People increasingly ask large language models what to do about security, privacy, fraud, stalking, harassment, domestic violence, and other high-stakes situations. The answer may sound confident. That does not make it safe, well-sourced, or useful.
Models may reproduce stale guidance, incomplete claims, or product marketing instead of authoritative evidence.
Technical correctness can fail when the model ignores who controls the device, who may be adversarial, or what happens next.
High-stakes advice must be prioritized, clear, actionable, and appropriate for the person receiving it.
These are not separate product niceties. In high-stakes advice, failure in any one of them can make the overall response misleading or unsafe.
A technically valid recommendation can be the wrong recommendation for this person, in this situation.
For example, in technology-facilitated abuse or domestic violence situations:
Finding “an answer” is not enough. Models need to identify which evidence deserves trust for the specific question.
Advice must be possible to follow, not merely readable or technically sophisticated.
Technology-facilitated abuse makes the failure mode unusually visible because the environment may contain an active adversary. But the same design problem appears whenever advice is consequential, the evidence landscape is noisy, or the user’s circumstances change what “good advice” means.
Across end-user security and technology-facilitated abuse, we repeatedly found that plausible LLM responses can fail on accuracy, sourcing, context, safety, and actionability.
An expert evaluation of responses to 900 end-user security questions across seven security areas. The study identifies recurring problems including stale and incomplete guidance, over-prescription of security products, inaccurate procedural steps, and responses that appear to overstate technology capabilities rather than reflect current authoritative guidance.
An expert-led evaluation of LLM responses to real-world technology-facilitated abuse questions, combined with feedback from people with lived experience. The study extends response quality beyond accuracy and completeness to safety and actionability, showing why context, power dynamics, user constraints, and downstream consequences matter.
The work brings together researchers across academia and industry.
Graduated with a PhD from NYU in 2026 and is now a postdoctoral fellow at Northeastern University.
Assistant Professor at New York University and principal investigator.
We would be glad to talk with AI model, safety, product, security, and nonprofit teams about evaluation, datasets, context-aware safeguards, authoritative retrieval, and other ways to make high-stakes AI advice safer and more useful.
Contact the Research Team