How Spain relies on the VioGén algorithm to score the likelihood of recurring domestic violence and decide protection measures, with at times fatal consequences
Spain has become reliant on an algorithm to score how likely a domestic violence victim is to be abused again … X: @tdietterich , @stevebholt , @ichbinilya , @noisejockey , @ichbinilya , @jonathancohn , @onherownlife , and @satariano X: Thomas G. Dietterich / @tdietterich : Important article on algorithmic decisionmaking. https://www.nytimes.com/... However, I was disappointed that they didn't discuss the false positive errors as well. That discussion would probably show that *all* women would benefit from protection to greater or lesser degree. Steve B. Holt / @stevebholt : I tend to support data-informed decision-making and experimenting in this direction, but policymakers really need to think about the context first. Domestic violence is pretty well-known for having thin, bad, incomplete data and high stakes. https://www.nytimes.com/... Ilya Lozovsky / @ichbinilya : The NYT just published this long feature story about an algorithm the Spanish police use to determine whether a woman who has experienced domestic violence is likely to be abused again. The point is to focus police resources on the most vulnerable. https://www.nytimes.com/... Nathan Moody / @noisejockey : We are annoyed by algorithmic ads/feeds and erroneous LLM output. So why is anyone surprised that ceding public safety to an algorithm may not be a great idea? https://www.nytimes.com/... Ilya Lozovsky / @ichbinilya : But... is that actually an indicative example? In the fifth paragraph, after the stabbing, we have this line. “At its best, the system has helped police protect vulnerable women and, overall, has reduced the number of repeat attacks in domestic violence cases.” [image] @jonathancohn : Outsourcing decision-making to algorithms can have fatal consequences. https://www.nytimes.com/... @onherownlife : Violence is not algorithmic. A set of questions and a decision tree (esp w/o broad expert input) can't accurately predict human nature. Tools like these are only as good as the info they get - and we don't yet know what all of the questions are. https://www.nytimes.com/... Adam Satariano / @satariano : In Spain, an algorithm scores every reported victim of gender violence on how likely they are to be abused again. At its best, the system helps police protect vulnerable women. But when the score is wrong the consequences have been fatal. https://www.nytimes.com/...
Context & Ripple Effects
Spain’s use of VioGén places domestic-violence protection within the broader expansion of predictive systems into police, sentencing and probation decisions. The distinction here is consequential: the score helps direct protective interventions for people reporting abuse, not merely an administrative workflow.
Earlier scrutiny of COMPAS risk-assessment shortcomings established why error rates, data quality and how staff use a score matter as much as a model’s stated objective. VioGén’s reported benefits therefore coexist with unusually high stakes when a score is wrong or misapplied.
First-order effects
- Reported victims’ protection measures and police attention are prioritized through VioGén’s assessed recurrence risk, making the score an immediate input to frontline safety decisions.
- A false low score can leave a victim with insufficient protection, while a false high score can direct limited police resources toward cases that may not need the same level of intervention.
Second-order effects
- Police leaders face pressure to examine whether officers override, validate or rely too heavily on scores, since model use and implementation can each produce harmful outcomes.
- Because VioGén helps allocate protective capacity, its error trade-offs become operational: reducing missed high-risk cases may require accepting more interventions for people the system rates as high risk.
Third-order effects
- If public-safety agencies continue embedding risk scores in protection decisions, explainability, error monitoring and meaningful human accountability are likely to become central tests of legitimacy rather than optional model features.
- The case points to a broader shift from algorithms as advisory analytics to algorithms that shape access to state protection—raising the threshold for evidence that such systems improve outcomes across error types.
The trend: Public-safety AI is moving from prediction toward operational allocation of scarce protective resources, intensifying scrutiny of who bears the cost of model errors.