Your Smartwatch, Your Psychiatrist, and the AI In Between

An AI Found 41 Depression Clues Hiding in Wearable Data

Your smartwatch is recording something your doctor has never looked at — and an AI just found 41 reasons to start.

Imagine dumping three years of someone's Fitbit data on a table and asking a scientist to find anything that tracks with depression. The data is enormous, messy, and cross-referencing it by hand would take months. CoDaS — short for AI Co-Data-Scientist — does that scan automatically and then, crucially, argues with itself about which findings are real. The system runs through wearable sensor data in phases: it generates hypotheses, runs statistics, then deploys a separate adversarial agent to try to knock down each finding. Think of it like a cooking competition where you must make a dish, then defend it to a panel actively trying to find flaws. Across 9,279 participant-observations from three datasets, CoDaS flagged 41 candidate signals — measurable patterns that track with depression. Two kept surfacing independently: how much your sleep duration varies night to night, and how unpredictably your sleep onset shifts. The system found a correlation of 0.25 between sleep duration variability and depression in one cohort, and a similar pattern replicated in a second, unrelated dataset. Why does this matter? If these signals hold up, a consumer wearable could flag early depression risk before you or your doctor would notice anything. Passive, cheap, always-on. The catch: 'candidate' is doing a lot of heavy lifting in that sentence. The prediction improvement was real but modest — about 4% more explained variance. Correlation doesn't tell you whether erratic sleep causes depression, results from it, or both spring from a third thing entirely. These candidates need prospective clinical validation before anyone acts on them. This is a promising lead, not a diagnostic tool.

AI Psychiatrists Look Right on Paper but Are Wrong About Everyone

When AI plays psychiatrist, each individual patient it invents looks convincing — but the whole crowd is quietly wrong.

The PsychBench team ran a large audit. They asked four major AI systems — GPT-4o-mini, Gemini, DeepSeek-V3, and GLM-4.7 — to simulate 28,800 psychiatric patients spread across 120 demographic groups, crossing race, gender, income level, and relationship status. Then they checked whether the AI-generated 'population' matched what real epidemiological surveys of the US population actually look like. Here is the analogy. Imagine asking someone to paint a crowd scene. Each face looks like a real person — convincing features, natural expression, believable. But step back, and everyone in the crowd is suspiciously similar in height and age. That is what these AI systems are doing with mental illness. Every individual AI-generated patient passed clinical plausibility checks — no one violated the basic rules for a diagnosis. But the realistic spread from 'doing fine' to 'in crisis' was dramatically flattened. Depending on the model, between 14% and 62% of the real-world variation was squeezed out. The skews are specific and troubling. Most models overestimated depression severity for the average person by 3.6 to 6.1 PHQ points — that is a clinically significant gap. Trans women's symptoms were systematically underestimated, capturing only 8–46% of their documented elevated risk. The bias patterns appeared in both US-developed and Chinese-developed models alike. Why does this matter? Researchers increasingly use AI-simulated patients to test screening tools and train other AI systems. A skewed training population produces skewed outputs downstream. The catch: this paper identifies the problem, it does not fix it. And none of these models are deployed in clinical settings yet. Whether the distortion causes real harm remains an open question.

An AI Rated Depression Severity as Well as Human Clinicians

Two clinicians listening to the same interview can disagree enough to change a treatment decision — an AI just closed some of that gap.

Rating how depressed or anxious someone is from a clinical interview is, honestly, messier than most people realize. A clinician listens, weighs what they hear, and assigns a score. Two experienced clinicians doing this on the same recording can land in meaningfully different places. That inconsistency has real consequences for research and for care. The team behind ADAPTS tackled this differently from previous attempts. Instead of asking one AI model to assess an entire interview at once, they broke the task apart: separate agents each reason about a single symptom — sleep disturbance, loss of interest, psychomotor changes — like a tasting panel where each judge evaluates only one element of the dish before the scores are combined. The system processed 204 clinical interviews from two independent datasets that used structurally different assessment scales. On the subset of interviews where original human raters disagreed most with expert-level scores, ADAPTS reached an absolute error of 22 versus the human raters' error of 26 — meaningfully closer to the expert benchmark. With an extended protocol that incorporated qualitative clinical conventions, agreement reached an ICC of 0.877, which falls in the range considered excellent for inter-rater reliability in psychiatry. Why does this matter? Consistent scoring is the foundation of both clinical trials and individual treatment plans. A tool that stabilizes ratings could improve both. The catch: 204 interviews is a small sample, and research recordings are tidier than real ward conditions. The team is explicit that this is a support tool, not a replacement for clinicians. Edge cases — patients who communicate in non-standard ways — remain uncharted.

Look at these three stories together and a pattern emerges that is more interesting than any individual finding. AI is getting genuinely better at detecting mental health signals — from wrists, from voices, from transcripts. But today's papers keep bumping into the same wall: accuracy at the individual level does not guarantee fidelity at the population level. CoDaS finds real patterns in wearable data, but modest ones. ADAPTS matches expert raters under controlled conditions, but the sample is small and the real world is messier. PsychBench shows AI systems generating convincing individual patient profiles while systematically distorting the population they collectively represent. The groups that get squeezed out of AI models tend to be the most vulnerable — severe cases, underrepresented identities, the statistical edges. In mental health, those edges are exactly where the stakes are highest. The field does not have a detection problem. It has a calibration and equity problem, and today's papers make that clearer than usual.

The DAIC-WoZ clinical interview dataset appears in at least three papers this week — ADAPTS, the voice biomarker study I did not feature today, and the PsyGAT depression detection model. That dataset is doing enormous work in this field and was originally collected from a few hundred people; watch for a formal expansion or a challenger dataset to emerge, because right now a significant slice of AI mental health research is benchmarking itself against the same small pond. More immediately: the circadian instability signals that CoDaS identified need external replication — a prospective study using sleep duration variability specifically as a depression predictor would be a significant test of today's most optimistic finding.

• Wikipedia: Digital biomarker — background on what counts as one and why it matters
• Wikipedia: Intraclass correlation — what those ICC numbers in the ADAPTS paper actually mean