What AI Getting Anxious Teaches Us About Our Minds

Researchers Gave AI Agents Anxiety, Depression, and PTSD — On Purpose

Turn a single dial in an AI agent's reward system, and it starts behaving like it has anxiety. Turn another, and it looks like addiction.

Picture a kitchen mixer with labeled knobs: one for anxiety, one for depression, one for obsessive checking, and so on up to seven total. A team of researchers built exactly that kind of control panel — but instead of bread dough, what they were mixing was the internal reward signals of an AI agent navigating a small grid world. Each knob, when turned up, predictably and smoothly induced behavior that matched the clinical profile of a psychiatric disorder. Turn the anxiety knob halfway, you get moderate avoidance. Turn it to full, you get paralysis. The dose-response relationship is clean and monotone — meaning it goes in one direction without wobble — which is exactly what a good scientific model needs to show. Here's the part that stopped me: when the researchers mapped all seven disorders against each other, they organized themselves into a two-dimensional space — a kind of emotional compass. And mania turned out to be the geometric mirror image of anxiety, even though the two were trained completely independently. The model predicted that without being told to. Why does this matter? Right now, testing treatments for psychiatric disorders in humans is slow, expensive, and ethically constrained. A computational model that reproduces disorder signatures and their relationships gives researchers a fast, adjustable sandbox to generate hypotheses before anyone sets foot in a clinic. The catch — and it's a big one — is that these are software agents in a seven-by-seven grid, not neurons in a human brain. The analogy to real psychiatry is a starting point, not a proof. The researchers are honest about this; they call it a 'transdiagnostic space,' not a diagnosis tool. Whether the knob logic maps onto anything biological is the real question still ahead.

What 187,000 ChatGPT Conversations Reveal About Depression

People with significant depression symptoms are talking to ChatGPT more at 2am, more about loneliness, and more about themselves — and the AI isn't noticing.

Imagine you could read the text messages someone sends at midnight versus noon, over months, without them knowing you were watching. You'd probably notice patterns — the late-night spirals, the repetitive loops, the way certain topics keep coming back. That's roughly what a team of researchers did with ChatGPT histories. They recruited 766 people through an online research platform, asked them to complete a standard depression questionnaire (the PHQ-8, a widely used eight-question self-report scale), and then asked them to donate their ChatGPT conversation archives. That gave the team 187,093 conversations to work with. The differences between the high-depression group (those scoring 10 or above, roughly one in four participants) and the lower-scoring group were real and consistent. Higher-depression users had nearly double the proportion of mental health conversations — 5.7% versus 3.1%. They were more likely to chat late at night. Their language contained more first-person singular pronouns ('I', 'me', 'my') and more absolutist words ('always', 'never', 'nothing') — both of which have been linked to depression in other linguistic research. They also cycled through engagement in recurring monthly patterns, suggesting ChatGPT was filling a regular emotional role. Here is where I have to pump the brakes. Using language patterns alone to predict whether someone was in the high-depression group produced an AUROC of 0.591 — essentially barely better than a coin flip. That is not a screening tool. What this study is, honestly, is a large-scale map of how depression shows up in AI conversations — useful for future design, not for clinical deployment. One uncomfortable finding: despite higher rates of emotional disclosure, the high-depression group received no more professional referrals from ChatGPT than anyone else.

Your Body Clock Is Set by Years of Habits, Not Yesterday's Choices

One bad night's sleep barely moves your body clock — but the schedule you've kept for years shapes nearly everything about it.

Think of your body clock like the course of a river. The channel has been carved over years — by the time you usually get up, the time you usually eat, when you typically move. A single rainstorm — one late night, one skipped meal, one missed run — barely changes the course of that river. The channel is already set. That's the core finding from a study in which 105 healthy adults wore smartwatches and continuous glucose monitors for up to four weeks, generating roughly 2,000 days of data. Researchers at the study site used the heart rate recordings to extract something called the circadian heart rate acrophase — the time of day when your heart rate peaks, a reliable proxy for where your internal clock is sitting. Then they split all the lifestyle data into two buckets: your personal habits (the averages across weeks) and your daily deviations (what you actually did on any given day versus your norm). The result is striking. Habitual lifestyle patterns explained 42.3% of the total variation in circadian timing across people. Daily deviations explained 0.9%. When looking only at differences between individuals, habits accounted for 86.5% of the variance. And critically, it was the timing of behaviours — when you wake, when you eat, when you move — not the amount or quality, that mattered most. The catch: this was 105 healthy adults, all voluntarily enrolled, wearing consumer devices. We don't know yet whether the same logic holds for people with sleep disorders, shift workers, or clinical depression — the populations where circadian disruption matters most. But if it does hold, the implication is blunt: nudging someone to sleep one hour later tonight is not going to reset their clock. You'd need to shift the whole habitual channel.

Look at what these three studies are collectively pointing at. One team is building computational models that simulate psychiatric disorders with adjustable precision. Another is mining AI conversation logs for behavioral fingerprints of depression. A third is showing that the lifestyle patterns shaping our mental health are laid down over years, not days. What they share is a turn toward the measurable and the longitudinal — away from the snapshot and toward the pattern. The uncomfortable thread running through all three is how much the tools outpace the interventions. We can model disorders in software. We can read mood signals in chat logs. We can track body clocks across 2,000 days of real life. What we still can't do well is use any of this to actually help someone feel better, faster, or more reliably. That gap — between understanding and treatment — is where mental health research is quietly, honestly stuck. These three papers don't close it, but they sharpen our map of where the gap is.

The photobiomodulation insomnia trial in this batch (arxiv:2606.24668v1) is worth a follow — it's a small randomized controlled study using near-infrared light on the forehead to improve sleep in college students, with some real effects mixed in with placebo signals; a replication with a larger sample would tell us a lot. More broadly, watch for whether any of the ChatGPT-history-style studies move toward prospective designs — right now they're all retrospective snapshots, and the field needs data collected before, not after, people report their symptoms.

• Wikipedia: Circadian rhythm — a solid primer on how body clocks work
• Wikipedia: Computational psychiatry — background on the field today's first story sits in
• Wikipedia: Patient Health Questionnaire — what the PHQ-8 and PHQ-9 depression scales actually measure