99% of their words pasted from an external source. This started with the fourth [article](https://www.erininthemorning.com/p/european-parliament-votes-overwhelmingly?ref=theneedlenews.com) of this time period, eight days after the new model release.

Since then, pasted sentences are on average 4.28 grade levels harder to read than her non-pasted sentences. 

The average reading level of a typed sentence is 12.55, slightly above the level expected of a high school senior. The average reading level of a pasted sentence in this period is 16.83, a level of difficulty only expected of academics talking to other academics. 

#### Sentences From an External Source Are Over 4 Grade Levels Harder to Read

Flesch-Kincaid grade-levels, compared between sentences with and without straight apostrophes/quotation marks. Feb. 9 to Aug. 31, 2026.

Average FK grade for pasted vs non-pasted sentences Pasted sentences have an average FK grade of 16.83 vs. 12.55 for non-pasted, a difference of 4.28 grade levels. p = 5.6 times 10 to the negative 69. High-schoolCollege standardGraduate / academic papers 0 5 10 15 Flesch-Kinciad Grade Level Pasted Non-pasted 16.83 12.55 +4.28 p = 5.6 × 10⁻⁶⁹ Cohen’s d = 0.64 (medium) 

Sentences with paste markers (n = 1,137) Sentences without paste markers (n = 2,989) 

Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

To consistently write at such a difficult to understand grade level implies bad writing, and/or a lack of respect for your audience. *More on this in Section 7.1.2.*

Starting Feb. 9, sentences with pasted content are 73.7% longer on average than sentences without them. Typed sentences are on average 15.88 words long, whereas pasted sentences are 27.58 words long on average.

#### In 2026, Pasted Sentences Are Over 70% Longer Than Typed Sentences

Sentence-length comparison between paragraphs with and without paste markers in Erin Reed’s writing. Feb. 9 to Aug. 31, 2026.

Average sentence length for pasted vs non-pasted paragraphs Pasted paragraphs average 27.58 words per sentence vs. 15.88 for non-pasted sentences, with a difference of 11.70 words. p = 7.5 times 10 to the negative 52. 0 10 20 Words per sentence Pasted Non-pasted 27.58 15.88 +73.7% p = 7.5 × 10⁻⁵² Cohen’s d = 1.14 (large) 

Sentences with paste markers (n = 463) Sentences without paste markers (n = 517) 

Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

One possible source of straight quotes are court records and other official sources that she’s pasting in to quote from. However, even assuming that all official sources were written with straight marks (they aren’t), that alone would not explain why the pasted text is so different from the typed text.

### **6.2.7: The 99%+ Pasted Articles**

In 2026, she published 10 articles which were at least 99% pasted. These are as follows:

#### The Ten Most Heavily Pasted Articles of 2026

__The 2026 Erin Reed articles with the largest percent of their words inside of pasted paragraphs. Columns: share pasted, date published, and the article, whose title links to the original.__
| Pasted | Published      | Article                                                                                                                                                                                                    |
| ------ | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| 99.8%  | March 3, 2026  | [AG Letitia James Just Stood Up For Trans Youth In New York. Now, Will Mayor Mamdani Act?](https://www.erininthemorning.com/p/ag-letitia-james-just-stood-up-for?ref=theneedlenews.com)                    |
| 99.8%  | April 23, 2026 | [Texas Tech Issues Ban On Students Writing On LGBTQ+ Topics](https://www.erininthemorning.com/p/texas-tech-issues-ban-on-students?ref=theneedlenews.com)                                                   |
| 99.7%  | Feb. 27, 2026  | [ACLU Files Lawsuit Against Kansas Trans Drivers License Revocations, Bathroom Bounty Hunter Law](https://www.erininthemorning.com/p/aclu-files-lawsuit-against-kansas?ref=theneedlenews.com)              |
| 99.7%  | April 3, 2026  | [Demonstrators Arrested After Sit-In In Idaho Protesting Extreme Felony Trans Bathroom Ban](https://www.erininthemorning.com/p/demonstrators-arrested-after-sit?ref=theneedlenews.com)                     |
| 99.6%  | April 29, 2026 | [New Colorado Conversion Therapy Ban With Clever Mechanism Close To Passing](https://www.erininthemorning.com/p/new-colorado-conversion-therapy-ban?ref=theneedlenews.com)                                 |
| 99.3%  | Feb. 17, 2026  | [“You Outlaw It”: Heritage Foundation President Announces Intent To Outlaw All Trans Adult Care](https://www.erininthemorning.com/p/you-outlaw-it-heritage-foundation?ref=theneedlenews.com)               |
| 99.2%  | Feb. 13, 2026  | [European Parliament Votes Overwhelmingly For “The Full Recognition Of Trans Women As Women”](https://www.erininthemorning.com/p/european-parliament-votes-overwhelmingly?ref=theneedlenews.com)           |
| 99.2%  | March 27, 2026 | [Idaho Passes Most Extreme Bathroom Ban In Nation, Creating Path To Life In Prison](https://www.erininthemorning.com/p/idaho-passes-most-extreme-bathroom?ref=theneedlenews.com)                           |
| 99.1%  | March 12, 2026 | [Florida Passes Bill That Lets DeSantis Remove Local Elected Officials From Office For “Promoting” Pride](https://www.erininthemorning.com/p/florida-passes-bill-that-lets-desantis?ref=theneedlenews.com) |
| 99.0%  | May 28, 2026   | [New York State Budget Released: $0 For Transgender Care Access](https://www.erininthemorning.com/p/new-york-state-budget-released-0?ref=theneedlenews.com)                                                |

A paragraph counts as pasted when it carries any straight quotation mark or apostrophe, whatever else is in it. One row per article, ranked by share; ties broken by date. Titles link to the original article. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

It is notable that the articles have no shared theme. These 10 articles cover topics as different as state legislation, politician coverage, national coverage, and international news.

The most extreme example of this is the article ‘[AG Letitia James Just Stood Up For Trans Youth In New York. Now, Will Mayor Mamdani Act?](https://www.erininthemorning.com/p/ag-letitia-james-just-stood-up-for?ref=theneedlenews.com)’. This article has 99.8% of its words inside fully pasted paragraphs. 

There are only two words in the article that show evidence of a human touch. The thumbnail has a caption reading “*AG’s Office.*” The apostrophe after “*AG*” is curly, meaning it was probably typed. 

![](https://storage.ghost.io/c/07/58/0758f73b-8a0d-4fa1-a055-4d836e2c23c5/content/images/2026/09/data-src-image-2677f33e-64b4-4e65-831b-4c84e5d24ab5.png)

AG's Office

That’s it.

#### The Rate of Straight Marks Spikes After Release of Claude Opus 4.6

0% 25% 50% 75% 100% Jan Feb Mar Apr May Jun Jul Aug 2026 Share of the article’s words in a pasted paragraph Feb. 9, 2026 

One articleFeb. 9, 2026

106 Articles 

Published under Reed’s byline in 2026

1 of 15 

Articles before Feb. 9, 2026 carrying any pasted text

82 of 91 

Articles from Feb. 9, 2026 onwards carrying any pasted text

79.1% median 

Median share of an article’s words inside a pasted paragraph from Feb. 9 onwards

A paragraph counts as pasted when it carries any straight quotation mark or apostrophe, whatever else is in it. One dot per article; articles published the same day sit on the same vertical. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

### **6.2.8: The One**

Across the 754 articles that Erin Reed wrote alone at EITM, only one of them shows 100% of the writing being pasted from an external source: her [July 2023 announcement](https://www.erininthemorning.com/p/erin-in-the-morning-to-return-monday?ref=theneedlenews.com) that she was returning from vacation with her then-girlfriend and [now-wife](https://www.erininthemorning.com/p/our-wedding-from-the-woman-who-loves?ref=theneedlenews.com), Montana State Representative Zooey Zephyr.

