Your Phone Knows Your Mood Before You Do

A Phone That Predicts When Cancer Survivors Need Emotional Help

Cancer survivors often stop filling in mood diaries at exactly the moments they're struggling most — so what if the phone could notice instead?

There's a cruel irony in asking distressed people to report their distress. Researchers call it the diary paradox: the people who most need to flag that they're struggling are the least likely to open an app and type it in. A team building a system called PULSE tried to get around this by watching the phone instead of asking the person. Think of it like a building manager who doesn't knock on your door to ask how you're doing, but checks whether the lights are on at odd hours, whether the front door opened, whether the heating system is running normally. PULSE monitors movement, location, screen usage, sleep rhythms, and social communication patterns — all passively, without the person having to do anything. The clever part is the reasoning layer on top. Instead of applying a fixed formula to the sensor data, the system uses a large language model — the same family of AI behind chatbots — equipped with eight purpose-built tools to investigate each case like a detective: query the data, compare it to the person's own baseline, check population-level patterns, then make a call. Tested on 50 cancer survivors, the system predicted with 74.3% balanced accuracy when someone wanted help regulating their emotions — a meaningful jump over the 52–60% ceiling traditional machine learning had hit on the same type of task. The biggest driver of improvement wasn't the richer data; it was the agentic reasoning. The catch is real: 50 people is a small group, all cancer survivors, all from one study. We don't know whether this holds for other conditions, other populations, or whether acting on these predictions actually helps anyone. It's a proof of concept, not a deployed product.

Depression Screening AI That Stops Leaking Your Age and Gender

Say ten words into a mental health app and the AI can probably guess your gender with 92% accuracy — even if you never said who you are.

Voice is extraordinarily leaky. The pitch, rhythm, and texture of how you speak reveal your sex, rough age, probably your emotional state, and potentially much more. When a mental health app uses your voice to screen for depression, it doesn't just get the depression signal — it gets a demographic profile of you as a side effect. That's a privacy problem nobody has cleanly solved. A team at this year's preprint sprint tackled it with a system called InfoShield. The idea is like noise-cancelling headphones: you want to hear the conversation clearly, but you want the traffic noise stripped out. Here, the 'conversation' is the depression signal and the 'traffic' is your gender and age. InfoShield does this by compressing the voice representation down to the minimum needed for depression detection, then actively penalising the system every time it can still infer demographic information. The penalty is mathematically defined using a concept called mutual information — essentially, how much knowing the voice representation tells you about gender or age. The numbers are striking. Gender inference accuracy dropped from 92.6% to 55.5% — near-random for a binary guess. Age inference dropped from 55.7% to 30.3%. Meanwhile, depression classification held up: F1 of 0.784, actually better than the previous state of the art at 0.723, with only a 6% cost from the privacy constraints. The honest limit: this was tested on a single dataset called the Androids Corpus, and the sample size isn't reported in the paper. One dataset, one team, no external validation yet. Promising engineering, but a long road to clinical trust.

AI Misses Anxiety Diagnoses When You Mention You Have Good Friends

An AI reading your mental health interview might look straight at your anxiety symptoms — and still miss the diagnosis because you mentioned you have a supportive family.

A team built a benchmark of 555 real semi-structured mental health interviews, each tagged with clinician-verified diagnoses using the SCID — the gold-standard structured interview psychiatrists use. Then they fed those interviews to five large language models and asked each one to screen for anxiety disorder, major depressive disorder, PTSD, and any mental health condition. The results were sobering. Accuracy ranged from 0.49 to 0.86 depending on the model and the task — that lower end is basically a coin flip. The best performers were GPT-4.1 Mini and GPT-5 Mini, but even they showed a pattern that should give you pause. When the researchers dug into the cases the AI got wrong, they found a recurring fingerprint: the false negatives — cases where someone had clear symptoms but the AI said 'no diagnosis' — almost always contained what they call protective-context language. The person mentioned good coping skills. Or a supportive social network. Or that they were still functioning at work. The AI apparently treated these as evidence against a diagnosis, even when the clinical symptoms were sitting right there in the same interview. Think of it like a smoke detector that stops alarming when you crack a window open. The ventilation might help, but the fire is still real. A human clinician knows to hold both facts simultaneously — this person has symptoms AND protective factors. The AI, at least in these zero-shot tests, couldn't. One important caveat: these models weren't fine-tuned for psychiatry. They were tested as-is, off the shelf. A trained clinical AI might behave differently. But given how many companies are already deploying these tools, the off-the-shelf behaviour matters.

Three papers, three different angles on the same project: teaching machines to recognise human distress. And taken together, they sketch something more honest than the usual AI-in-mental-health narrative. The PULSE work says passive sensing can catch what self-reporting misses — that's genuinely useful. InfoShield says we can scrub demographic leakage from voice data without losing diagnostic signal — that's a real engineering advance. The LLM screening paper says the models we're already deploying make a specific, human-shaped mistake: they're too impressed by coping language. Notice what connects all three. None of them are ready for unsupervised clinical use. All three are honest about that. What they collectively suggest is that AI in mental health is becoming more precise about specific, narrow tasks — predicting one emotion, stripping one attribute, reading one transcript — while the question of how you safely string those pieces together into something a real person can trust remains wide open. That gap is where the real work lives.

The LLM screening bias finding deserves a follow-up with fine-tuned clinical models — if someone runs that experiment, it'll tell us whether this is a fundamental reasoning failure or a training data problem. On the passive sensing side, watch for whether PULSE or similar systems move into a randomised trial: predicting distress is one thing, but whether acting on those predictions improves outcomes is a completely different question nobody has answered yet.

• Wikipedia: Passive sensing in mobile health research
• Wikipedia: Structured Clinical Interview for DSM Disorders (SCID)
• Scientific American: The privacy risks of mental health apps