Your smartphone knows more about your mental health than you might think. GPS traces, call patterns, typing speed, sleep schedules, the sound of your voice, the words you choose — all of these can signal a shift in mood, sometimes before you notice it yourself. The science of reading those signals is called digital phenotyping, and it now spans two channels: what your behavior reveals and what your conversations reveal.

What digital phenotyping means

The term was coined by Thomas Insel, former director of the U.S. National Institute of Mental Health, in a 2017 paper in JAMA (318:1215–1216). The idea is straightforward: the phone you carry everywhere accumulates vast amounts of behavioral data, and if we learn to read it correctly, we can build a continuous portrait of a person's mental state — far more informative than a checkup once every three months.

Insel identified three types of data. First, sensor data — GPS location, accelerometry, and the timing of activity. Second, voice and speech — sentiment and prosody, the melody of how someone talks. Third, human–computer interaction — the latency between keystrokes, scrolling intervals, how long a screen holds attention. Each channel is a window onto behavior that a person would never think to report in a clinic.

What makes this powerful is that the signals move ahead of self-awareness. In a pilot study of 17 patients with schizophrenia, Barnett, Torous, and colleagues (2018, Neuropsychopharmacology, 43:1660–1666) found statistically significant anomalies in mobility patterns several days before a relapse. The phone flagged an approaching crisis before the person — or their doctor — had noticed anything wrong. Insel later reached for a metaphor from another field of medicine: continuous mental-health monitoring, he suggested, could work like a "smoke alarm" (2018, World Psychiatry), an early warning rather than a diagnosis after the fact.

Specialized platforms make this research possible. Beiwe, developed by John Torous and colleagues at Harvard Medical School, collects GPS, accelerometry, call and message logs, and voice samples while remaining HIPAA-compliant; at scale its cost falls from roughly $80 to about $3.50 per subject per month. A sister platform, mindLAMP (Beth Israel Deaconess/Harvard), combines this passive stream with active check-ins such as surveys and cognitive tests. A 2025 systematic review by Linardon and colleagues of 112 studies found that GPS is the most-used sensor, that the average study runs 14.3 weeks, and that most research is done on Android — a snapshot of a field that is real but still young.

One subtle finding shapes how all of this is analyzed. Torous and colleagues (2021, World Psychiatry, 20:318–335) showed that variational measures — the entropy of a routine, or how far today deviates from a person's own personal average — are more informative than absolute sensor values. It is not that you visited two places today; it is that you usually visit six. Mental state shows up in the change from your own baseline, not in a universal threshold.

What your conversations reveal

Behavior is only half the picture. Depression also changes the way people speak and write, and a second research thread reads that channel directly.

The voice is a surprisingly rich marker. Depression tends to flatten it: speech becomes more monotone, articulation slows, and the pauses between words stretch. Briganti and Lechien (2025, Journal of Voice) reviewed 12 studies covering 16,872 participants and found that vocal biomarkers distinguish depression from a healthy baseline with an AUC ranging from 0.71 to 0.93 — from modest to strong, depending on the method and the population. One commercial tool, Kintsugi Voice, identifies depression or anxiety from just 20 seconds of free speech with roughly 80% accuracy (Mazur and colleagues, 2025, Annals of Family Medicine, 23:60–65).

Text carries its own signals. A long series of research competitions — the AVEC (Audio/Visual Emotion Challenge, 2013–2019) — drove much of the work on multimodal depression detection, and its DAIC-WOZ dataset, recordings of clinical interviews with a virtual agent, became the standard benchmark. On the language side, Tauscher and colleagues (2023, Psychiatric Services) applied a fine-tuned BERT model to text messages between patients with serious mental illness and their clinicians, reaching an F1 of 0.62 — essentially matching human raters at 0.63.

Depression also shows on the face. Reduced variability of facial expression is one of its recognized behavioral markers, and the AVEC challenges deliberately paired audio with video precisely because no single channel tells the whole story. The most robust research systems fuse several at once — voice, words, and facial movement — on the logic that a marker missed by one modality is often caught by another. That redundancy is also what makes these systems hard to fool by accident: a flat voice on a bad day means less when the language and expression look ordinary.

Language assessment can even happen live, inside an ordinary chat. A system called MoPHES, described in IEEE in 2025, pairs two compact 0.5-billion-parameter models: one silently rates anxiety and depression every few turns, the other carries the conversation informed by that rating. It reached 80.5% accuracy for anxiety and 63% for depression while running entirely on the user's device — notably outperforming models many times its size. The depression figure is lower for a reason that recurs across this whole field: apathy, slowing, and avoidance are harder to hear in words than the fear and tension of anxiety.

