Crisis counselors face one of the hardest judgment calls in mental health care: identifying, in real time, which person reaching out is at imminent risk of suicide. Language is often the only available signal. A tool developed by researchers at MIT’s McGovern Institute for Brain Research is designed to make that signal legible at scale, by systematically scanning crisis conversations for words and phrases tied to established suicide risk factors.
How a Lexicon Becomes a Risk Assessment
The tool was built by Daniel Low, a former graduate student in Senior Research Scientist Satra Ghosh’s Senseable Intelligence Group, now a research scientist at the Child Mind Institute where he leads its AI, Risk, and Contemplative Science Lab, and also a visiting scholar at Harvard University. The core of the system is a custom lexicon: a structured list of roughly 60 words or phrases for each of 49 recognized suicide risk factors. Those risk factors span psychiatric conditions such as depression, borderline personality disorder, and post-traumatic stress disorder, as well as social and environmental stressors including poverty, incarceration, discrimination, and loneliness.
Building the lexicon was not a purely automated process. The team used a large language model to generate a preliminary list of relevant terms, then manually reviewed and refined it, with each entry confirmed by expert clinicians. The result is a vocabulary that connects everyday language to clinical categories, making it possible for a machine learning model to scan a text conversation and estimate where a person falls on a risk spectrum.
That model was trained and validated on a substantial dataset: approximately 16,000 de-identified conversations from Crisis Text Line, a global nonprofit providing free, round-the-clock, confidential text-based mental health support. Crisis Text Line provided specialized training and controlled access to this restricted dataset. Each conversation had already been assessed and grouped into one of three categories: non-suicidal, suicidal ideation without imminent risk, and imminent risk. The imminent risk group, defined as individuals with a plan for suicide or an intent to die within the next 48 hours, was the primary focus. The findings were published in the Journal of Psychopathology and Clinical Science.
What the Data Revealed About Risk
The results confirmed some clinical intuitions and complicated others. Depression is widely understood as a risk factor for suicidal ideation, but the model found that mentions of lethal means and substance use were more strongly associated with the highest-risk group than expressions of depressed mood or fatigue. References to specific means, such as “cut” or “pills,” were weighted heavily by the model. Active expressions of suicidal ideation and self-injury were also strong predictors. Anxiety, PTSD, and emotional pain emerged as intermediate predictors.
This matters because it shifts attention away from a general sense of distress toward specific behavioral and situational signals. A person who mentions hopelessness contributes to a risk score, but a person who mentions a method contributes far more. The model makes that hierarchy explicit and auditable: rather than producing a black-box risk score, it flags the specific words and phrases that drove its assessment, so a human counselor can see exactly what triggered the alert and respond accordingly.
That interpretability is not incidental. It is a deliberate design choice. The model is described by its creators as “lightweight,” capable of running on a personal computer without the computational infrastructure required by large language models. This reduces cost and, critically, limits privacy exposure. The researchers note that they often run the lexicon-based model in parallel with large language model approaches, using the simpler system to guarantee that certain high-risk terms are always flagged.
Why This Approach Signals Something Broader
Here is what most coverage of AI in mental health tends to miss: the value of this tool is not that it replaces clinical judgment. Ghosh is explicit on this point, stating that having a human in the loop will be critical for a long time given the complexity of the space. The tool’s role is to handle the volume and consistency that human attention alone cannot guarantee, surfacing patterns across thousands of conversations so that counselors can focus their judgment where it is most needed.
The research also points toward a methodological shift in how suicide risk is studied. Traditional epidemiological surveys ask people to recall their symptoms after a crisis has passed. Crisis text data captures language as the crisis unfolds. That difference in timing changes what can be learned. The lexicon is already being applied to other text sources, from social media to electronic health records, to explore how risk signals appear across different contexts.
Ghosh and Low are sharing both the suicide risk lexicon and the software package used to build it, so that researchers can construct similar tools for other mental health conditions. The infrastructure, not just the output, is being made available.
In Short
A research team at MIT’s McGovern Institute built a text-based tool that scans crisis conversations for language tied to 49 suicide risk factors, trained on roughly 16,000 real conversations from Crisis Text Line. The tool found that mentions of lethal means and substance use are stronger predictors of imminent risk than depression alone. It is deliberately simple, interpretable, and designed to support human counselors rather than replace their judgment. The lexicon and the software to build it are being shared openly, with potential applications extending well beyond crisis lines.
Based on reporting from MIT News AI.