Your voice, a fake waiting room, and the scroll-anxiety link

An AI listened to 34,000 people's voices and learned to spot depression

You can hear when a friend is off — flatter tone, slower rhythm, something you can't name — and it turns out a machine can too.

A team trained deep learning models on 863 hours of recorded speech from roughly 34,000 people across the United States. The goal was to detect depression and anxiety not from what people said, but from the hidden texture of how they said it. Think of it the way you can tell someone is exhausted the moment they pick up the phone — before a single meaningful word — just from the weight in their voice. The model (built on Whisper, the same backbone behind automatic transcription services) was fed raw audio, no transcripts. It learned to pull out acoustic patterns that track with how people scored on two standard self-report questionnaires — PHQ-9 for depression and GAD-7 for anxiety. A second layer added language features: what was actually said, processed by a BERT-style model. Combined, the system reached 71% on both sensitivity and specificity — meaning roughly 7 in 10 people with elevated symptoms were correctly flagged, and 7 in 10 without symptoms were correctly cleared — tested on about 5,000 held-out subjects. A predecessor paper from this group (Mazur et al., 2025, Annals of Family Medicine) already cleared peer review, which gives this work some credibility lineage. That said: 71% means 1 in 3 calls is still wrong. The labels come from self-reported questionnaires, not clinical diagnosis. The training data is proprietary and U.S.-based, so performance across languages, accents, or different healthcare contexts is unknown. And a screening tool that misses 29% of people in distress can cause real harm if deployed without a clinician in the loop. This is a meaningful step — not a finished product.

AI writes convincing fake patients — but gets the whole crowd wrong

Every AI-generated patient file looked clinically real — but the fake waiting room looked nothing like an actual one.

Researchers generated 28,800 simulated psychiatric patient profiles using four major AI systems — GPT-4o-mini, DeepSeek-V3, Gemini-3-Flash, and GLM-4.7 — across 120 demographic combinations (different races, income levels, gender identities, relationship statuses). They then compared those fake populations against real U.S. health survey data from NHANES and NESARC-III. The individual-level result was striking in one direction: zero of the 28,800 simulated patients violated the basic clinical rules of how depression symptoms cluster together. Every fake file looked plausible. But zoom out to the full population, and things fall apart. Think of it like a novelist who can write any single convincing character but, asked to fill a stadium, produces thousands of near-identical extras — no extremes, no outliers, no genuine variety. The study measures this as variance compression: DeepSeek-V3 flattened the real-world spread of depression severity by 62%. And 37% of simulated patients would flip their diagnosis — depressed or not — if you ran the same AI twice on the same prompt. The demographic failures are specific and serious: depression scores for most groups were overestimated by 3 to 6 points on a 27-point scale, while transgender women's scores were underestimated by more than 5 points. The bias patterns were consistent across both U.S.-built and Chinese-built models. Why does this matter to you? Because mental health researchers and companies are increasingly using AI-generated patient profiles to test and train new tools. If the simulated crowd is a caricature of the real one, every conclusion drawn from it is built on sand.

Social media and anxiety: a real but weak link, honestly measured

551 people, six patterns, and one honest finding: the link between scrolling and anxiety exists — but it's smaller than your feed would have you believe.

Researchers surveyed 551 people about their social media habits and psychological well-being, then used a technique called K-Means clustering to let the data self-organize into groups. Imagine sorting a bag of mixed buttons by color and size without deciding in advance how many categories to use — you just let natural groupings emerge. Six clusters appeared, running from younger heavy users who never stepped away from their feeds to older moderate users with calmer habits. The headline number: a correlation of 0.28 between daily social media hours and self-reported anxiety scores. One cluster stood out as the most distinct — younger users with high personal use, 68% of whom regularly took breaks from their feeds — and this group appeared to report fewer negative effects than similarly heavy users who never paused. Now, the catch — and here it matters a lot. A correlation of 0.28 means social media hours explain roughly 8% of the variation in anxiety. The other 92% comes from everything else in someone's life — sleep, relationships, finances, genetics. The cluster quality score (Silhouette Score of 0.32, on a 0-to-1 scale) indicates the six groups blur significantly at their edges; this isn't six clearly distinct human types. Most critically, this is a one-time snapshot of 551 people, not a follow-up study. We cannot say from this data that social media causes anxiety. What we can say, modestly and honestly: heavier use without breaks appears linked to higher reported anxiety in this sample. That's a reasonable signal worth pursuing in better-designed, longer studies. It's not a verdict.

Pull back, and today's three stories are really one story told three ways: we are getting better at measuring mental health from a distance — through voice, through text, through survey patterns — and the tools are improving fast enough that the gaps are now visible, which is progress. The voice biomarker work shows real detection signal in audio alone. The social media study shows a real, if modest, anxiety correlation. But PsychBench throws cold water on the assumption that AI already understands human mental health populations well enough to simulate them safely. That's the uncomfortable centre: detection is advancing faster than understanding, and the tools being built on top of that detection are already being tested on AI-generated populations that don't reflect reality. The next hard problem isn't building a model that spots depression in a voice clip. It's knowing whether the populations those models were trained and tested on look anything like the people who will actually use them. Right now, the evidence says they often don't.

The voice biomarker team's prior paper (Mazur et al., 2025) is already in Annals of Family Medicine — watch for clinical responses to that publication, which will signal whether the medical community is ready to take voice-based screening seriously as a triage tool. On the simulation side, the PsychBench methodology is now public; I'd expect other groups to run the same audit on newer models within weeks. The open question I want answered: does variance compression get better or worse as models scale up?

• Wikipedia: PHQ-9 (Patient Health Questionnaire)
• Scientific American: Can AI Detect Mental Illness?
• Wikipedia: K-means clustering