Your therapy chat knows how depressed you are.

An AI reads your therapy chat and estimates how depressed you are.

What if your therapy app already knew your depression score — without you ever filling in a questionnaire?

The Ash AI platform runs AI-powered therapy conversations. Researchers there asked whether a large language model could read those transcripts and estimate how depressed a user is, passively — no extra forms, no interruption. Think of it like a skilled sommelier who can name the vintage just from the taste, without seeing the bottle. The model — a fine-tuned version of Qwen3.5-27B, a language model with 27 billion parameters — was trained to predict PHQ-9 scores, the standard 0-to-27 depression scale clinicians use worldwide. Here is what they did. They started with 3,111 real PHQ-9 scores from platform users, then used another AI — Claude Opus — to generate 'pseudo-labels', educated guesses, for a further 3,172 users, nearly doubling the training set. The final model predicted PHQ-9 scores with a Pearson correlation of 0.80 against real answers, and hit an AUC of 0.91 at the clinical threshold of PHQ-9 ≥10, the point where a clinician would consider a formal evaluation. Why does this matter? Many people in therapy never fill in a questionnaire between sessions. Passive monitoring — inferring severity from natural conversation — could catch someone sliding before their next appointment. The catch is real, though. Every conversation here was with an AI therapist, not a human one. The model learned from one platform's language patterns and hasn't been tested on face-to-face session notes, a different app, or a different language. The pseudo-labels used in training are themselves AI guesses, not clinician judgments. And there are no confidence intervals or cross-validation results in the paper. One platform, one setting — promising, not proven.

Why fancy AI keeps losing to basic math in medical data.

Across 140 medical prediction tasks, deep learning kept losing to a 1970s regression formula — and a mathematician just proved why.

A research team — whose proofs were machine-checked using the Lean proof assistant, a software that catches logical errors the way a compiler catches code bugs — showed that across 140 prediction tasks using UK Biobank health data, complex machine learning models (deep neural networks, gradient-boosted trees) match or lose to basic linear regression. Not sometimes. Consistently. The reason, they argue, is not that the fancy models are bad. It is that biomedical measurements are inherently noisy. Imagine trying to photograph someone's face through a frosted glass window. A better camera does not help — the blur is in the glass, not the lens. In medical data, every measure — a questionnaire score, a blood protein level, a wearable reading — carries error. The team's theoretical framework, which they call an excess-risk identity, shows that complex nonlinear patterns need multiple noisy inputs multiplied together, and each multiplication compounds the noise problem. Linear relationships use each feature only once, so they degrade more slowly. For mental health, this matters a lot. Depression research keeps hoping that some ML model will unlock hidden patterns in questionnaires, brain scans, or wearable sensor streams. This paper says: if your measurements are unreliable, no model rescues you. And crucially, more data does not fix it — a larger dataset just gives you a more precise estimate of the blurry picture, not the sharp one underneath. The honest limit: the result is observational and theoretical, and the team did not test every possible feature engineering approach. There will be domains where reliability is high enough for complex models to win. But biomedical tabular data — the bread and butter of depression research — is precisely the regime they studied.

Smartwatches and cycling helped veterans manage PTSD — barely enough people to tell.

Thirteen veterans, a long cycling event, and a smartwatch app acting like a check-engine light for PTSD stress spikes.

Project Hero is a real endurance-cycling program for veterans. Researchers randomized participants to one of two groups: cycle while using a digital app that detected hyperarousal — moments of acute PTSD-linked stress — in real time, or cycle with no app at all. A third, non-randomized group stayed home and wore the watch passively for comparison. The app combined heart rate and movement data to flag stress spikes, like a smoke detector that goes off not for fire but for your nervous system overheating. The digital group showed more stable symptom trajectories over the study period, tracked weekly using standard PTSD and anxiety questionnaires. The cycling-only group showed a late rise in stress scores. Both cycling groups improved acutely during the main event — consistent with what we already know about exercise and mood. The at-home group slowly declined without either support. Why does this matter? It is one of the first attempts to combine physical activity, real-time biofeedback, and a structured PTSD intervention in a setting where veterans actually want to be. The wearable was not just observing — it was trying to intervene. The catch is the numbers. Seven people in the treatment arm. Three in the cycling-only arm. Four analyzable in the home group. At this scale, one participant's bad week can look like a group trend. The authors are honest about this: it is a pilot, designed to test feasibility and hunt for signals, not to prove the intervention works. What it does is lay out a credible protocol worth testing at a proper scale — and that is a genuinely useful thing for a pilot to do.

Read these three stories together and a single tension comes into focus: we are building increasingly sophisticated tools to detect mental state from passive signals — conversation, wearable sensors, brain recordings — and yet the most rigorous paper today is the one proving those signals may not be reliable enough to matter. The LLM depression-detection work is genuinely impressive, but it lives or dies by whether the language patterns in one AI therapy platform generalize to the real clinical world. The veterans trial is honest about its own smallness. And the measurement-noise paper is a quiet warning to all of it: better models will not save you if the underlying data is blurry. The field is not moving fast in one direction — it is moving fast in two directions simultaneously, toward more ambitious tools and toward a clearer understanding of their structural limits. That tension is worth holding.

The next test for LLM-based passive depression screening is whether any of these models get validated on human-therapist transcripts in a clinical setting — watch for that replication attempt, probably from academic hospital groups rather than platform companies. On the wearable side, a properly powered PTSD trial using continuous biofeedback would be the obvious follow-up to the Project Hero pilot; the open question I'd most want answered is whether the app's real-time alerts actually change behavior in the moment, or just raise awareness after the fact.

• Wikipedia: PHQ-9 depression scale explained
• Wikipedia: UK Biobank — the dataset behind the measurement-noise paper
• Wikipedia: Post-traumatic stress disorder — background for the veterans trial