MIT Team Builds a Lightweight, Auditable Model for Suicide Risk in Crisis Texts

Researchers at MIT's McGovern Institute for Brain Research say they have built a text-analysis tool that estimates suicide risk from conversations with crisis counselors, and that it predicts risk severity in conversations it has not seen before. The approach is notable less for raw capability than for its architecture: a large language model was used to draft the vocabulary, but the model doing the prediction is small enough to run on a laptop and reports which words drove each score. The work was published in the Journal of Psychopathology and Clinical Science.

For anyone building AI into high-stakes health workflows, the design choices here are the story. The team explicitly traded some modeling power for interpretability, privacy and cost, and framed the result as a research instrument that still needs validation before it goes near clinical use.

What the team built

The tool was developed by Daniel Low, a former graduate student in Senior Research Scientist Satra Ghosh's Senseable Intelligence Group. Low is now a research scientist at the Child Mind Institute, where he leads its AI, Risk, and Contemplative Science Lab, and a visiting scholar at Harvard University.

The core artifact is a lexicon. According to MIT, the researchers first used AI to generate a preliminary list of words and phrases tied to established suicide risk factors, spanning factors linked to suicidal ideation, suicide attempt and suicide death. They then reviewed and curated that list by hand. The final version contains roughly 60 words or phrases for each of 49 risk factors, with the relevance of each item confirmed by expert clinicians. The paper's title indicates the goal was content validity, meaning the lexicon should actually cover the construct it claims to measure rather than whatever happens to correlate in a given dataset.

A machine learning model then scans conversations for lexicon terms and uses them to estimate risk. Because every term maps back to a named risk factor, the model's weights can be read as statements about which factors matter. The specific model class, feature representation and performance metrics were not described in the MIT account.

The dataset, and why it is unusual

The team worked with Crisis Text Line, a global nonprofit whose trained volunteers provide free, confidential, around-the-clock text-based support. Crisis Text Line provided specialized training and controlled access to a restricted dataset, and the researchers analyzed de-identified texts from approximately 16,000 conversations.

Crisis Text Line's own assessments sorted those conversations into three levels: non-suicidal, suicidal ideation without imminent risk, and imminent risk, defined as having a plan for suicide or intent to die within 48 hours. That third group was the target.

Low's argument for the data is methodological. Most prior work on which risk factors matter has relied on epidemiological surveys that ask people to recall symptoms and experiences, often after a crisis has passed. Crisis line transcripts capture language as the crisis is happening.

That framing also flags the main dependency: the labels are Crisis Text Line's assessments, not verified outcomes such as attempts or deaths. The model is learning to reproduce a counselor-assigned severity category. That is a reasonable and possibly the only practical target at this scale, but it is a different thing from predicting behavior, and the MIT description does not claim otherwise.

What the model surfaced

The findings were broadly consistent with earlier research, though MIT notes they are not always intuitive. Depression is a well-established risk factor for suicidal ideation, but the model found that mentions of lethal means and of substance use were more characteristic of the highest-risk group than depressed mood or fatigue. Expressions of active suicidal ideation and self-injury were also strong predictors. Anxiety, post-traumatic stress disorder and emotional pain landed in the intermediate tier.

The weighting is legible by design. Terms referring to lethal means, such as "cut" or "pills," carry heavy weight. Hopelessness language such as "don't know what to do" or "hopeless" contributes less.

There is an obvious caveat that follows from the design: a lexicon that rewards explicit mentions of means and method will separate people who name them from people who do not. Whether that distinction generalizes across populations, cultures and channels is an empirical question the announcement does not answer.

Low describes the broader problem as underdetermined. Dozens of risk factors interact in ways that are not well understood, and even trained clinicians struggle to identify who among those with suicidal ideation will attempt suicide. Psychiatric conditions including depression, borderline personality disorder and PTSD sit alongside environmental and social stressors such as poverty, incarceration, discrimination and loneliness.

The case for the small model

The researchers are candid about what a lexicon cannot do. It ignores context, and it misses terms that resemble entries in the list without matching them. Large language models can reason over context, and Low's group has used LLMs to detect suicide risk in other projects.

They still run the lexicon in parallel, for two reasons: it guarantees that specific terms get flagged, and it keeps data local. The prediction model is lightweight enough to run on a personal computer, which cuts cost and reduces the privacy exposure of sending crisis transcripts to an external inference endpoint. That is a real consideration for any organization handling protected mental health data.

Interpretability is the other argument. Rather than emitting a score alone, the tool flags the words behind it, so a human reviewer can see the basis for an assessment and act on it. The team is pursuing similar explainability methods for LLM-based approaches.

Ghosh, who directs the Open Data in Neuroscience Initiative at the McGovern Institute, says keeping a human in the loop will be necessary for a long time. The researchers add that any predictive model must be thoroughly validated before clinical use and may need continual refinement as language use and target populations shift. That last point is a maintenance burden many deployment plans underestimate: a lexicon built on 2020s crisis-line slang will drift.

The team is releasing both the suicide risk lexicon and the software package used to build it, so other researchers can construct lexicons for other mental health conditions. Licensing terms and release timing were not specified. Low says the lexicon is already being applied to test whether text from sources including social media and electronic health records can help estimate risk, which raises questions about consent and secondary use that the announcement does not address.

What to watch

• Whether external validation on data outside Crisis Text Line reproduces the results, and whether accuracy figures are published in a form that allows comparison with clinician judgment.
• How the lexicon holds up against a context-aware LLM baseline on the same conversations, and whether the interpretability gain costs measurable accuracy.
• Whether any operator moves from research use to live triage, and what human review requirements accompany it.
• The terms of the open release, and whether independent teams can reproduce the lexicon-building pipeline with a different underlying language model.
• How applications to social media and electronic health records handle consent, given that those sources lack the help-seeking context of a crisis line.