Think of various examples of AI slop you’ve seen around, especially for blogs and other similar posts. Then read the opening paragraphs from this announcement:

> I'm thrilled to share with the faithful readers of 'Erin In The Morning' that our regular reporting will be back on track come Monday. There are so many stories I need to bring to the forefront, inspired by the numerous updates you've shared with me during my vacation. I'm deeply appreciative of my readers and the invaluable information you provide. I'm equally thankful for the robust support you've given while I’ve been gone. Together, we've created a powerful platform here - a means to push essential stories into mainstream media and ensure trans voices resonate in LGBTQ+ news reporting.  
>  
> The past week has been an extraordinary journey for my partner Zooey and me. My first ever overseas trip, and France couldn't have been a more magnificent destination. We began in Paris, specifically the LGBTQ+ hub of Le Marais district, and stumbled upon a delightful queer bar where we met some of our Parisian readers. Our adventure then took us northwest to St. Malo and Mont St. Michel, where we caught the international space station gliding across a star-lit sky while standing on the ramparts. Our journey is now drawing to a close with a visit to one of the largest renaissance festivals in Europe, set in an actual castle town. It's been a trip of dreams. 

If you don’t see the tells, that’s okay. If you do, it becomes a lot easier to see those tells in the rest of her work.

### **6.2.9: Reed vs. Baum**

To find a comparable writer, we ran similar tests on S. Baum, the only other full-time writer at EITM. 

In 2025, most (81.1%) of Baum’s straight marks were inside of quoted sources. In 2026, Baum had no straight marks at all. 

By contrast, in 2026, most of Reed’s straight marks are found in writing which she presents as coming from her own voice. 

While some pasted marks are from acknowledged sources, these are in the minority. Only 25.6% of Reed’s straight marks are explained by quoted paragraphs. The rest have no clear source.

*The evidence we used to detect AI in Reed’s writing does not hold for Baum’s writing. Therefore, we have no evidence that Baum’s writing at EITM is written with (or influenced by) an LLM.*

*S. Baum did not respond to five separate requests for comment.*

### **6.2.10: This is Not a Coincidence**

Reed's pasted content is consistently harder to read and uses more words than the stuff we can't prove she pasted. 

If Reed were pasting her own writing from an app that defaults to straight marks (such as the Notepad apps of Windows and iOS), the differences in the writing style would not be this drastic.

If the rise in straight quotation marks in Feb. 2026 were caused by Reed changing the default version in her word processor, then the rate of straight quotation marks would shoot up to 100% of all quote marks, or slightly below that figure. 

The difference between the two isn't explained by her occasionally quoting courts and legislation. We controlled for that when doing our calculations. The difference is in the difference between two sources of writing, both of which she attempts to pass off as her own voice.

### **6.3: Mixed-Human-AI Writing**

*The Needle* ran tests on Erin Reed’s writing to look for signs of mixed writing between a human and an AI author.

Since AI detectors are known to be [very unreliable](https://lawlibguides.sandiego.edu/c.php?g=1443311&p=10721367&ref=theneedlenews.com), you might be wondering, ‘Can this premise even be tested?’

According to the scientific papers we reviewed, the short answer is yes: the tells of mixed human-AI can be tested.

The longer answer is that nothing on its own is a smoking gun, but there are several tells that can be combined with one another to get an answer. Most of these are in the patterns of writing across large bodies of words. Since we are analyzing almost three-quarters of a million words under Reed’s byline, this is far more than enough.

For evaluating mixes of human and AI writing, some of the papers we reviewed showed that function words (words that serve specific grammatical functions in writing) can indicate mixed human-AI prose. 

The first type of word to change usage are [demonstratives](https://www.ef.com/wwen/english-resources/english-grammar/demonstratives/?ref=theneedlenews.com), which point at shared context between the author and the reader. Reed’s use of demonstratives such as ‘this’ and ‘these’ fell, while her use of ‘the’ rose. [Multiple](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0335369&ref=theneedlenews.com)[ scientific](https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2026.1921847/full?ref=theneedlenews.com)[ papers](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0288453&ref=theneedlenews.com) were able to identify similar trends, among other stylometrics, to point at mixed prose using human and AI writing, one of which did so with 98.1% accuracy. 

In short, the way she drops some demonstratives and raises others is a huge tell that Reed is mixing AI into her writing process.

Another study [claims](https://aclanthology.org/2020.emnlp-main.673/?ref=theneedlenews.com) that even the best-trained LLMs still have stylistic differences that can be assessed.

Over the course of Reed’s writing, the word ‘would’ fell, while ‘or’ and ‘but’ increased. This is another sign.

#### Using Function Words as Tells of Mixed Human-AI Writing

The drift here shows a mix of human and AI writing under a single byline.

Change in the rate of eight function words in Erin Reed’s writing, 2022 compared with 2026.Eight function words shift between 2022 and 2026\. Among demonstratives, “this” falls 9.7 per 1,000 words and “these” falls 2.6, while “the” rises 16.9\. Among depersonalisation markers, “its” rises 1.4 and “their” falls 0.4\. Among modals and connectors, “would” falls 4.1 while “but” rises 0.3 and “or” rises 1.9\. Each shift is labelled with its value. Four are significant at p below 0.001 and two more at p below 0.01; the shifts for “their” and “but” are drawn as hollow dots because they are not significant. \-10 \-5 0 +5 +10 +15 +20 Demonstratives this \-9.7 these \-2.6 the +16.9 Depersonalisation its +1.4 their \-0.4 Modals & Connectors would \-4.1 but +0.3 or +1.9 no change ← used less often used more often → 

Change in rate per 1,000 words, 2022 to 2026\. Four of the eight shifts are significant at p < 0.001 (Wilcoxon rank-sum); “would” and “or” are p < 0.01; “their” (p = 0.36) and “but” (p = 0.18), the hollow dots, are not significant. Pre: 2022, Jul. 9 to Dec. 31\. Post: 2026, Jan. 5 to Aug. 31\. n = 46 pre, 106 post. 2023 to 2025 excluded. Hollow dots show words whose changes are not statistically significant. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

According to [one paper](https://www.nature.com/articles/s41599-025-05986-3?ref=theneedlenews.com), a function word shift like this is listed among the “*clear and consistent stylistic*” signs of LLM-written text. 

Her text also depersonalizes herself more often than is common of human writers. Some of this depersonalization is expected in a mainstream journalistic style. However, this goes beyond normal journalistic standards.

Each of these shifts is statistically significant as well.

