Your Voice, Your Phone, Your Symptoms — Who Hears What?

Screening for Depression by Voice, Without Leaking Your Identity

Your voice can betray your age and gender to an algorithm in under a second — and that's a problem if the same algorithm is also screening you for depression.

Here is the tension. Researchers have spent years building tools that listen to your voice and look for signs of depression — tremor in the pitch, irregular rhythm, flattened emotion. These tools work reasonably well. But the same acoustic features that hint at your mental state also carry a lot of other personal information: your age, your gender, your likely ethnicity. Feed a recording into a standard model and it will classify your gender correctly about 93% of the time, almost as a side effect. The InfoShield team built a kind of frosted-glass layer between your voice and the algorithm. Think of frosted glass on a bathroom window: someone outside can tell there's a person in there, but they cannot make out your face. InfoShield does something mathematically similar — it keeps the acoustic signal that carries depression-relevant information while actively scrubbing out the patterns that reveal who you are. The technical name for the scrubbing process is mutual information minimization, which just means: measure how much the processed signal still 'leaks' about your age or gender, then train the system to leak as little as possible. The results are real. Gender inference accuracy dropped from 92.6% to 55.5% — essentially chance for a binary guess. Age inference dropped from 55.7% to 30.3%. And depression classification actually improved slightly compared to the previous best published system, hitting an F1 score of 0.784 against a prior benchmark of 0.723. The catch: this was tested on a single dataset, the Androids Corpus. One dataset is a starting point, not a proof. Privacy protection that holds on one population may crack on another. And 0.784 F1 is promising but nowhere near the reliability you'd need before deploying this in any clinical setting. A small but real step.

When AI Misses Anxiety Because You Seem Like You're Coping

An AI reads a transcript packed with anxiety symptoms — and still marks the person as probably fine, because they mentioned they have good friends and still go to work.

A team of researchers fed 555 real psychiatric interview transcripts into five different large language models — including GPT-4o Mini and two versions of GPT-5 Mini — and asked each one to screen for anxiety, depression, PTSD, and general mental health disorders. No training, no fine-tuning: the models just read the transcripts and gave a verdict, the same way you might hand a document to someone and ask for their opinion. The overall accuracy ranged from 0.49 to 0.86 depending on the model and the condition. But the number that caught my attention sits inside that range: the Matthews correlation coefficient — a stricter measure of whether the model is genuinely tracking the disorder rather than just guessing the common answer — ran from 0.16 to 0.38. That is weak to modest at best. These models are doing something real, but not reliably enough to trust. Think of a smoke alarm that keeps quiet when the window is open. The smoke is still in the room. But a ventilation signal confuses the detector into inaction. That is roughly what happened with anxiety and PTSD false negatives: the interview transcripts contained clear symptom language, but they also mentioned preserved functioning, social support, and coping ability. The models consistently weighted those protective-context words heavily enough to override the symptom evidence and issue a negative classification. This matters enormously if you are imagining LLMs as triage tools, because the people most able to articulate their coping strategies are not necessarily the people least in need of help. Depression classification also showed higher accuracy for male than female participants — a bias the authors flag without yet resolving. The catch: the rationale text the models produced was analyzed to explain the false negatives, but the researchers are careful to note that what a model writes in its explanation does not necessarily reflect how it actually reached its decision. Honest, that.

Your Phone's Passive Sensors May Predict Your Anxiety Score

What if predicting your anxiety level required no questionnaire — just the motion sensor and sleep log your phone already records?

Your phone knows things you have not told anyone. It knows you woke at 3 a.m. four nights this week. It knows your steps dropped by half on Tuesday. It knows your screen-on time spiked at midnight. These are passive sensing signals — data collected without you actively doing anything — and researchers have been trying to turn them into mental health predictions for years. The problem is that sensor data from one research study rarely transfers well to another: different phones, different study populations, different definitions of 'a good week.' The TimeSRL team tackled that transfer problem with a two-stage approach. Stage one: take the raw numbers and convert them into natural language descriptions — something like 'this person had fragmented sleep and low activity mid-week, with a sharp rebound on the weekend.' Stage two: feed those text descriptions into a language model that predicts the person's score on a standard anxiety scale. The key insight is that natural language might travel across datasets better than raw sensor values, the way a description of someone's energy travels better than a spreadsheet of their step counts. To test this, the team ran a leave-one-dataset-out protocol: train on multiple datasets, predict on the one you left out. That is the hard version of the test, the one that mimics what would happen in a real deployment. Against non-LLM machine learning baselines, TimeSRL reduced prediction error for anxiety by 3.1 to 10.1 percent. Against other LLM-based approaches, the reduction was 9.5 to 44.1 percent. Depression results were similar. The catch: these are still prediction errors on research datasets, not clinical trials. The model tells you someone's probable score on a questionnaire — it does not diagnose them, and it does not tell a clinician what to do next. That gap is large and mostly unsolved.

Step back and look at what these three papers are collectively doing. They are all trying to replace or augment the most fragile part of mental health care: the moment when someone has to accurately describe how they feel, in words, to another person or a form. TimeSRL listens to your phone. InfoShield listens to your voice. The LLM screening paper tries to read your interview transcript. None of them asks you to introspect perfectly. But each one hits the same wall from a different direction. TimeSRL's signal is real but refuses to generalise cleanly. InfoShield's privacy protection works on one dataset and may not hold on the next. The LLM screener misses exactly the people who are distressed but present as coping. The field is not failing — these are genuine advances. But the shared lesson is that passive, automated detection keeps bumping into the irreducible complexity of the person on the other end. The algorithm sees the pattern. It still cannot see you.

The LLM screening paper used GPT-5 Mini, which is barely weeks old as a public tool — it would be worth watching whether any team runs a systematic replication as access widens. More broadly, the FDA's Digital Health Center of Excellence has been slowly developing guidance on AI-based psychiatric tools; any updated framework there would immediately change what any of these three approaches can legally claim to do. The open question I'd most want answered: does InfoShield's privacy protection hold when you move from the Androids Corpus to a clinical population in a different country?

• Wikipedia: Computational psychiatry — background on the broader field these papers sit in
• Wikipedia: Passive sensing in mental health research
• Nature News: Can AI detect mental illness from your social media posts? (2023)