A research team led by scientists at Harvard Medical School has developed AI models capable of identifying patients at elevated risk of intimate partner violence using data already collected in electronic health records. The models flagged potential risk, on average, more than three years before patients enrolled in a hospital-based domestic abuse intervention program. The technical achievement is real. What it means for survivors is a harder question.
What the Algorithm Actually Sees
The models were trained on routinely collected health record data from female patients at a single U.S. academic health centre, then validated across three additional patient cohorts. The researchers themselves noted that careful performance evaluation on the broader population of female patients should be conducted before any clinical implementation.
That caveat matters. The gap between a validated model and a deployed tool is where most of the ethical weight sits. The AI does not observe violence. It infers risk from patterns in medical data, patterns that correlate with intimate partner violence without directly revealing it. Detection and understanding are not the same thing, and neither is detection and response.
The Problem With Flagging Someone Who Has Not Spoken
Here is what most coverage of this kind of research misses: for someone experiencing intimate partner violence, silence is often not passivity. It is strategy.
Deciding whether to disclose abuse involves a range of factors that no algorithm can fully account for. Safety concerns, financial dependence on an abusive partner, distrust of social services, fear of not being believed, and the very real possibility that disclosure could trigger child removal proceedings. In some situations, staying silent is how a person keeps themselves and their children safe.
Trauma- and violence-informed care, the framework that guides best practice in this area, is built on principles of safety, choice, collaboration, connection and capacity-building. Those principles exist precisely because survivors need to retain control over their own stories. A system that infers abuse before a person has chosen to share it shifts that control away from them, however well-intentioned the technology behind it.
One of the researchers involved has stated that the tool’s goal is not to pressure patients into disclosure, but to help clinicians communicate more compassionately and connect patients with support. That framing is important. It also leaves open a critical question: what actually happens once a patient is flagged? Is the clinician trained to respond in a trauma-informed way? Are appropriate services available and accessible? The research does not yet answer those questions.
Why Detection Is Not the Same as Care
There is a second risk that runs in the opposite direction. If health systems become over-reliant on AI-generated risk scores, the absence of a flag could discourage providers from asking about violence at all. Research on clinical decision-support systems has already identified this pattern: automation bias, the tendency of clinicians to defer to automated recommendations even when those recommendations are wrong.
Applied to intimate partner violence, automation bias could mean that a survivor who does not match the model’s risk profile is less likely to be asked about their experience. The algorithm becomes a filter, and some people fall through it.
This points to a structural issue that AI cannot resolve on its own. Even if a model correctly identifies elevated risk, the survivor will still encounter the same gaps in services, safety planning and resources that exist today. Better identification does not automatically produce better outcomes. The infrastructure of response has to be there first.
A Question About Who Gets to Tell the Story
The broader significance of this research is not really about accuracy rates or model validation. It is about a shift in the relationship between patients and health systems, one that AI is accelerating across many domains but that carries particular weight here.
Intimate partner violence is defined, in part, by coercion and control. The ability to decide when, how and whether to speak about that experience is not a procedural detail. For many survivors, it is one of the few areas of autonomy they have. A tool that removes that choice, even in the name of early intervention, reproduces a dynamic that trauma-informed care is specifically designed to disrupt.
The researchers’ own framing acknowledges this tension. Any implementation of AI-based risk detection, they argue, must be evaluated not only for predictive accuracy but against the principles of safety, choice, collaboration, connection and capacity-building. That is a meaningful standard. It also means that deploying this technology responsibly requires far more than a well-performing model.
In Short
AI can now identify potential intimate partner violence risk years before survivors choose to disclose it. The technical capability is real and the intention behind it is to improve care. The risk is that early detection, without the right clinical training, survivor-centred protocols and accessible services, could undermine the very agency it aims to protect. Detecting a problem and responding to it well are two entirely different things. The gap between them is where the real work begins.
Based on reporting from The Conversation.