#### Statistical Significance of Mixed Human-AI Prose

These measures are reported in p-values.

| Word  | Category            | 2022, per 1,000 | 2026, per 1,000 | Change | p           |
| ----- | ------------------- | --------------- | --------------- | ------ | ----------- |
| this  | Demonstratives      | 14.45           | 4.74            | \-9.72 | 1.8 × 10−14 |
| these | Demonstratives      | 3.99            | 1.42            | \-2.57 | 7.3 × 10−7  |
| the   | Demonstratives      | 38.84           | 55.79           | +16.95 | 7.3 × 10−11 |
| its   | Depersonalisation   | 1.31            | 2.73            | +1.42  | 3.8 × 10−4  |
| their | Depersonalisation   | 5.31            | 4.95            | \-0.37 | 3.6 × 10−1  |
| would | Modals & Connectors | 7.06            | 2.93            | \-4.13 | 2.0 × 10−3  |
| but   | Modals & Connectors | 2.27            | 2.55            | +0.28  | 1.8 × 10−1  |
| or    | Modals & Connectors | 2.50            | 4.35            | +1.85  | 3.9 × 10−3  |

Wilcoxon rank-sum, two-sided. n = 46 pre (2022), 106 post (2026). 2023 to 2025 excluded. A p at or above 0.05 is not significant. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

LLMs don’t understand shared context in the way that humans do. They focus on mathematical representations, and translate them into text. To them, everything is just an object represented in the form of code and statistics. 

An LLM has a default writing style that humans read as lacking personality. That’s because LLMs don’t have personalities in the same ways that humans do. The difference between human and AI personalities is often obvious when looking at writing from the two side by side. 

That’s how the untrained eye is often able to detect that something is ‘off’ about LLM-written text, even if they don't know what is ‘off’ about it.

### **6.4: Model Drift**

The patterns in Reed's writing indicate that she began using ChatGPT in her writing no later than Spring 2023, and that she switched to Claude no more than four days after the release of Claude Opus 4.6.

This section looks at the flip between the two models using a different test, made by the same authors of the test detailed above. 

Our ‘pasted content test' pointed to her using Claude in 2026, and using ChatGPT in 2023-24, both with near-certainty. However, it couldn’t determine what model she was using in 2025.

In this latest test, *The Needle* replicated a peer-reviewed, scientifically validated[ method](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0335369&ref=theneedlenews.com) that was able to distinguish fully human writing from fully AI writing with reported accuracy of about 99.8%. This method was validated on short-form texts in the Japanese language. However, the actual tests are transferable to English language writing.

*The Needle* took our own writing across three groups as a baseline, since we know our work to be 100% human. Those three groups were Artemis’ own byline, Jane’s own byline, and their joint byline (in either order). 

We then had Claude Opus 5 rewrite each of Erin Reed’s articles for the 2025-26 period in two different ways. *Using AI in this way is required for the test to work, and there are no alternatives to replicate this method.*

First, we had Claude write articles while telling it not to sound like Erin Reed. Second, we gave it Erin Reed’s headlines and word counts, and told it to write an article based on nothing else. We used each as the 100% machine-written baseline, which we need for this test. 

We then trained a classical topic model (an old-school machine learning model) on each of the groups in the known-human and known-machine texts, and compared the results of both sets to Reed’s actual writing.

It spit out a score and a distribution for each group it measured. What we found was that Erin’s writing was inching closer to that of Claude Opus 5 over time. Dec. 2022 to Dec. 2023, 2024, 2025 to Feb. 5, 2026, and Feb. 9, 2026 onward were each a little closer to the way Claude currently writes than the preceding era was.

#### Erin Reed’s Writing Grows Closer to That of an LLM Over Time

Detector score per article for each group the model measured 0.00 0.25 0.50 0.75 1.00 Detector score per article (0 = human bylines, 1 = Claude Opus 5) Average Reed, before ChatGPT n = 17 0.228 Reed, 2022-23 n = 248 0.267 Reed, 2024 n = 195 0.284 Reed, 2025-26 n = 274 0.302 Claude Opus 5, paraphrase n = 52 0.605 Claude Opus 5, headline n = 52 0.747 The Needle News, Brigham n = 68 0.181 The Needle News, Douglas n = 40 0.214 Joint Brigham/Douglas Byline n = 37 0.177 

Erin Reed, by period Claude Opus 5 reference texts The Needle News, control bylines 

19 of her 2025-26 articles score above Claude’s paraphrase median. 18 of those are from Feb. 2026 onwards. Source: Erin Reed’s corpus, a synthetic corpus, and The Needle’s own corpus. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

This means that her writing has been getting closer to that of LLMs over time. 

The strongest evidence in favor of Reed not using AI is from the same six-phased test that shows her writing drifts towards Claude, and this 17 model benchmark can only reliably tell whether something is 100% AI if the writing in question is actually 100% written by AI. Since we know that Reed mixes the LLM with her own writing, this cannot be used to exonerate her.

This test was run to be as charitable to Reed as possible. We wanted to be wrong in this investigation, and comparing her writing to 100% machine writing across 17 different models was a way of testing the most extreme possible outcome: that she is using 100% unedited machine writing in what she publishes. We can prove this is not the case (with the exception of a single article from 2023).

Either way, the cumulative evidence shows her writing drifting over time towards the style of the machines. 

It’s also concerning that 2 of her articles in 2025-26 are more noticeably like the writing of Claude Opus 5 than anything we got Claude to write as part of this test. Worse still, 21 of her articles in the same period read as being more likely to have been written by Claude than the average of the articles we got Claude to write. All 21 of these occur after she switched to Claude.

#### Model Pull by Article: 2023 to 2024

Writing-style lean, pasted vs typed Reed articles, 2023 to 2024, tested against Claude Haiku (2023) and GPT-4 (2023). On a scale between Claude vs ChatGPT’s GPT-4, Reed’s writing mostly sits as a cloud just left of the no-lean middle point. This means her writing in this era was slightly more like GPT-4. more like GPT-4 ← → more like Claude no lean \-0.5\-0.4\-0.3\-0.2\-0.10+0.1+0.2+0.3+0.4+0.5 Pasted n=18 Typed n=193 median −0.093 median −0.081 

Value = P(LLM | Claude anchor) - P(LLM | GPT-4 anchor).

Positive = leans toward Claude. Each dot is one Reed article. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

Her writing from 2023 to 2025 aligned more closely with the era-specific version of ChatGPT than that of Claude.

However, starting in 2026, her writing begins to resemble the writing of Claude more closely than it resembles the writing of ChatGPT.

#### Erin Reed’s 2026 Writing Resembles That of Claude

Writing-style lean, pasted vs typed Reed articles, Feb to August 2026, tested against Claude Opus 4.5 (2026) and GPT-5.2 (2026). On a scale between Claude Opus 5 vs ChatGPT-5.2, Reed’s writing mostly sits as a cloud just right of the no-lean middle point. This means her writing in this era was slightly more like Claude. more like GPT-5.2 ← → more like Claude Opus 5 no lean \-0.5\-0.4\-0.3\-0.2\-0.10+0.1+0.2+0.3+0.4+0.5 Pasted n=39 Typed n=24 median +0.136 median +0.157 

Value = P(LLM | Claude anchor) - P(LLM | GPT-5.2 anchor). Positive = leans toward Claude. Each dot is one Reed article. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

In other words, this is more evidence that she switched models from ChatGPT to Claude, and that this change can be seen in her aggregate writing style. 

### **6.5: Em Dashes**

As LLMs have grown more prominent in online writing, so too have their stylistic quirks been increasingly evident in Erin Reed’s writing. This can be seen in the sharp rise of em dashes in her work, and the fact that their rise coincides with the releases of ChatGPT and various models of Claude.

This rise is coupled with a drop in Erin Reed’s own pre-ChatGPT writing style of space-hyphen-space. This is when she puts a regular space, a regular hyphen, and a regular space in that order in her writing.

