← Blog

What an algorithm cannot yet say about suicide risk

A commentary on my article about artificial intelligence and the prediction of suicidal behaviour in university students.

In recent years machine learning has been presented as one of the most promising ways to improve the prediction of suicidal behaviour, a field in which fifty years of research have not produced clear answers about risk factors, or about the configuration of factors that precedes a death by suicide (Franklin et al., 2017). The promise is especially appealing in universities, where mental health problems are common and where institutions have data, research teams and a legitimate concern for their students. In 2025 I published a paper in the Revista Iberoamericana de Psicología in which I tried to pause on that promise and ask what ethical safeguards its implementation requires (Fuentes-Ferrada, 2025). What follows is a commentary on that work, built around the three considerations that, as I see it, still fully hold.

Predicting ideation and preventing suicide

Most people who experience suicidal ideation do not go on to suicidal behaviour, and the factors traditionally associated with suicide (such as depression, hopelessness or impulsivity) predict ideation rather than attempts (Klonsky & May, 2014). Even so, many studies with university students train algorithms to predict suicidal ideation and then present their results as a contribution to suicide prevention, which assumes that ideation is the first step on a path that leads to an attempt. The first consideration concerns precisely this, what the models actually predict. Claiming that predicting ideation helps prevent attempts or deaths by suicide may therefore amount to a false promise, one that is neither empirically nor ethically consistent.

Added to this, suicide research has been oriented mainly towards deficit (psychopathology and risk factors), losing sight of the protective factors that directly reduce risk. So-called positive suicidology has begun to explore the role of reasons for living, meaning in life and psychological wellbeing, and some of the studies I review in the article even found that people with positive mental health are less likely to attempt suicide despite having ideation. If algorithms are fed only with deficit variables, they will probably reproduce, with greater sophistication, the same problems that have stalled predictive capacity over recent decades. Using these analyses to identify the mechanisms that sustain wellbeing could instead become a promising field of development.

After screening

The second consideration means going beyond accurate screening. If a team implements a machine learning strategy to detect suicide risk at a university, it also needs protocols, established in advance, to guide the support and follow-up of those who are identified. One example I find worth following is a study with Chilean university students in which I took part (Langer et al., 2024). Students who reported suicidal ideation received a phone call, a risk assessment by a specialist, weekly follow-up in cases of moderate or high risk, and referral to emergency services when the risk was imminent. Even with accurate screening, an inaccurate response is potentially fatal, so detection commits the institution to a responsibility that goes far beyond data analysis.

Who assesses

In a systematic review of machine learning studies, none had managed to report imminent risk of suicidal behaviour (Burke et al., 2019). The third consideration therefore concerns the agent of assessment (who, or what, determines the risk). Even if algorithms improved enough to show more clearly who is more likely to attempt to take their own life, they could not show when someone at risk might act. Speaking of imminent risk requires exploring specific, idiographic factors, those that belong to one person and their situation, and that calls for data and also for a conversation, a relationship and a clinical judgement that artificial intelligence is still far from acquiring. In the article I proposed, as a minimum ethical criterion, that after screening every student at risk be connected with trained human assessors, whether in a university health service or in the wider health network.

I wrote that article as a clinical psychologist with experience in suicide risk assessment in primary care, university settings and psychiatric units, and I recognise that this familiarity with clinical procedures may bias my reading in favour of assessment between humans. Even so, I still think we are facing an overlap between the crisis in suicide research and the innovation of predictive analytics, one that invites a much needed dialogue between both disciplines. If advances in artificial intelligence do not engage with new developments in suicidology, we will have sophisticated analyses that repeat old problems. If they do, it becomes possible to move towards a more precise, and at the same time ethically responsible, understanding of suicidal behaviour in university students.

If you or someone close to you is going through a difficult time, in Spain you can call 024 and in Chile *4141, both free and available 24 hours a day. Elsewhere, your local emergency number or crisis line can help.

References

Burke, T. A., Ammerman, B. A., & Jacobucci, R. (2019). The use of machine learning in the study of suicidal and non-suicidal self-injurious thoughts and behaviors: A systematic review. Journal of Affective Disorders, 245, 869-884. https://doi.org/10.1016/j.jad.2018.11.073

Franklin, J. C., Ribeiro, J. D., Fox, K. R., Bentley, K. H., et al. (2017). Risk factors for suicidal thoughts and behaviors: A meta-analysis of 50 years of research. Psychological Bulletin, 143(2), 187-232. https://doi.org/10.1037/bul0000084

Fuentes-Ferrada, R. (2025). Inteligencia artificial y predicción de la conducta suicida en jóvenes universitarios, algunas consideraciones éticas. Revista Iberoamericana de Psicología, 18(2), 87-97. https://doi.org/10.33881/2027-1786.rip.18207

Klonsky, E. D., & May, A. M. (2014). Differentiating suicide attempters from suicide ideators: A critical frontier for suicidology research. Suicide and Life-Threatening Behavior, 44(1), 1-5. https://doi.org/10.1111/sltb.12068

Langer, Á. I., Schmidt, C., Martínez, V., Mac-Ginty, S., Fuentes, R., & Núñez, D. (2024). Suicidal ideation, non-suicidal self-injury and psychopathology in university students: Examining the moderating role of experiential avoidance. Journal of Affective Disorders Reports, 16, 100734. https://doi.org/10.1016/j.jadr.2024.100734

reinerfuentes7@gmail.com

Write to me

What would you like to talk about?