AI chatbots fell short on accuracy, completeness or safety in 86% of 268 responses assessed in a study of technology-facilitated abuse, while 60% omitted critical safety warnings.
The NYU-led research also found that specialist intimate-partner-violence chatbots did not outperform general-purpose systems in the study's testing conditions.
When Technically Correct Advice Can Still Be Dangerous
In technology-facilitated abuse, a device is not simply a device. It may be a tracker, a witness, a lock, or the place where an abuser still has access.
That distinction matters because changing a password or resetting a phone could alert an abuser, destroy evidence or leave a survivor less safe. An answer may sound helpful and still omit the context that makes it safe to follow.
Researchers at NYU Tandon selected 193 questions from academic literature and online discussions covering spyware, account takeovers, location tracking and other forms of technology-enabled abuse.
They then assessed 268 chatbot responses against authoritative sources for accuracy, completeness and safety. The study evaluated GPT-4o, Claude 3.7 Sonnet and two specialist systems, Aimee and Ruth, although the models did not receive equal numbers of expert-evaluated questions.
The researchers also surveyed 114 people with lived experience of technology-facilitated abuse about how actionable the advice appeared. That survey measured perceived usefulness, not real-world outcomes.
Specialist Systems Did Not Outperform General Models
The specialist systems did not solve the problem. Aimee's responses were rated unsafe about 74% of the time, compared with 65% for Claude, 61% for Ruth and 55% for GPT-4o.
Those figures describe the responses assessed in the research and should not be treated as universal failure rates for every later version of the systems.
The researchers identified recommendations involving largely ineffective radio-frequency detectors, VPNs in situations involving spyware or account hijacking, GPS jammers that would be illegal in the US, and password changes that could prove ineffective or risky where an abuser retained access to shared accounts or devices.
None of the systems warned that spyware installed on the system partition of a rooted or jailbroken phone could survive a factory reset. The models also frequently failed to explain that changing settings or removing access could alert an abuser or affect the preservation of evidence.
Researchers further tested whether the systems could answer questions from a potential perpetrator. All four could be induced to respond, with Ruth and Aimee answering roughly two-thirds of those questions.
Danny Huang, an assistant professor at NYU Tandon and the study's principal investigator, said that technically correct advice could still leave out critical context.
'In high-risk situations, that distinction can have very real consequences,' Huang said.
Why the Findings Matter in the UK
The study evaluated US-focused questions and did not establish how UK users or UK-specific systems would perform. Its findings nevertheless arrive as British authorities and support organisations address technology-facilitated abuse.
The Mayor of London's office cited National Police Chiefs' Council figures showing at least 123,515 violence-against-women-and-girls offences with an online or technology-enabled element between August 2022 and July 2023.
The Mayor's office also cited Refuge figures showing a 207% increase in technology-abuse referrals between 2018 and 2024. In a separate statement, Refuge reported a 78% rise in referrals to its specialist Technology-Facilitated Abuse and Economic Empowerment team over the period it described as the previous year.
Those figures concern different measures and reporting periods. They should not be treated as a single continuous series.
A recent TechTimes report on online violence against women described related UK findings on coercive control, stalking, monitoring, harassment and image-based abuse.
Ofcom says women and girls face disproportionate risks from online harms including stalking, coercive control, coordinated harassment and image-based sexual abuse. Its Online Safety Act work requires regulated services to address specified online harms, while its wider guidance encourages safety-by-design
The NYU findings point to a related but distinct concern: the reliability of advice generated by AI systems when people seek help with technology-enabled abuse. The study did not examine whether the systems complied with UK online-safety legislation.
The researchers recommend treating AI-generated safety and security advice with caution, and say support organisations should independently evaluate systems before relying on them for high-stakes guidance.