[Em dashes](https://www.merriam-webster.com/grammar/em-dash-en-dash-how-to-use?ref=theneedlenews.com) are a grammatical tool to separate different parts of a sentence, not unlike a comma. While they have been used for centuries, they have recently been adopted by LLMs, and their use has become closely associated with LLM-generated writing in general. 

Here is an example of a dash from her [earliest risk map](https://www.erininthemorning.com/p/new-map-anti-transgender-legislative?ref=theneedlenews.com) on July 18, 2022:

> The map overlays well with the abortion ban maps and this should not be surprising - the two issues are very closely linked and fought for by the same organizations. 

Pay attention to the dash between the two clauses, and how there are spaces on either side of it. This is clearly different from a [more recent article](https://www.erininthemorning.com/p/trans-sports-ban-with-genital-inspections?ref=theneedlenews.com) with an actual formatted em dash:

> On January 3, 2026, the far-right political group Let's Go Washington announced that it had collected more than 400,000 signatures—enough to place a transgender sports ban for high school students on the ballot with genital inspections as a primary verification method.

Erin herself has [said that](https://bsky.app/profile/erininthemorning.com/post/3ls2bjqrol22e?ref=theneedlenews.com) “*I use Em dashes a lot and always have*”. However, her early work for EITM used regular dashes in space-hyphen-space format instead, while also using these less often than she now uses em dashes.

It is possible that Reed doesn't know the difference between a dash and an em dash. However, that wouldn't explain why dashes declined so rapidly in her writing, and why it took so long for em dashes to replace them.

The presence of em dashes in a text is often used as evidence that the text was AI-generated. We should be clear that the presence of em dashes alone does not prove that an LLM made a text. Some writers simply like to use them, and that is fine.

However, that doesn't explain why the em dashes got more common at the same time that ChatGPT began using them excessively.

### **6.5.1: The Rate of Em Dashes**

However, it is undeniable that all major LLMs used em dashes far more commonly than human writers; [until July](https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing?ref=theneedlenews.com) of this year. [OpenAI CEO Sam Altman](https://podscripts.co/podcasts/this-past-weekend/599-sam-altman?ref=theneedlenews.com) admitted it, and the [scientific](https://arxiv.org/abs/2606.29540?ref=theneedlenews.com) and [other](https://msukhareva.substack.com/p/the-mystery-of-emdashes-part-two) literature backs it up. As [more online text](https://www.pewresearch.org/data-labs/2026/08/20/how-much-of-the-internet-is-written-with-ai/?ref=theneedlenews.com) is being written by LLMs, the amount of em dashes on the internet has increased as a result.

Erin Reed's writing at EITM is an excellent way to track this change. She began writing before LLMs were commercially available, and has continued to publish as LLMs have grown in popularity.

This change can be seen in the rise of em dashes in her writing.

In 2022, before the release of the first commercially available LLM (ChatGPT), and in the initial month after its release, her articles did not contain a single em dash. This means that Reed wrote for six months without ever feeling compelled to use one.

#### Erin Reed’s Em Dash Rate

Em‑dashes per 1,000 words. July 2022 to Aug. 2026 02468101220222023202420252026Em dashes per 1,000 wordsMonth of publicationNo data ChatGPT released (Nov 30, 2022)Claude Opus 4.6 released (Feb. 5, 2026) 

Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

Instead, she used the space-hyphen-space ‘ - ’ convention. However, she used this far less than she currently uses em dashes ‘—’. In Jan. 2023, she used the first em dash of her career at EITM. 

In some months, her em dash usage has decreased, but it has never gone below the baseline from Summer 2024.

#### Dashes Fall, While Em Dashes Rise

Spaced hyphens and em dashes per 1,000 words. July 2022 to Aug. 2026 Reed’s spaced hyphen rate runs between 2 and 9 per 1,000 words through 2022, falls away over the first half of 2023, and sits at or near zero from 2025 onward. Her em dash rate is zero until 2023, stays under 2 per 1,000 words through 2024, then jumps in February 2025, and holds between 4 and 12 for the rest of the period. The two turning points are around two years apart. Both lines fall to 0.3 in the final month, August 2026; but Reed published only 4 articles and 3,346 words that month, with no publication at all from Aug. 1 to Aug. 18, so that last point is too thin to read as a trend. 02468101220222023202420252026Dashes per 1,000 wordsMonth of publicationOnly 4 Articles 

Spaced hyphen ( - ) Em dash (—) 

Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

At her peak rate in October 2025, Reed used 12.05 em dashes per 1,000 words. This is one for every 83 words.

From Jan. 2025 to the end of Jan. 2026, all but three of her articles have at least one em dash. For contrast, her first six months at EITM have none.

Put the two eras of 2022-23 and 2025-26 side by side. If you draw one article at random from each side, the article from the 2025-26 side will have a higher rate of em dashes 95.7% of the time.

#### Percent of Erin Reed’s Articles Without Em Dashes

Percent of Erin Reed’s articles with zero em dashes.Most of Reed’s Articles in 2023 had no em dashes at all. None of her articles in 2022 had them either. Her share of articles without em dashes cratered to zero by Feb. 2025, and has barely changed since.0%20%40%60%80%100%2022-07: 5 of 5 articles (100%)2022-08: 3 of 3 articles (100%)2022-09: 2 of 2 articles (100%)2022-10: 1 of 1 articles (100%)2022-11: 11 of 11 articles (100%)2022-12: 24 of 24 articles (100%)2023-01: 26 of 27 articles (96%)2023-02: 25 of 27 articles (93%)2023-03: 26 of 26 articles (100%)2023-04: 19 of 23 articles (83%)2023-05: 16 of 18 articles (89%)2023-06: 14 of 18 articles (78%)2023-07: 8 of 15 articles (53%)2023-08: 9 of 17 articles (53%)2023-09: 5 of 15 articles (33%)2023-10: 6 of 17 articles (35%)2023-11: 10 of 13 articles (77%)2023-12: 13 of 19 articles (68%)2024-01: 20 of 21 articles (95%)2024-02: 13 of 19 articles (68%)2024-03: 10 of 19 articles (53%)2024-04: 7 of 16 articles (44%)2024-05: 17 of 18 articles (94%)2024-06: 5 of 12 articles (42%)2024-07: 6 of 14 articles (43%)2024-08: 7 of 14 articles (50%)2024-09: 2 of 12 articles (17%)2024-10: 6 of 19 articles (32%)2024-11: 1 of 19 articles (5%)2024-12: 4 of 13 articles (31%)2025-01: 3 of 14 articles (21%)2026-03: 1 of 21 articles (5%)2026-05: 1 of 12 articles (8%)2026-08: 3 of 4 articles (75%)Jul 2022Jan 2023Jan 2024Jan 2025Jan 2026Jul 2026 

In her first six months, she never used an em dash. In the twelve months after Feb. 2025, she never published anything without an em dash. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

### **6.5.2: Em Dashes in Claude’s Writing**

As of late July, em dashes are no longer a reliable tell for AI writing from most new models, as they are trained to use them less frequently. The most notable exception is Claude, which all evidence indicates she is currently using.

Recent editions of ChatGPT [have tried](https://techcrunch.com/2025/11/14/openai-says-its-fixed-chatgpts-em-dash-problem/?ref=theneedlenews.com) to reduce the amount of em dashes in their outputs, as the public now sees them as a surefire way of identifying if a text was AI-generated.

