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:
A naturalistic study of 66,786 posts from 992 Reddit users with self-identified borderline personality disorder, drawn from the two largest online BPD communities and spanning 2011 to 2019. Self-harm events were manually coded by trained raters rather than inferred by keyword, and sixteen LIWC-22 categories mapped onto four psychosocial dimensions: self-processes, emotion, social processes, and cognition. The design allows two things most work in this area cannot do at once, which is to describe who is at risk from stable language patterns and to describe when, from language change over time.
At the person level, self-focused language, negative emotion, sadness and anger were associated with the frequency of both suicidality and self-injury disclosures, with swear words and absolutist language additionally marking suicidality. These correlations are small, in the range of r = .10 to .16, and the authors present them as descriptive rather than predictive.
The temporal analysis is where the study is most informative. Tracking language weekly from three weeks before an event to three weeks after, anxiety language rose sharply two weeks before suicidality events and stayed elevated, with sadness and swearing rising in the week immediately preceding, and third-person references falling away in the same window. Affiliation language then increased in the week after. Self-injury followed a different and counterintuitive pattern: sadness language rose two weeks out and then dropped sharply in the week immediately before the event, which the authors read, given persistently high anxiety, as a possible period of emotional numbing or dissociation. Anger language rose steeply in the week following, suggesting the behaviour generated further dysregulation rather than resolving it.
The study also examined how the community responded. Posts disclosing recent suicidality received more upvotes than non-disclosure posts; posts disclosing recent self-injury received fewer replies. Across the full corpus, posts scoring higher on anxiety, sadness, anger and swearing received more upvotes. The authors note this cuts both ways, since the community appears to reward the expressive register that precedes self-harm, and flag the causal question as untested.
Relevance:
Co-authored by Ryan Boyd, academic advisor to Receptiviti. Measurement throughout is LIWC-22, the instrument underlying Receptiviti's dimensions.
Two properties matter for measuring what AI systems do to the people using them. The first is temporal resolution. The signal that distinguishes the weeks before an event from baseline is not present in any single post and would not survive periodic self-report, both because it moves faster than most assessment schedules and because, as the authors note, people in acute distress often lack insight into their own psychosocial state. Reading it from language is what makes it observable at all. The second is that the direction of change carries information the level does not: sadness language falling in the week before self-injury is the opposite of what a threshold-based monitor would flag, and only trajectory makes it legible.