PUBLISHED · NPJ MENTAL HEALTH RESEARCH · 2025
Psychosocial dynamics of suicidality and nonsuicidal self-injury: a digital linguistic perspective
Charlotte Entwistle, Katie Hoemann, Sophie J. Nightingale, Ryan L. Boyd
Finding:
In a large naturalistic study of 66,786 posts from 992 Reddit users with self-identified borderline personality disorder, the researchers found that measurable changes in language preceded disclosures of suicidality and nonsuicidal self-injury (NSSI). The study tracked sixteen LIWC-22 categories spanning self-processes, emotion, social processes, and cognition, with self-harm events manually coded by trained raters rather than inferred from language alone.
At the person level, self-focused language, negative emotion, sadness, and anger were associated with the frequency of both suicidality and NSSI disclosures, while swearing and absolutist language were additionally associated with recent suicidality. These relationships were small, generally around r = .10 to .16, and the authors treat this analysis as descriptive rather than predictive.
The temporal analysis is the most consequential finding. Tracking language weekly from three weeks before an event to three weeks after, anxiety language rose sharply two weeks before suicidality and remained elevated in the immediately preceding week. Sadness and swearing increased in the week immediately before the event, while third-person references decreased. After the event, affiliation language increased.
NSSI showed a different trajectory. Sadness language increased two weeks before an event and then fell sharply in the immediately preceding week. Anger increased substantially in the week afterward, while affiliation language declined in the weeks leading up to the event. These patterns show why change over time can carry information that a single measurement may not: the direction and sequence of psychological signals can matter as much as their absolute level.
Relevance:
Co-authored by Ryan Boyd, academic advisor to Receptiviti. Measurement throughout is LIWC-22, the instrument underlying Receptiviti’s dimensions.
The study demonstrates something particularly important for AI safety: psychologically relevant risk can appear as change in a person’s language over time, before an adverse event occurs. In the suicidality analysis, anxiety language increased two weeks before the event and remained elevated in the preceding week, while sadness and swearing increased in the week immediately before it. These were trajectories measured relative to the individual’s earlier language, not simply the presence of obviously concerning words in the event disclosure itself.
Two properties matter for measuring what AI systems do to the people using them. The first is trajectory. A person’s direction of change can contain information that a point-in-time assessment misses. The NSSI results illustrate this particularly well: sadness increased two weeks before the event and then fell sharply in the immediately preceding week. A safety system looking only for elevated sadness could therefore interpret a falling score very differently from one able to see the trajectory that preceded it.
The second is operationalization. Language is produced continuously during conversational AI interactions. That creates the possibility of measuring psychologically relevant changes as the interaction unfolds, independently of the model’s own characterization of the user. The Entwistle study did not test a real-time intervention system, so it does not establish that these signals can predict or prevent harm. What it does establish is the underlying premise: measurable psycholinguistic changes can precede clinically important events. That makes real-time trajectory measurement a credible direction for developing systems that recognize emerging risk early enough to adjust their behavior or trigger safeguards before harm occurs.