In November 2025, Sam Altman announced that users of ChatGPT would be able to successfully prompt it to not use em dashes. Altman even [tweeted](https://x.com/sama/status/1989193813043069219?ref=theneedlenews.com) this out as a "*small-but-happy win*". Newer models of ChatGPT now use em dashes at a lower rate than humans.

Similar changes are true of other models. ChatGPT, Gemini, and Grok have all reduced their use of em dashes, according to [*The Economist*](https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing?ref=theneedlenews.com).

The notable exception to this trend are current Claude models, which have made no similar changes to their use of em dashes.

If she were using models other than Claude, we would expect her use of em dashes to crater or stop completely over the course of 2026\. In reality, her average usage of em dashes is roughly what it was last year.

This is evidence that she is using Claude as her LLM of choice.

There is also evidence that Reed knows that large sections of the public view em dashes as a sign of LLM writing, even in the absence of corroborating evidence. 

Just this month, Reed defended Former President Barack Obama from accusations that he wrote [a post](https://bsky.app/profile/erininthemorning.com/post/3mvjjillw2223?ref=theneedlenews.com) about AI regulation with an LLM, rather than his own voice. The only evidence that he used AI were the em dashes in his writing. Replying to a [post saying](https://bsky.app/profile/quarks.bsky.social/post/3mvjjmw3mik2c?ref=theneedlenews.com) “*That em dash makes me suspicious*”, [Reed responded](https://bsky.app/profile/erininthemorning.com/post/3mvjk5sfsms2o?ref=theneedlenews.com) “*NO! Em dashes are good! :(* ”

While em dashes are sometimes used by humans, the ways that Reed has used them over the course of her career are not consistent with fully-human writing.

### **6.5.3: Em Dashes in the Future**

On Sep. 22, Anthropic released [Claude Opus 5.5](https://platform.claude.com/docs/en/models/opus-5-5/overview?ref=theneedlenews.com). According to [one source](https://x.com/DevaiahShrithan/status/2102767477310844980?ref=theneedlenews.com) who used it, it does not use em dashes by default. 

Assuming this finding holds, we can predict how Reed's writing will use em dashes in the future in different scenarios.

1. \-If Reed switches to Claude Opus 5.5, we would expect her use of em dashes to crater or stop completely.
2. \-If she switches models away from Claude, we would expect her use of em dashes to crater or stop completely.
3. \-If Reed were to stick with the Claude model she was using in Aug. 2026, we would expect her em dash usage to remain within the range it has been in since the beginning of 2025.
4. \-If Reed stops using AI in her journalism in any capacity, then assuming she fully returns to her writing style from before commercially available LLMs, we would expect her to never use em dashes again.

### **6.6: Authorial Drift**

*The Needle* used [a test](https://forensicspot.com/glossary/burrowss-delta?ref=theneedlenews.com) for telling apart the writing of separate authors. That test could separate Reed’s 2022-23 articles from her 2025-26 articles 93.5% of the time.

Reed’s current writing is difficult to identify as being from the same author as her early work at EITM. In other words, her writing has changed so dramatically that they don't read as being from the same person.

This kind of method, called stylometrics, is used by journalists and courts in disputes over authorship. It is also used to identify anonymous writing. 

In journalism, this type of stylometrics is how the *Sunday Times* [figured out](https://www.scientificamerican.com/article/how-a-computer-program-helped-show-jk-rowling-write-a-cuckoos-calling/?ref=theneedlenews.com) J.K. Rowling was [writing under](https://www.npr.org/2013/07/26/205794448/uncovering-the-mystery-of-j-k-rowlings-latest-novel?ref=theneedlenews.com) the mannish pseudonym of Robert Galbraith. As for language forensics, the [highest profile](https://www.npr.org/2017/08/22/545122205/fbi-profiler-says-linguistic-work-was-pivotal-in-capture-of-unabomber?ref=theneedlenews.com) case was how [The Unabomber](https://theconversation.com/how-the-unabombers-unique-linguistic-fingerprints-led-to-his-capture-207681?ref=theneedlenews.com) was identified based on his writing style.

Using a [classical method](https://www.researchgate.net/publication/240956478%5F'Delta'%5FA%5FMeasure%5Fof%5FStylistic%5FDifference%5Fand%5Fa%5FGuide%5Fto%5FLikely%5FAuthorship?ref=theneedlenews.com) to do a similar form of [stylometrics](https://www.academia.edu/21241069/Towards%5Fa%5Fbetter%5Funderstanding%5Fof%5FBurrowss%5FDelta%5Fin%5Fliterary%5Fauthorship%5Fattribution?ref=theneedlenews.com), the same techniques would not be able to tell that Erin Reed is the author of both her 2022-23 and 2025-26 articles.

This method can also measure how far apart different writers are from one another. For example, Reed in 2025-26 is farther away from herself in 2022 than she is from her own employee today. 

#### Erin Reed Today is More Similar to Other Writers on the Same Beat Than She is to Her Own Writing from 2022

Five periods of Reed's writing. Every line carries its own Burrows’ Delta. The dots are sized by how many pieces are in each group.

ERIN IN THE MORNING 0.278 0.171 0.216 0.178 0.241 0.265 Reed pre-ChatGPT 19 pieces Reed Nov. 2022 to 2023 252 pieces Reed 2024 196 pieces Reed 2025 to Feb. 5, 2026 186 pieces Reed after Feb. 9, 2026 91 pieces Baum 206 pieces Lazine 30 pieces MEASURED FROM REED AFTER FEB. 9, 2026 to Baum, her staff writer 0.241 to Lazine, her staff writer 0.265 HER OWN DRIFT, FOR SCALE to Reed pre-ChatGPT 0.413 0.403 that group’s own noise floor, n = 19 0.100 DELTA Anything narrower than 0.100 Delta is basically the same author fingerprint. 

Reed, earliest periodReed, later periodsBaumLazine

0.413 Delta 

Reed pre-ChatGPT to Reed after Feb. 9, 2026

0.241 Delta 

Reed after Feb. 9, 2026 to Baum

0.178 Delta 

The February 2026 switch itself

97.6% of the way 

Already travelled by Feb. 5, 2026

A node and edge map. The five periods of Reed's writing run as a trail from the lower left to the upper middle, each step labelled with its Burrows’ Delta. Positions are arranged for legibility only. Every distance is the number printed on its line. These centroid distances do not flatten into two dimensions, so the map is not drawn to scale and carries no ruler. The strip below is one axis from one origin, and is drawn to scale. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

Reed was a measurably different author in each period we looked at: before the release of ChatGPT, from Dec. 2022 to 2023, during 2024, 2025 until Feb. 5, 2026, and Feb 9., 2026 until the end of our data collection. Each subsequent era is increasingly different from her pre-ChatGPT self. Despite being under the same byline, the stylometrics cannot identify all of these time periods as belonging to the same author.

It is clear that Reed’s journalism has little authorial voice which stays consistent between the different eras listed. It is also notable that she has been drifting in the same direction, i.e. further away from her pre-GPT self, in every subsequent period.

This is not a subjective claim. In each era, Reed writes with an author fingerprint that can be separated from her other eras with the same certainty with which you’d expect her to be separated from another author on the same beat.

