If you are thinking about suicide or self-harm right now, please reach out to a human. Contact your local emergency services or a crisis line immediately, and tell someone you trust. This article is about the technology of risk detection — it is not a crisis service and cannot help you in an emergency. Trained people can, and they are available.

Psychiatrists have long known an uncomfortable truth: traditional suicide risk scales perform only marginally better than chance. A meta-analysis of 365 studies across 50 years (Franklin et al., 2017) found that the predictive power of classical risk factors sits near AUC 0.58 — nearly useless for real clinical decisions. That failure is precisely what pushed researchers toward machine learning and natural language processing.

How text reveals suicide risk

Suicidal thoughts leave traces not so much in the words "I want to die" as in the structure of speech. Studies by John Pestian's group at Cincinnati Children's Hospital showed that models trained on interview transcripts distinguish suicidal from non-suicidal adolescents with roughly 85% accuracy — not by relying on direct statements, but on patterns: reduced cognitive complexity, a rise in absolutist phrasing ("always," "never"), a narrowing time horizon, a shift of pronouns toward "I" combined with emotional dissociation.

Al-Mosaiwi and Johnstone (2018) analyzed over 6,400 posts on English-language forums and found that the share of absolutist words in depression and anxiety communities was 50% higher than in controls — and 80% higher in communities focused on suicidal ideation. This is the kind of signal hard to catch by ear, but easy to measure statistically. Related NLP work on clinical text shows how far this has come: Tauscher and colleagues (2023, Psychiatric Services) applied a fine-tuned BERT model to messages between patients with serious mental illness and their clinicians and reached an F1 of 0.62 — essentially matching human raters at 0.63.

At scale, the numbers get more striking. Walsh, Ribeiro, and Franklin (2017) trained a model on the electronic health records of 5,167 patients and achieved AUC 0.84 for predicting a suicide attempt within the next 7 days — far above any clinical scale. Similar results come from social-media data: the annual CLPsych shared tasks use Reddit posts (the SuicideWatch subreddit) as a labeled corpus, with the best systems reaching F1 scores of 0.55–0.60 on risk-level classification. Since 2017, Facebook has deployed a system that detects suicidal signals in posts and live streams; by the company's own reporting, it triggered more than 3,500 wellness checks in its first year, and Instagram and TikTok have rolled out similar algorithms. In 2023, JAMA Psychiatry published a systematic review of 54 ML studies: the mean AUC was 0.81, making NLP the most accurate known method for short-horizon prediction.

From detection to intervention

Detecting risk is only useful if something helpful happens next, and that "next" is its own field of research. The framework is called a just-in-time adaptive intervention (JITAI), defined by Nahum-Shani and colleagues (2018, Annals of Behavioral Medicine): an intervention designed to deliver the right type and amount of support at the right moment by adapting to a person's changing state. A JITAI has six components — a distal outcome, a proximal outcome, decision points, intervention options, tailoring variables, and decision rules — which together form a closed loop: sense the signal, decide, act, observe the result, adjust.

Thomas Insel captured the ambition with a metaphor: continuous monitoring as a "smoke alarm" for mental health — an early warning that gives someone time to act. But the honest state of the evidence is modest. A 2025 meta-analysis in BMJ Open (K = 23 studies, N = 2,563) found only a small effect for JITAIs (g = 0.15), and a systematic review in Frontiers in Digital Health (2025) counted just five fully implemented mental-health JITAIs to date. The loop is real and promising, but it is early, and the leap from "we detected something" to "we helped" is exactly where the science is thinnest. We go deeper into the closed-loop idea in our piece on just-in-time interventions in AI crisis support.

False positives and the cost of errors

High accuracy is only half the story. The base rate of suicide attempts is so low that even a model with 90% sensitivity and 90% specificity will produce dozens of false positives for every true case in the population. This isn't a flaw of the algorithm — it's the mathematics of rare events, and no amount of model tuning makes it disappear. Every system that flags risk must choose a point on the sensitivity–specificity trade-off, and both directions cost something: miss real crises, or drown real ones in false alarms.

A quick illustration makes the problem concrete. Suppose 1 in 1,000 people in a given week is at genuine acute risk. Run that model on 100,000 people and it correctly flags most of the ~100 true cases — but 10% of the ~99,900 others get flagged too, producing nearly 10,000 false alarms. For every real crisis, roughly a hundred people are wrongly labeled. That ratio is why an automated, high-stakes response to a raw flag is so dangerous, and why the threshold you choose is an ethical decision, not just a technical one.

From this flow practical harms. First, stigma: a false "high risk" label in a health record can affect insurance, employment, or parental rights. Second, cultural blind spots: nearly all training corpora come from English-speaking patients in the US and UK, and models transfer poorly to other languages and cultural idioms of distress. Third, distribution shift: patterns of expression change over time, and a model trained in 2019 may be miscalibrated by 2024.

There is also a deeper question that no accuracy figure answers: even a perfect detector doesn't decide what to do with the signal. Dispatch emergency services without consent? Show a banner with a helpline number? Notify a loved one? Each choice carries its own ethical cost, and research on which interventions actually reduce risk after detection is still scarce. A detector that fires without a careful, humane response plan can do more harm than good.

What responsible products do

When a tool works with someone in a vulnerable state, risk detection isn't a feature you switch on and forget. It's an obligation: to listen more carefully, respond more cautiously, acknowledge the limits of your own competence, and hand the person off to trained humans when the signals cross a threshold. A responsible AI companion doesn't compete with a crisis line — it helps someone reach one in time, and it is transparent about being software rather than a clinician. The broader set of safety mechanisms that separate a careful product from a raw chatbot — crisis detection, human escalation, honest framing — is the subject of our guide to guardrails for mental-health AI.

Some of the strongest evidence points toward AI assisting human responders rather than replacing them. The HAILEY system (Sharma et al., 2020, 2023) sits in the loop of a peer-support platform, suggesting edits to volunteers' messages instead of speaking to people in distress directly. In a randomized study it raised counselors' expressed empathy by 19.6% overall, and by 38.9% among those who were struggling most to respond well. That is the pattern worth aiming for: the machine catches signals and sharpens the human response, while a person keeps their hand on the decision. Detection without that human anchor is the part that goes wrong.

The technology can catch what escapes the person themselves. But what to do with what's caught remains a decision in which a human must take part.

FAQ

Can AI detect suicidal thoughts?

To a meaningful degree, yes — better than traditional questionnaires. NLP models pick up on structural patterns in language (absolutist words, a narrowing time horizon, pronoun shifts) and reach a mean AUC around 0.81 across dozens of studies, versus roughly 0.58 for classical risk scales. But "detect a statistical signal" is not the same as "know that a specific person is in danger." These are screening tools that raise a flag, not oracles that make a diagnosis.

What happens when an AI flags crisis risk?

That depends entirely on the product, and it's the most important question to ask. A responsible system treats a flag as a prompt to respond more carefully and to surface real help — a crisis line, an emergency number, a trusted person — not as grounds to act unilaterally. Because false positives vastly outnumber true crises at the population level, an automated response like dispatching services without consent can itself cause harm. The right design keeps a human in the loop and points the person toward trained help.

A reminder: if you or someone you know is in crisis, contact local emergency services or a crisis line now. AI can help notice a problem — it cannot be the help itself. Reaching a trained human is what matters most.

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