Why does reading state from conversation matter so much? According to the WHO's 2022 World Mental Health Report, roughly one billion people live with a mental disorder — about 13% of the world — and more than 70% never receive effective care. Part of that gap is a shortage of specialists, but part is that many people simply don't recognize they need help or can't gauge how serious their state is. A tool that assesses mood from an ordinary conversation changes the point of entry: no questionnaire, no appointment, no need to first admit that something is wrong. You just talk, and the assessment rides along quietly with the dialogue.

Sensors vs. words: what works better

Neither channel is strictly better; they see different things at different costs.

Signal type	What it detects	Evidence strength	Privacy cost
GPS / mobility	Withdrawal, routine collapse, relapse risk	Days-ahead relapse signal in small cohorts; needs personal baseline	High — continuous location history
Typing / interaction	Sleep disruption, agitation, slowing	Emerging, less validated	Moderate — behavioral metadata
Voice / prosody	Depression, anxiety from speech quality	AUC 0.71–0.93 across 16,872 people	Moderate — audio is identifying
Text / language	Mood, cognitive distortions, distress	BERT F1 ≈ human raters (0.62 vs 0.63)	High — content of private messages

Sensor data is passive and continuous but blunt: it tells you that something changed, rarely what. Language is sharper about content and emotion but sparser — it only exists when someone speaks or writes. The most convincing systems combine them, and they lean on the variational insight above: the goal is to detect a departure from a person's own normal, not to match a population-wide cutoff. To understand the models that turn these raw signals into a picture of mental state, see our overview of computational models of mental disorders; for the specific, higher-stakes case of spotting crisis language, see how AI detects suicide risk in text.

Limits and ethics

Passive measurement raises hard questions, and honesty about them is part of doing it responsibly.

The first is false positives. A detector that flags an approaching low mood will sometimes be wrong, and every wrong alarm carries a cost — needless worry, eroded trust, or an intervention no one wanted. Sensitivity and specificity trade off against each other; there is no setting that catches everything and cries wolf never.

The second is privacy. GPS traces, communication patterns, sleep, and social activity are among the most sensitive data a person produces, and the content of messages or the sound of a voice is directly identifying. This is why the location of processing matters so much: a model that runs on your device and keeps data local has a far smaller breach surface than one streaming everything to a server. We go deeper into that trade-off in on-device AI therapy and privacy. Researchers are also developing "dynamic consent" models that let people choose, moment to moment, which data they are willing to share.

The third is the line between monitoring and diagnosis. Digital phenotyping is a monitoring tool. It can tell a clinician that something deserves attention; it does not, on its own, make a diagnosis, and screening-grade accuracy is not clinical-grade accuracy. A responsible product treats a signal as a prompt to check in and, where needed, to route someone toward real help — never as a verdict. Our piece on guardrails for mental-health AI covers what that responsible handling looks like in practice.

Digital health technologies are advancing quickly, and services like Nearby use evidence-based approaches to support mental health. Digital phenotyping is still making its way from the lab to everyday practice, but the tools built on cognitive behavioral therapy are already available to anyone — right on the phone that carries all these signals.

FAQ

Can a smartphone really detect depression?

It can detect signals associated with depression — changes in movement, sleep, voice, and language that shift when mood shifts. In research, mobility anomalies have preceded relapse by days, and vocal biomarkers separate depression from a healthy baseline with an AUC as high as 0.93. But detecting a signal is not the same as diagnosing an illness. A phone can raise a flag; a clinician makes the call.

Does AI read my messages to assess my mood?

Some systems analyze language, and the best of them are designed so this happens on your device rather than on a company's servers. In the MoPHES architecture, for example, assessment runs locally and nothing is sent to the cloud. Where and how your text is processed is exactly the question to ask any app that offers this — the answer determines both the privacy risk and, often, whether you can use it offline.

How accurate is digital phenotyping?

It depends on the signal. Voice biomarkers reach an AUC of 0.71–0.93; text models can match human raters (F1 around 0.62); on-device conversation models reach roughly 80% for anxiety but lower for depression, which is harder to read in words. Across the board, accuracy improves when a model compares you to your own baseline rather than to a fixed threshold. This is promising screening technology, not a substitute for professional evaluation.

Nearby is a support tool that uses evidence-based psychology. It does not replace a psychologist, psychotherapist, psychiatrist, or emergency service.