#### Erin Reed Writes More Similarly to Other Trans Journalists Than to Her Pre-LLM Self

Reed is closer to three of the four other bylines we compared her against then she is to her pre-ChatGPT self. Erin in the Morning The Needle News 0.273 0.167 0.210 0.174 0.236 0.258 0.355 0.438 Reed pre-ChatGPT 19 pieces Reed Nov. 2022 to 2023 252 pieces Reed 2024 196 pieces Reed 2025 to Feb. 5, 2026 186 pieces Reed after Feb. 9, 2026 91 pieces Baum 206 pieces Lazine 30 pieces Douglas 44 pieces Brigham 84 pieces MEASURED FROM REED AFTER FEB. 9, 2026 to Baum, Erin in the Morning 0.236 to Lazine, Erin in the Morning 0.258 to Douglas, The Needle News 0.355 to Brigham, The Needle News 0.438 HER OWN DRIFT, FOR SCALE to Reed pre-ChatGPT 0.412 0.403 that group’s own noise floor, n = 19 0.100 DELTA Anything narrower than 0.100 Delta is basically the same author fingerprint. 

Reed, earliest periodReed, later periodsBaumLazineDouglasBrigham

0.412 Delta 

Reed pre-ChatGPT to Reed after Feb. 9, 2026

0.236 Delta 

Reed after Feb. 9, 2026 to Baum

0.355 Delta 

Reed after Feb. 9, 2026 to Douglas

0.438 Delta 

Reed after Feb. 9, 2026 to Brigham

Five periods of Reed's writing, compared against four known-human bylines. Two bylines are at her own outlet, and two are at our own. Every line carries its own Burrows’ Delta. Brigham is the only byline farther from Reed than her own past self. Positions are arranged for legibility only. Every distance is the number printed on its line. These centroid distances do not flatten into two dimensions, so the map is not drawn to scale and carries no ruler. The strip below is one axis from one origin, and is drawn to scale. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

Authorial drift can happen for reasons other than AI usage. It could mean evolution as a writer. It could mean learning more. It could mean one author working with another author enough that their styles grow closer together.

However, this much of a drift this fast, combined with our other evidence, leads to our claim that Erin Reed’s aggressive authorial shifts after ChatGPT was launched, and again after she switched to Claude, are further evidence of an AI coming between her and what she publishes. 

The machines that Reed uses to write for her are increasingly infecting her authorial voice. Her voice is growing less human and more algorithmic.

### **6.7: Preserving the Record**

It is possible that Erin Reed’s articles may be edited in the future to remove signs of LLM writing. 

The database of her writing used for this article was created from Aug. 31 to Sep. 1, 2026, and includes all articles prior to that point. 

To see the database as we saw it, we preserved every article from July 2022 to Aug. 2026 under Erin Reed’s solo byline at EITM on the [Wayback Machine](https://web.archive.org/) on Sep. 15\. We also preserved every article EITM has online under any byline for the same time period on Sep. 16\. This will allow readers to see her work as we saw it while working on this article.

## **7: How AI has Affected Erin Reed's Journalism**

Erin Reed’s reliance on LLMs is making her an observably worse journalist.

Her writing is less clear because of the AI she increasingly relies on. She is deskilling herself by relying on her AI to phrase her writing. This is leading her to have more errors in her work, and making her less willing to correct and retract untrue information.

### **7.1: Erin Reed’s Writing is Getting Harder to Read**

A key job of a journalist is taking complicated information and breaking it down into a format that [readers can understand](https://substack.com/@johnganz/p-215484423?ref=theneedlenews.com). If the writing is not understandable by its readers, then the journalist has failed to fulfill their role.

There are ways to score a text based on how hard it is for the reader to understand it. The most popular of these is the [Flesch-Kincaid Score](https://readable.com/readability/flesch-reading-ease-flesch-kincaid-grade-level/?ref=theneedlenews.com). This score is commonly used in primary and secondary schools to determine the reading levels of books, based on factors like the length of words and sentences. 

These scores are closely associated with U.S. grade levels. For example, a book with a score of 1 would be at the reading level of a first grader. Likewise, an article with a score of 11 would be at grade level for a high school junior.

Writing for the general public, including news writing, is often pitched at around a [score of 8](https://readable.com/readability/flesch-reading-ease-flesch-kincaid-grade-level/?ref=theneedlenews.com). The [words of U.S presidents](https://www.cmu.edu/news/stories/archives/2016/march/speechifying.html?ref=theneedlenews.com) are usually lower.

#### Erin Reed’s Writing is Getting Harder to Read

Journalists are meant to write for general audiences. Reed’s writing has the complexity expected for college and postgraduate-level readers.Flesch-Kincaid grade level of Erin Reed’s articles. Background bands mark reading difficulty: green for high-school level (grades 10 to 12), yellow for college (12 to 15), and red for postgraduate (15 and above).10111213141516High schoolCollegePostgraduate20222023202420252026Flesch-Kincaid Grade: 13.112-month trailing mean: 14.4ChatGPT launchLikely start of Reed’s AI writing11.113.914.215.7Pre-launch baseline is 11.1\. The peak is +4.6 grade levels above that.Flesch-Kincaid GradeFewer than five articles that monthEra average12-month trailing mean 

Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark "The Needle".TNThe Needle 

Before the release of ChatGPT, Reed's writing hovered around the level of a high school junior or senior. Over the course of a few months in early 2023, the difficulty of her writing shot up by two grade levels, placing it firmly at the level expected of college students. 

Over the next few years, her writing difficulty continued to rise, eventually peaking at 15.71, a level normally reserved for academics; and 4.64 grade levels above her own pre-LLM baseline. 

Erin Reed, the most popular trans journalist on the planet, whose writing is the single biggest source for trans people in the United States to understand their situation, has spent most of her career writing at a difficulty that would normally take at least a university junior or senior to fully understand. 

In addition, she spent six separate months writing at an average complexity typical of postgraduate-level academic papers.

Most native English speakers are not trained to understand texts at this level unless they have a postgraduate education, which few do. Even those with this education will lose the ability to parse these difficult texts if they do not practice it. 

For a writer to write at this complexity is to leave most readers guessing what they meant to say; or leaving them to gloss over the parts they can't understand, filling in the parts they can't understand with their imagination.

### **7.1.1: Sentence Length**

The rising difficulty of her writing is not because she started writing like an academic. Reed is not using bigger words that are harder to parse. 

The difficulty of her writing comes from how long her sentences have become. From her first six months of writing, to her latest six months of writing, her sentences have gone from being around 17 to around 23 words long, a 33.6% increase.

She's taking longer to get to her point. 

Remember that paragraphs showing signs of being pasted into her word processor from an external source are 53 words longer on average, and are 4.2 grade levels harder to read on average than paragraphs without these signs. This means her pasted paragraphs are 64.4% longer by word count, with a 36.6% higher grade level, and 33.6% more words per sentence.

#### Her Writing Getting Harder Because Her Sentences are Getting Longer

Three metrics, each indexed to Erin Reed’s pre-ChatGPT average (100%).

Three style measures for Erin Reed’s articles, each indexed to its own pre-ChatGPT average, by month, July 2022 to August 2026.Reed’s sentence length and reading grade are increasing well above her Pre-AI baseline, but her vocabulary barely moves at all. 75% 100% 125% 150% pre-ChatGPT average = 100% ChatGPT released, Nov 30, 2022 142% 116% 104% Jul 2022 Jan 2023 Jan 2024 Jan 2025 Jan 2026 Aug 2026 

Words per sentenceFlesch–Kincaid scoreVocabulary range (MATTR)

Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

This is notable, since both her average word count and average readability score began creeping upwards several months after the release of ChatGPT.

