Body Clocks, Light Therapy, and AI Agents With OCD

What Happens When You Give an AI Agent OCD

What if you could turn a single dial to give an AI agent something like OCD — and a different dial to give it depression or mania?

That is, essentially, what this paper does. The researchers built software agents that navigate a simple grid world, learning to make decisions by chasing rewards. Then they embedded a set of single-parameter 'knobs' into each agent's reward-sensing system — one knob per disorder analogue. Turn the anxiety knob up, and the agent starts avoiding any situation that carries even a small threat signal, even when the threat is long gone. Turn the OCD knob up, and it compulsively rechecks corners of the map. Turn the mania knob, and it charges into high-risk situations recklessly. Think of it like adjusting the sensitivity dial on a smoke alarm: turn it too high, and it fires at burnt toast; turn it the other way, and the alarm barely registers actual smoke. Across more than 1,000 test runs with preregistered behavioral assays, every knob produced a smooth, graded curve — more dial, more disorder-like behavior — and no control condition reproduced any of these patterns by accident. One result stands out: mania and anxiety turn out to be mirror images of each other in this model, produced by opposite settings of the same knob. That is a testable prediction about human neuroscience that didn't exist before this paper. The catch is significant. These are software agents in a flat grid world. The 'disorders' are behavioral patterns that look like OCD or PTSD from the outside — they are not the lived experience of illness, and they don't involve neurons, hormones, or trauma histories. The authors call them 'disorder-like phenotypes' deliberately. Whether this computational map reflects anything true about human psychiatry is, honestly, unknown. What it does offer is a controlled sandbox where you can induce, measure, and reverse disorder analogues with reproducible precision — something you cannot do ethically in humans.

A Weekly Light Session on Your Forehead Improved Sleep

Pressing a near-infrared laser to your forehead for ten minutes a day, seven days in a row, improved sleep quality — and the effect kept growing after the sessions stopped.

The device looks a bit like a fitness headband. It emits near-infrared light at 980 nanometers — invisible to the eye — aimed at the prefrontal cortex, the part of your brain just behind your forehead that handles decision-making and emotional regulation. The technique is called transcranial photobiomodulation, or tPBM: using light to nudge brain activity from the outside, without surgery or drugs. Think of it like warming up a cold engine before you ask it to run smoothly — the light appears to shift the brain toward a calmer, more sleep-ready baseline. Researchers randomized 37 college students who met clinical insomnia criteria into either real sessions or a sham device that looked identical. After seven daily ten-minute sessions, sleep quality scores in the active group improved significantly — with effect sizes growing at follow-up one week later and again two weeks later. Brain recordings showed that slow, sleep-associated electrical activity decreased while calming alpha-wave activity increased. The researchers, working in China, also ran mediation analyses suggesting the brain changes were doing the work, not just a placebo effect. The catch: 37 people is a small pilot. The population is exclusively Chinese university students. The trial was single-blind — meaning patients didn't know which group they were in, but the clinicians running sessions did, which introduces a risk of unintentional bias. And one key measure, mood, showed no difference between groups at all. One measure of cognitive processing also improved in both groups, suggesting a general practice effect rather than something specific to the light. This is a promising signal from a well-designed small study — not a conclusion you can act on yet.

One Number From Your Daily Routine Can Screen for Depression

Your sleep duration, your nap habits, how much you move, and how often you leave the house can be compressed into one number — and that number screens for depression almost as well as the whole bundle.

Think of it like a credit score, but instead of your bill-payment history, it tracks how well your body is following its natural daily rhythm. Researchers working with data from 15,233 adults in the China Health and Retirement Longitudinal Study built a tool called the Circadian Rhythm Score, or CRS. It takes several behavioral inputs — sleep timing, napping, physical activity, social engagement — and fuses them into a single composite number using a machine learning method that deliberately preserves the positive, protective meaning of each input. That last detail matters: the math was constrained so that the score always moves in an interpretable direction. The resulting score identified people with depression at an accuracy level (AUC 0.825) that came close to using all the raw behavioral measures separately — with almost no loss of information. In adults aged 70 to 79, it actually nudged slightly higher, to 0.83. The analysis also surfaced specific behavioral thresholds: around 300 MET-minutes of exercise per week — roughly five 60-minute brisk walks — is where depression risk started dropping. A nap of about 65 minutes looked protective for sleep-deprived individuals. A sleep window of roughly six hours showed up as a sweet spot. The catch is real. This dataset is from China and skews toward older adults, so we don't know how well the score travels to different ages, cultures, or healthcare systems. More importantly, the study is a cross-sectional snapshot — everyone was measured at one point in time. That means we can't tell whether a disrupted rhythm caused depression or depression caused a disrupted rhythm. Almost certainly the relationship runs in both directions. That matters enormously if you're thinking about using this score for early warning or intervention. Prospective validation — following people over time — is the missing piece.

Three stories, one thread: measurement. A circadian rhythm score turns your daily routine into a single screenable number. A near-infrared light device shifts your brain's sleep architecture measurably in a week. An AI sandbox lets researchers induce, grade, and reverse disorder-like states with a controllable dial. None of these is a treatment breakthrough. But each one chips away at the same fundamental bottleneck: mental health conditions have always been hard to quantify, which makes them hard to study, which makes them hard to treat. What I notice across today's papers is that the action has shifted upstream — toward instruments, not interventions. Who develops better measurement tools right now is probably more important than who runs the next drug trial. The body-clock score is cheap enough to scale. The light device is non-invasive enough for a GP's office. The AI sandbox is reproducible enough to preregister and replicate. That combination of cheap, safe, and reproducible is rare in this field.

On the tPBM front, 37 people is a proof-of-concept number — the authors note a larger follow-up trial is the logical next step, and I'd watch for a multicenter version. On the circadian rhythm score, the key test is prospective: does a low CRS today predict who will develop depression six months from now? That validation study doesn't exist yet, and it's the one that would make this clinically meaningful. The AI disorder sandbox paper was preregistered, which means replication attempts by independent groups are straightforward — watch for those over the next year.

• Wikipedia: Computational psychiatry — background on using math and AI to model mental illness
• Wikipedia: Photobiomodulation — what near-infrared light does to cells and tissue
• Wikipedia: Circadian rhythm — how your body's internal clock works