The New York Times changed how it covers trans people.

When the paper of record treats your existence as a debate, the framing does not stay on the page.

The New York Times effectively sets the terms of what is considered respectable mainstream liberal opinion. What the Times treats as settled, the country treats as settled. What it frames as an open question shows up in courtrooms, in statehouses, in exam rooms, and in the minds of parents deciding whether to believe their own kids. I'm a trans person and a lawyer; I have watched that framing land on people I know.

I'm not the first to notice, or to count. Trans journalists and watchdogs have been documenting this for years. Assigned Media tracks how rarely the Times quotes trans people compared with other outlets; Lee Hurley has chronicled the same turn in the British press; and Media Matters and GLAAD found the Times left trans voices out of most of its own stories on anti-trans legislation. For years, trans readers and writers, including hundreds who signed an open letter to the paper in 2023, said the coverage had shifted: more doubt about medical care, more room for opponents, less room for trans people to describe their own lives. The paper's leadership said nothing had changed, and that they were simply following the facts. Both sides were arguing from the same handful of articles when they came out, rinse and repeat.

The subtle shift in framing hides in the aggregate, across more stories than any single person could reasonably analyze.

And that's the issue the Times relies on to defend its reporting. Any single article can be defended on its own terms, and a critic and an editor will each cite the facts that prove their point. The broader editorial change only shows up in the pattern across thousands of articles. This is exactly what no one can see by reading a single article on a given day. So instead of criticizing any particular article, I analyzed all of them.

Every article I could find, read the same way.

First, the corpus. I pulled every Times article on transgender topics published since January 2014 using two of the paper's own public data feeds. The Article Search API, queried for 22 terms (transgender, gender-affirming care, puberty blockers, trans youth, detransition, and so on). And the Archive API, walked month by month and filtered against the Times' own editor-applied subject tags, including the descriptors the paper itself files this coverage under, "Transgender and Transsexuals" and "Transgender." This means the corpus automatically includes every article the NYT has itself classified as being about trans people, 2,266 of them, not just whatever my search wording happened to match. After de-duplicating across both feeds, that's a total of 3,242 articles; I recovered the full text of 3,200 from public web archives. This is roughly the best total you can reasonably accumulate given the nature of the tools available. The only stories it can miss are ones that carry neither the Times' tag nor any of the search words. There's also the issue of some stories being podcasts, videos, and wire stories; many of those were excluded from the corpus.

Then, the hard part: analyzing 3,242 articles consistently. No person can hold a fixed standard across twelve years and millions of words of text, and a plain keyword count can't tell an article condemning anti-trans laws from one promoting them. I wrote a detailed rubric as a prompt to apply the same questions to every article. Does the article take a side on gender-affirming care, or on trans rights? Whose voices carry the story? What does it treat as settled, and what as doubtful? Is the tone one of conflict, or of ordinary life? I had three separate large language models apply that rubric, one article at a time.

I want to be precise about what I used LLMs here for. Much of the metadata already contained information about authors, publication time, placement, and other key details. The more nuanced questions of framing and tone are far harder to pull out of a dataset, and that is something LLMs are genuinely decent at evaluating. They are a way to apply one human-written standard at a scale I couldn't reach by hand. To keep any single model's quirks from steering the result, every article is scored independently by three models from three different companies. One of them is Qwen 32B, an open-weights model that runs on a home computer, so anyone can rerun it and check the numbers without my cooperation. The models were fed the body text itself along with the prompt containing the rubric.

And because "an AI said so" is not evidence, I ran the whole corpus through an old-fashioned natural-language-processing pass (VADER) that does plain word counts and rule-based sentiment. On conflict framing it lines up with the models. On the subtler question of which side a story takes, it falls apart in a way that reinforces the choice to use models for the more nuanced reads: the NLP scores an op-ed condemning bigotry as "negative" because of words like assault and erase, the exact mistake the models are better able to avoid. That gap, between counting words and reading framing, is the whole reason I leaned on models for the parts a word-counter can't do.

Around 2022 the coverage turned. All three models saw it independently.

My initial expectation was a slow, ambiguous drift I'd end up having to argue about, with individual story framings as tentpoles. What the data shows is a sharp, measurable change. Around 2022, the framing and tone used by the Times turned, and it turns on nearly every measure at once. I didn't have to pick a favorable model; all three, scored separately, agreed on the direction. These are the four shifts that hold up.

• 01Protective framing fell. Coverage that used to lean toward trans rights and gender-affirming care drifted toward "neutral," and later into the negative.
• 02The voices traded places. In the paper's own coverage of trans people, the prominence of opponents climbed year after year, until it nearly matched the prominence of trans people themselves.
• 03Conflict crowded out ordinary life. The language of battle and controversy went up; the language of trans people simply living their lives went down.
• 04The medical beat moved the most. In coverage of health care, the same kind of story came to be told a very different way. It's the single sharpest change in the dataset.

There are notable changes in 2022 that align with a shift in editorial stance at the Times. It was the first major year in which hundreds of bills targeting transgender rights were introduced around the country. The paper itself changed: A.G. Sulzberger became chairman in January 2021, and Joseph Kahn became executive editor in June 2022. In February 2022, the Times' most prolific pro-transgender voice, Jennifer Finney Boylan, left after her contract was not renewed, while Pamela Paul, who would become the most prolific anti-trans voice on the editorial board, came into the opinion section. Around the same time, Azeen Ghorayshi took charge of the gender-affirming-care beat, becoming the paper's most prominent and most critical voice on that medicine. None of these changes can definitively prove why the Times changed, but they help show how it did.

It's important to note that none of the findings rests on one article. That's the point of a corpus-wide analysis: it's a pattern across thousands, the thing I couldn't prove by reading a story in isolation. I'm not asking you to take my word for it. The next page is the evidence, and you can take it apart.

I have a point of view. The findings don't depend on it.