### **7.1.2: Readability and Pasted Writing**

In addition, these months also saw the average number of marked sentences in her writing greatly increase. 

Before April 2023, her monthly average Flesch-Kincaid score was 11.28\. This is just under halfway between the reading level of a high school junior and senior. 

From April 2023 forward, her monthly average Flesch-Kincaid score is 14.23\. 

#### Sentences From an External Source Are Over 4 Grade Levels Harder to Read

High-schoolCollege standardGraduate / academic papers 0 5 10 15 Flesch-Kinciad Grade Level Pasted Non-pasted 16.83 12.55 +4.28 p = 5.6 × 10⁻⁶⁹ Cohen’s d = 0.64 (medium) 

Sentences with paste markers (n = 1,137) Sentences without paste markers (n = 2,989) 

Flesch-Kincaid grade-level comparison between sentences with and without paste markers (straight apostrophes/quotation marks) in Erin Reed’s byline. Feb. 9 to Aug. 31, 2026\. Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark “The Needle”.TNThe Needle

Over the same time frame, straight (pasted) apostrophes go from 9.94% of apostrophes before April 2023, to 34.21% of all apostrophes afterwards.

The trend for both began shortly after ChatGPT became available.

While there is no clear scientific literature on whether AI prose is correlated to higher Flesh-Kincaid scores or [sentence length](https://arxiv.org/abs/2510.05136?ref=theneedlenews.com) directly, AIs do tend to be longer and [more verbose](https://arxiv.org/abs/2301.07597?ref=theneedlenews.com) than humans when responding to questions. 

This is especially true of flagship models. Their training process [rewards AI bots](https://arxiv.org/abs/2310.03716?ref=theneedlenews.com) for longer outputs that use more tokens, thus requiring users to buy more tokens from the AI company. 

Her pasted writing is a direct cause of the rising difficulty of her writing.

### **7.1.3: Readability and its Effect on the Reader**

The difficulty of reading her articles is not because she started writing like an academic. Rather, it is because her articles are increasingly ghostwritten by soulless automatons that don't know how humans think, much less how we read.

#### In 2026, Reed's Pasted Paragraphs are Almost 7 Grade Levels Harder to Read

Median Flesch-Kincaid grade level per paragraph under Erin Reed’s byline, 2026.

Median Flesch-Kincaid grade level per paragraph under Erin Reed’s byline. Paragraphs carrying paste markers read at a median Flesch-Kincaid grade of 15.1\. Paragraphs without them read at a median of 8.4, a gap of 6.7 grade levels, across 1,622 paragraphs.05101515.18.4PastedNon-pasted+6.7 grade levelsFlesch-Kincaid grade levelp = 4.5 × 10⁻⁶⁸ 

Paragraphs with paste markers (n = 464)Paragraphs without them (n = 1,158)

Source: Erin Reed’s Substack archive. In house analysis by The Needle.

The NeedleA vial with a blue cap, purple contents, a pink label carrying the letters TN, next to the wordmark "The Needle".TNThe Needle 

Her outlet is the primary source that hundreds of thousands of trans people use to understand what civil rights they have. These people rely on her to inform them in a way they can actually understand. 

Both of the people involved in this investigation often struggled to understand what she was saying in her recent articles, despite both having postgraduate education. We can’t imagine how someone with a high school degree or less would be able to get any meaningful information out of her recent work.

If she was writing academic treatises on trans rights, this could be forgiven. However, Reed is a journalist whose work (at least tries to be) targeted towards a general audience. 

Her poor writing, as measured by its increasing complexity and obtuseness, is a failure of her responsibility to inform her audience. That failure is a direct and measurable consequence of her use of LLMs.

### **7.2: Deskilling**

A subjective read of Reed’s early work by Jane Migliara Brigham finds that in the time between when she quit her day job (Dec. 2022) and when the first major signs of AI appeared in her writing (April/May 2023), her writing was of a quality far higher than it is today.

Reading her early work, it's not hard to see how Reed made a name for herself. She wrote at a high quality that was easy to understand, and she gained respect as a result.

What makes a journalist worth reading isn’t just the strength of their analysis. No, it’s the strength of their connection to the community they write for, multiplied by their ability to skillfully weave facts into plain language, such that almost anyone can understand why those facts matter. 

An AI can find papers. It can surface quotes from sources. But it can’t understand, or think, or feel, in the ways you and I can. It doesn’t understand context clues. It can’t understand the visceral feeling of [facing your own genocide](https://www.lemkininstitute.com/single-post/experts-warn-u-s-in-early-stages-of-genocide-against-trans-americans?ref=theneedlenews.com), or policies meant to make you and those like you extinct. 

Using AI to write for you, *or worse, to think for you*, puts you on a fast track to two major harms. 

One is personal: [it makes you dumber](https://arxiv.org/html/2605.03512v1?ref=theneedlenews.com#bib.bib7).

The second harm is collective: it [severs the ties between yourself](https://www.youtube.com/watch?v=YCPAIg7RUq8&ref=theneedlenews.com) and the rest of humanity.

Both harms can be damning for journalism. 

The deskilling that Erin Reed has subjected herself to can hopefully be undone with time and effort.

Untangling AI from her workflow is entirely possible, if she so chooses.

After reading what she was once capable of, going back to Reed’s more recent articles is jarring. The quality decreases, and the difficulty of understanding it increases. This isn't just seen in the readability scores; you can feel the difference when reading them.

In the years since Reed put out her best work, her reliance on AI has deskilled her as a journalist.

### **7.3: AI on the Brain**

While the science is young, and conflicted on whether one can regain full cognitive function after what might (uncharitably but fairly) be called ‘AI-induced brain damage’, the fact remains that the human brain and its plasticity are enduring, and it can recover from a great deal.

A more academic way of talking about[ AI-induced](https://www.psychologytoday.com/us/blog/psych-unseen/202605/your-brain-on-ai-cognitive-offloading-debt-and-atrophy?ref=theneedlenews.com) brain damage is the term[ “*cognitive debt*”](https://arxiv.org/html/2606.15078v1?ref=theneedlenews.com). It was defined as, roughly; a collection of burdens within the brain which accumulates when you use AI to substitute for your own thinking.

In other words, there is an academic and scientific argument that a person relying on AI to think for them makes them measurably dumber. 

The same paper also states that AI, when used in the right ways, can lessen this cognitive debt.

They argued that using AI to generate friction instead of convenience in one’s intellectual work can repay some of that debt. If you use AI not in place of traditional learning, but to force you to justify your thoughts, it can repay some of the cognitive debt incurred by using it to think or write for you.

Crucially, this is not the same as saying that using AI thinking and writing alongside one’s own is helpful to the person in question. The only way you can get any benefits out of this method is if you use AI as solely an oppositional force.

This is not what Reed is doing.

## **8: Journalism**

Good journalism matters because getting stories right matters. Getting stories right requires a human touch. 

Reading, recalling, thinking, revising, writing; hell, even tearing yourself up over the potential harms of a story, and weighing how to minimize them, while sticking to the facts so that you’re actually helping your community; all of this goes into good journalism.

An LLM cannot provide the humanity needed to do this.

