ChatGPT Logs, AI Disorders, and Light on Your Skull

Depressed users use ChatGPT very differently — and it doesn't notice

At 2 a.m., someone opens ChatGPT to talk about loneliness instead of trying to sleep — and that pattern showed up, at scale, in nearly 200,000 real conversations.

Think of your ChatGPT conversation history like a diary you didn't know you were keeping. A research team recruited 766 volunteers through Prolific — an online research platform — who agreed to donate their actual ChatGPT histories alongside a standard depression symptom questionnaire called the PHQ-8. The resulting dataset: 187,000 conversations, split between 571 people with low symptom scores and 195 people with moderate-to-severe symptoms. What they found is uncomfortable. People with higher depression scores were more than twice as likely to bring mental health topics into their chats — 5.7% of conversations versus 3.1%. They talked more about loneliness, interpersonal struggles, and their own inner lives. They used the word 'I' more. They leaned toward absolutist language — 'always,' 'never,' 'nothing.' And they showed up later at night, in recurring monthly patterns that echo the sleeplessness and rumination common in depression. Here is the catch. ChatGPT didn't respond to this need differently. The rate at which it redirected higher-symptom users to professional help was no higher than for anyone else — the tool wasn't picking up the signal and acting on it. And when the researchers tried to use conversation language alone to predict depression scores, the model achieved an AUROC of 0.591, which is barely better than a coin flip for clinical purposes. This is not a story about ChatGPT diagnosing depression. It's a story about how people are already using AI as an emotional outlet — and the AI, for now, largely doesn't know it's happening.

Scientists gave AI agents anxiety and depression — then tried to cure them

What if you could dial anxiety up and down in a mind like a volume knob — and use that to figure out why exposure therapy works?

Imagine a mixing board where each slider controls a different psychological state: anxiety over here, depression over there, obsessive checking, addiction, PTSD. A research team built something close to this — not for humans, but for software agents that learn by trial and error, the same underlying technology behind game-playing AI. The researchers introduced single 'disorder knobs' into how their agents evaluate reward and threat. Dial one knob and the agent starts over-avoiding, threat-fixated — anxious. Dial another and it chases reward compulsively, ignoring long-term consequences — addicted. They reproduced seven disorder profiles in total: anxiety, mania, OCD-like checking, depression, impulsivity, addiction, and PTSD. Each one showed a clean dose-response curve — more knob, more disorder — that none of their control conditions could reproduce. The finding that jumped out at me: they couldn't cure all disorders the same way. For reward-distortion disorders — addiction, mania, compulsive checking — simply removing the knob restored normal behavior. But for avoidance disorders — anxiety and PTSD — turning off the knob wasn't enough. The agents needed a graded exposure curriculum: reintroducing the feared situation slowly, step by step, over time. That mirrors exactly what works in human therapy for PTSD. The catch, and it's a real one: these are software agents, not people. The 'disorders' are behavioral analogies. The framework makes predictions about treatment that are now testable in humans — but it doesn't prove anything about the human brain yet. Think of it as a very sophisticated thought experiment that has just earned the right to run clinical trials.

Ten minutes of near-infrared light on your scalp improved insomnia — in 37 people

A near-infrared laser aimed at one spot on your scalp, ten minutes a day for a week — and sleep quality improved significantly, with effect sizes that would make a pharmacologist raise an eyebrow.

Transcranial photobiomodulation — tPBM for short — uses specific wavelengths of light to stimulate metabolic activity in brain tissue just beneath the skull. The wavelength here is 980 nanometers, well beyond what the human eye can see. The target: a frontal scalp site called AF8, sitting over a region of the prefrontal cortex known to be underactive in people with insomnia. A randomized, sham-controlled pilot trial enrolled 37 college students who all met clinical criteria for insomnia. Half got the real device; half got a device that looked and felt identical but emitted no active light. Ten minutes a day, seven days. Neither group knew which they had — think of it like a placebo pill study, except with hardware. The treated group improved substantially. Sleep quality scores dropped sharply compared to sham, with an effect size of d=1.27 at the final follow-up — that's a large difference, roughly equivalent to the gap between terrible sleep and decent sleep rather than just a marginal nudge. Brainwave recordings showed real changes: less slow-wave activity linked to disrupted sleep, more alpha-wave activity linked to calm relaxation. And these effects persisted two to three weeks after treatment ended. Now the honest part. Thirty-seven people is a very small group — about the size of a classroom. The assessors knew who got what, which is a known source of bias. Some improvements also appeared in the sham group, hinting at practice effects. This is a pilot: its job is to justify a bigger, better trial, not to become a recommendation. File it under 'intriguing, needs replication.'

These three papers share a question hiding underneath all the technical detail: can we find new leverage points for mental health, outside the traditional clinic? The ChatGPT study tells you that people are already using AI as an emotional outlet — whether anyone intended that or not — and the tool is largely unaware it's happening. The RL agents paper shows computational psychiatry maturing: we can now simulate disorder and recovery in ways that generate testable, specific predictions about treatment. The tPBM pilot suggests that stimulating the prefrontal cortex from the outside, non-invasively, might have real effects on sleep — but we are in early innings. Taken together: the tools are multiplying faster than the evidence base. That's not a crisis, but it is a responsibility. The choices being made right now — about what AI responds to, what it ignores, who gets access to these devices — will shape how millions of people navigate depression and sleeplessness in the next decade. Most of those people are already there.

The tPBM insomnia result needs a much larger replication trial — watch for whether the AF8 frontal-targeting approach gets picked up by a sleep medicine center with the resources to run 200-plus participants. On the AI side, the RL disorder framework is preregistered with specific behavioral assays, so there should be follow-up work testing whether the graded-exposure prediction holds in animal or human analogs. The open question I want answered: will anyone build the ChatGPT study's successor — a longitudinal version that follows the same people over months, not just a cross-sectional snapshot?

• Wikipedia: Computational psychiatry — background on the field the RL disorders paper belongs to
• Wikipedia: Photobiomodulation — what the light-on-skull technique actually is
• Wikipedia: PHQ-9 — the depression screening tool used in the ChatGPT study (PHQ-8 is a close variant)