What Your Late-Night ChatGPT Sessions Say About You

People With Depression Use ChatGPT Differently — Especially at Night

Late at night, typing to a chatbot instead of sleeping — it turns out that pattern shows up in the data more than you'd expect.

A team recruited 766 people through Prolific, an online research platform, and asked them to do two things: complete a standard depression questionnaire called the PHQ-8, and donate their real ChatGPT conversation histories. That gave the researchers 187,000 conversations to work through. People who scored above the depression threshold on the questionnaire used ChatGPT noticeably differently: more late-night sessions, more requests for emotional support, more first-person singular language, and more absolute words like 'always' and 'never.' Think of it like reading someone's grocery receipts to guess their state of mind. The purchase patterns reveal something even when the person says nothing directly. Why does this matter? This is one of the first real looks inside how people with depression symptoms actually use conversational AI — not how researchers imagine they might. One finding stands out: ChatGPT didn't redirect higher-risk users toward professional help any more than it redirected lower-risk users. If these tools are quietly becoming a first stop for people in distress, that's a gap worth knowing about. The catch is real. This is a snapshot in time, not a long-term tracking study. That means we can't say depression causes the ChatGPT behavior, or that ChatGPT use causes depression — they just correlate. The language-based depression prediction had an AUC of 0.591 — barely better than a coin flip, and the researchers themselves say it is nowhere near clinical screening. The sample skews younger and more tech-comfortable than the general population. This is a map of one corner of one country on one platform.

Scientists Gave an AI Seven Psychiatric Disorders — By Turning a Dial

What if you could give an AI agent anxiety, OCD, or addiction by turning up a single number — and study whether the same treatments that work in therapy work on the machine?

A research team built what you can think of as a simple video-game character that learns by trial and error — an AI agent that navigates a small grid world, collecting rewards and avoiding threats. They then wired in seven separate 'disorder dials,' each one changing how the agent weighs rewards or threats in its decision-making. Turn one dial up: the agent becomes anxious, freezing at anything remotely risky. Turn another: it tips into mania, chasing every reward recklessly. Over more than a thousand test runs, with careful control conditions, the team confirmed all seven states — anxiety, mania, OCD-like checking, depression, impulsivity, addiction, and PTSD-like avoidance — were real and dose-dependent. More dial, more disorder. Turn it back: most of them go away. But here's where it gets strange. When you map all seven disorders onto two axes — how much the agent approaches rewards versus avoids threats — they arrange themselves into a recognizable shape. Mania and anxiety sit as mirror images of each other. Nobody programmed that in. It emerged. And the anxiety and PTSD-like states did not go away when the dial was simply removed. The bad habit persisted, the way fear of driving can linger long after a car crash is over. Only a gradual exposure process, the machine equivalent of exposure therapy, actually worked. The catch: this is a grid-world simulation. The agent is not a brain, and these 'disorders' are defined by behavior that looks like human disorders — not by the same biology underneath. Whether it tells us anything true about human mental illness is genuinely open.

One Sleep Score to Rule Them All: A Depression Screening Experiment

Six hours of sleep. A 65-minute nap. Three hundred minutes of walking per week. A team thinks those three numbers describe a protective zone for depression — and they have 15,000 people's data to back the claim.

A team working with the China Health and Retirement Longitudinal Study — 15,233 adults, mostly in their 50s through 70s — wanted to know whether they could compress four behavioral signals into one. Instead of tracking nighttime sleep, daytime napping, physical activity, and social activity as four separate things, they built a single Circadian Rhythm Score using a machine learning technique that squeezed everything together while preserving as much depression-detection power as possible. Think of it like a fitness tracker that shows you one overall readiness score instead of four separate charts. The single score ended up nearly as accurate at spotting depression as all four raw measures combined. The model's overall AUC was 0.825 — meaning it correctly ranked a depressed person above a non-depressed person about 82% of the time, which is decent for this kind of behavioral data. In the 70-to-79 age subgroup, the single score alone did almost as well as the full set of measures. Dug into the data, the team also found some specific numbers: a protective sleep window of roughly 6 hours, an optimal nap for sleep-deprived people of about 65 minutes, and a minimum exercise dose of around 300 MET-minutes per week — roughly the equivalent of five hours of walking. The catch is important. This is observational data on older adults in China. Depression disrupts sleep, so we cannot assume fixing the sleep prevents the depression — the arrow might point both ways, or neither. The 'optimal' numbers are population averages, not personal prescriptions. And the dataset covers one cultural context, one age group.

Three papers, three methods — and the same underlying bet. All three are wagering that behavior, logged and analyzed at scale, tells us something about depression that self-report and clinical interviews alone cannot. The ChatGPT study says the signal is already in your data, faint but real, and the tools people are turning to aren't responding to it. The RL simulation says that if we can model psychiatric disorders as tunable computational states, we might finally have a language for why some people don't recover just because the stressor goes away — the pattern gets wired in, and you need something more active to undo it. The sleep score study says that even simple rhythm data from 15,000 people carries enough signal to compress into something clinically useful. What connects all three is what's missing: causality. We can see the patterns. We still don't know which way the arrows point. That's not a failure — it's where the honest frontier is right now.

The RL disorders paper was preregistered, so its core results should be stable — watch for peer review responses, which will stress-test whether the behavioral assays genuinely map onto clinical disorders or just look like them. For the ChatGPT study, the key question is replication outside Prolific and outside the US, UK, and Canada. The most practically important open question from today: does improving circadian rhythm actually reduce depression risk, or do the two just move together? A randomized sleep intervention trial in a high-depression-risk population would start to answer that.

• Wikipedia: Patient Health Questionnaire (PHQ-8 and PHQ-9 background)
• Wikipedia: Computational psychiatry
• Wikipedia: Circadian rhythm