According to the [proposed revision](https://www.spj.org/spj-ethics-code-revision-project-2026/?ref=theneedlenews.com) to the Society of Professional Journalists’ (SPJ) Code of Ethics:

> Nearly six in 10 (56%) of journalists use AI at least weekly, and AI-generated content is increasingly appearing in news, raising persistent ethical and practical questions about the use and misuse of technology, as public concern rises about AI use in society, including in news.

Both the proposed revisions and current SPJ Code of Ethics agree that journalists have a responsibility to be ethical and accurate. 

The current code opens by stating:

> Journalists should:  
>  
> \-Take responsibility for the accuracy of their work. Verify information before releasing it. Use original sources whenever possible.   
> \-Remember that neither speed nor format excuses inaccuracy.   
> \-Provide context. Take special care not to misrepresent or oversimplify in promoting, previewing or summarizing a story.   
> \-Gather, update and correct information throughout the life of a news story

The SPJ is trying to revise its code to address “*profound changes in how news and opinion journalism are produced, distributed and understood.*” From the proposed revised code:

> Journalists should:  
>  
> \-Explain ethical choices and processes to audiences. Listen respectfully while encouraging a civil dialogue with the public, and respond quickly to legitimate questions, concerns and criticism.  
> \-Acknowledge mistakes. Correct them promptly and prominently. Explain corrections and clarifications carefully and clearly.  
> **\-Expose unethical conduct in journalism**, including within their organizations. 

*Emphasis added by The Needle.*

### **8.1: The Part Where We Remind You That Money is Involved**

By allowing an LLM to write under her byline, and not disclosing this fact, Erin Reed isn’t *just* taking a slurry of all the world’s writing, molding it into a form approximating whatever was in her notes, and feeding the resulting slop to her readers, the way a bird pushes pre-chewed food down the throats of their children; she’s making a living doing it.

The news business where she works is brutal. Getting more content out, faster, is incentivized by the algorithms and news cycles that journalists are subjected to.

That’s why many journalists cut corners. That’s why writers settle for words that aren't their own. [It’s about the money](https://www.youtube.com/watch?v=yDp3cB5fHXQ&ref=theneedlenews.com).

Releasing bad work faster makes sense if you care less about the quality of writing than about the money you can make off of it.

### **8.2: Personal Experience**

Jane Migliara Brigham, one of the writers on this story, has spoken to Erin Reed about her journalism in the past.

In Jan. 2025, Brigham was interviewed by Erin Reed for a writing position at *Erin in the Morning*. She was offered $60,000 to write 4 articles a week at 600-800 words each. 

For contrast, in that same month, Artemis T. Douglas, the other writer of this story, interviewed at *Colorado Public Radio* for a role that would involve daily quick-turn stories, roughly one brief a day, about something that happened in Colorado, for $45,000-$50,000\. 

Douglas had prior journalism experience. Brigham had never reported a story in her life, and would not do so until later that year, when the two of us co-founded *The Needle*. 

This difference proves that EITM has [money to spend](https://theneedlenews.com/2025/12/you-get-the-news-that-someone-is-willing-to-pay-for/); far beyond what most journalists covering trans news have access to. As far as we are aware, there are only two trans journalists working exclusively on the trans beat in the United States while making a full time salary from their work: Erin Reed, and the employee she took on last year.

Reed got that money by pumping out articles faster than anyone else in trans journalism. 

Even now that there is actual competition in this field, EITM is still guzzling up all the dollars in the room.

That’s why she writes her articles for the people who have money to spend, not the people who need to be informed the most. 

The people with the most money aren't the people who need information the most.

When Brigham asked Reed in her interview about how Reed writes for a trans audience, Reed paused, looked conflicted, and responded with something to the effect of:

> Our main audience isn’t trans people; it's the cis family members of trans people.

## 9: Trans Voices

The Needle has no way of knowing what percent of EITM’s readership is trans versus cis. To be honest, we doubt that even Erin Reed herself has hard data on the topic. 

Whether or not more cis eyeballs get on her work than trans ones, it's clear that trans people place far more importance on her work than cis people do. Trans people use her reporting to learn what rights they have.

Cis people don’t.

Even today, EITM is still the only source of trans news that many trans people even know exists. Her own name now serves as a generic name for trans news at large, in the same way that many older people used the phrase ‘a Nintendo’ to refer to video games at large.

If Reed is the only source of trans news that many trans people know about, and she fucks up in her job as a journalist, then as far as those trans people know, every trans journalist has fucked up.

That’s why Reed using AI to write for her is a big deal. Like it or not, her brand, and that of trans news at large, are closely intertwined. If Reed’s brand falters, the entire brand of trans news falters with it. Everyone involved in a trans-centered alternative to the mainstream media’s bigotry is seen as ‘an Erin Reed’.

That misplaced trust which so many put into her is the central reason why her AI use is so important.

One trans woman *The Needle* interviewed said that “*if \[trans journalists\] are using GenAI or LLMs to write their entire article, they are engaging in \[journalistic\] dishonesty, and giving those who'd see us dead money and clout.*”

She followed up by tying it to a broader harm, arguing that “*it's just contributing to the horror that is the data center craze*”.

When asked directly what she’d say if it could be shown that a trans journalist was ‘pasting as much as 60% to 100% of their articles straight from an AI interface’, she said:

> It comes across as just plain laziness and as disingenuous.

A pattern of factual errors leads to burnout and doomscrolling.

Given the abysmal state of trans rights in the US today, making the situation appear even worse than it is makes it harder to distinguish between [true threats](https://theneedlenews.com/2026/05/trump-administration-proposes-barring-federal-agencies-federally-funded-organizations-from-promoting-gender-ideology/) and [fake news](https://theneedlenews.com/2026/01/alleged-fbi-letters-are-fake-2026-01-16t18-18-47-05-00copy-of-alleged/).

Journalists have a responsibility to tell the truth to the best of their abilities. When they fail to do so, the audience becomes unable to distinguish real threats from fake threats. 

If following the news exposes you to as much fake news as true news, then you would be less misinformed by not following the news at all.

We cannot allow this to be the norm.

A lot of the errors *The Needle* found in Reed’s reporting in 2026, especially the less obviously egregious ones, painted a worse picture for trans people than would be supported without those errors. 

There is arguably a direct line between Erin Reed’s writing and our community’s own burnout and general inability to properly understand the threats we face. 

Her work explains why, before we started *The Needle*, when we asked our friends and community what they wanted out of trans journalism, the most common thing that both of us were told was “*don’t be like Erin Reed*”.

When trans people read trans news written by a fellow trans person, there is an unspoken assumption that the author gives a shit about them. 

Erin Reed presents herself as someone who gives a shit about the people she makes a living covering. The soulless automaton that increasingly ghostwrites her articles doesn't even have the appearance of warmth. 

Whatever faults Reed has, she is capable of understanding that trans people are a community of living, breathing people worthy of dignity. This isn't just because she is trans herself; it’s because she is at least human.

To her AI, we will never be more than 1s and 0s. What Erin Reed publishes proves that AI can write stories about the plight of trans people, but it will never understand why those stories matter, or even give a shit about us.

If Erin Reed wants to serve the trans community, that means using a trans voice - *her voice* \- to do it.

*Erin Reed did not respond to our request for comment.*

[CLICK HERE TO SEE OUR METHODOLOGY FOR THIS INVESTIGATION](https://theneedlenews.com/2026/09/erin-reed-ai-methodology/)