AI Reads Your Words, Misreads the Crowd

AI Can Fake One Depressed Patient But Not a Real Population

One in three AI-generated mental health patients flips their diagnosis between two runs of the same test.

Imagine you hire four actors to improvise being a depressed patient. Each gives a convincing, emotionally coherent performance. No obvious mistakes. But when you line all four up next to a real population of depressed people, something is off: everyone sounds moderately depressed. Nobody sounds severely ill. Nobody sounds mildly affected. The full human range has been squished toward the middle. That is essentially what the team behind PsychBench found when they tested four leading AI models — GPT-4o-mini, DeepSeek-V3, Gemini 3 Flash, and GLM-4.7 — on their ability to simulate mental health patients. They generated 28,800 synthetic patient profiles using standardized psychiatric questionnaires, then compared those profiles to real population health surveys from the US (NHANES and NESARC-III). The results are worth sitting with. DeepSeek-V3 compressed the real range of symptom severity by 62 percent. Across almost every demographic group, models overestimated depression scores by 3.6 to 6 points. Transgender women — the group with the highest documented real-world burden — were underestimated by 5.4 points. That is the group most likely to be missed when accuracy matters most. Here is the catch, and it is important: not a single AI-generated profile violated the logical rules of psychiatric diagnosis. The models get individual patients right. They fail at the population level. That distinction matters enormously if you are using these models to test a therapy app, generate training data, or simulate clinical trials. You would be optimizing for a population that does not actually exist. The team calls this coherence-fidelity dissociation — coherent individuals, unfaithful crowds. It replicates across US-built and Chinese-built AI architectures, which suggests this is not a one-company problem.

How You Structure Your Story Predicts Your Mental Health

It is not what you write about your trauma — it is whether your story has a clear beginning, middle, and end.

Starting in 2018, therapists and researchers in China began collecting short written exercises from people going through difficult experiences. By 2024 a team had assembled 830 of these samples, spanning clinical settings, post-disaster communities, schools, and online groups, covering ages 9 to 50. They then asked a question most researchers in this space had not thought to ask: does the architecture of a story — not just the words in it — predict someone's mental health? Think of a detective reading a statement. A good detective notices not only what you say but how you say it. Does your account jump around in time? Do you keep circling back to the same moment without moving forward? Is there a clear before, a turning point, and an after? Those structural cues carry information independent of the emotional vocabulary you use. The team tested three levels of analysis: counting emotionally loaded words (like reading a label), measuring the overall emotional tone of the text (like tasting the dish), and analyzing the narrative architecture — whether there was a proper orientation, complication, and resolution, using a framework called Labov's story grammar and an approach called Rhetorical Structure Theory. That third level, analyzed with the help of large language models, substantially outperformed the other two for predicting depression, anxiety, and trauma severity across all 830 samples. Two signatures stood out: depressed writers tended to scramble their timelines; anxious writers struggled to anchor events to a specific place. The honest limit: all samples were in Chinese, from specific Chinese populations, and the full statistical methods were not published at the time this digest was written. Replication in other languages and contexts is the obvious and necessary next step before this touches anything clinical.

A Few Words You Type Could Measure Your Depression

Write down five words describing how you feel right now — that list might measure your depression as accurately as a clinical questionnaire.

In modern language AI, every word lives at a location in a vast mathematical space. Words with similar meanings cluster together. 'Sad' sits near 'grief' and 'hollow.' 'Energized' sits near 'alert' and 'motivated.' It is like a city where neighbourhoods form by meaning rather than by street plan. The researchers behind this paper exploited that geography. They took items from standard clinical questionnaires for depression and anxiety — phrases like 'I feel blue' and 'I feel tense' — and used them to define a direction in that word-space. Imagine drawing a line from the healthy neighbourhood to the distressed neighbourhood. Then they asked 247 people to do something simple: write a few words or phrases describing how they currently felt. Those responses were then measured against the line — how far along the 'distressed' direction does your language sit? The result was striking. For structured responses — a short list of chosen words or brief phrases — the score correlated up to r = 0.87 with standard clinical depression measures like the PHQ-9. That is an unusually strong match for an approach that requires no labeled training data and no clinical interview. You do not need to teach the system what depression looks like. You just need a compass. The practical implication is clear: a screener that works from a few typed words, cheap to adapt to new populations, requiring no expensive annotation. But here is where I want to slow you down. The sample was 247 people, all recruited online via the platform Prolific — not a clinical population. Long free-form text worked much worse than structured short formats. And correlating well with a questionnaire is not the same as clinical validity. This is a promising signal, not a deployable tool.

Put these three papers side by side and a single theme sharpens into focus: we are getting much better at reading mental health signals from language, but we are not yet honest enough about who we are reading and whether the reading is trustworthy. The semantic projection paper says: a few words can mirror a clinical questionnaire surprisingly well. The narrative paper says: even the shape of your story carries a signal. Those are genuine steps forward for low-cost, accessible screening. But PsychBench stands as a warning label on both. If we build AI screening tools and then train or test them on AI-generated patients — a practice that is already happening in the field — we risk optimising for a population that does not exist. Marginalised groups, the people with the highest real burden, get erased first. The honest position is this: language is a real window into mental health, and AI is learning to read it. What it cannot yet do is reliably read a crowd. That gap matters before any of this moves into a clinic or an app.

The most important open question hanging over all three papers is external validation on clinical populations — not Prolific users, not synthetic patients. Watch for replication studies, particularly on narrative analysis in non-Chinese languages. On the PsychBench side, it would be worth seeing whether fine-tuned clinical models (as opposed to general-purpose frontier models) show the same population-level distortion or whether targeted training helps. No specific conference or trial result is imminent that I can point to, but the pressure on AI mental health tools from regulators in the EU and the US is building, and these methodological audits are exactly the kind of work those conversations need.

• Wikipedia: PHQ-9 depression questionnaire (the benchmark used in several papers today)
• Wikipedia: Word embedding — how AI maps word meaning as locations in space
• Wikipedia: Narrative — background on story structure concepts like Labov's grammar