What the research found
In two studies, researchers from the Child Mind Institute, Harvard and MIT compared AI and natural-language methods for assessing psychological states on crisis-counselling conversations and public social-media posts, then used crisis conversations to model about 50 factors that predict later emergency-service intervention. Mentions of lethal means and substance use or overdose were among the strongest predictors. The authors conclude the tools cannot yet provide valid psychological assessments and set out standards for testing them.
Tools that score crisis or suicide risk from text are already on the market, and this says the evidence for trusting them without a clinician isn't there yet.