People suffering from a technological form of abuse - spyware on their phone, surveillance by a former partner or control over their accounts - can turn to artificial intelligence chatbots for help. A new study led by specialists at the NYU Tandon School of Engineering found that the advice from such systems is often inaccurate, incomplete or dangerous.
The researchers evaluated four AI systems, including two chatbots built specifically for victims of intimate partner violence (IPV). They found that 86% of the tested responses did not meet the criteria for accuracy, completeness or safety, and in 60% of cases important safety warnings were missing.
The results were presented in August 2026 at the USENIX Security symposium in Baltimore, where the work was named a finalist for the Distinguished Paper Award.
"A response can sound confident and reasonable and even be technically correct, yet miss critically important context or recommend actions that do not actually make a person safer," said Danny Huang, an associate professor at Tandon and the project's principal investigator. "In high-risk situations, that difference can have very real consequences."
The NYU team conducted the first expert evaluation of AI responses to real questions from victims about technology-facilitated abuse. The researchers compiled a set of 193 questions from academic literature, Reddit and Quora and put them to GPT-4o and Claude 3.7 Sonnet, as well as to the chatbots Aimee and Ruth, which were built specifically for IPV victims. The responses were rated for accuracy, completeness and safety based on authoritative sources. In addition, 114 people with experience of technology-facilitated abuse took part in a survey on how applicable the advice seemed.
The specialized chatbots for IPV victims did not outperform general-purpose models. Aimee's responses were judged unsafe in about 74% of cases, compared with 65% for Claude, 61% for Ruth and 55% for GPT-4o. Ruth had the highest share of responses rated as partially or minimally accurate - about 62%.
Among the recurring errors: recommending the use of a VPN against spyware and account hacking; advising radio-frequency detectors to find hidden devices, which are mostly ineffective; suggesting illegal GPS jammers; changing passwords when an abuser controls or has shared access to accounts or devices. None of the systems noted that spyware in the system partition of a rooted or jailbroken phone may survive a factory reset. The models also often failed to warn that changes could alert an abuser or escalate the situation, and did not consider the need to preserve a compromised device as evidence.
The researchers also tested whether the systems could be induced to help a potential perpetrator. All four systems complied, with Ruth and Aimee answering about two-thirds of such questions.
The problem extends beyond domestic abuse. In related work presented at a computer security applications conference in 2025, researchers examined 900 GPT responses to users' security questions in seven areas, including authentication, social engineering and antivirus software. Of these, 485 failed to meet at least one criterion: accuracy, completeness, relevance or directness. The same types of errors also appeared in responses from Llama and Gemini.
Together, the two studies show that the problem is not limited to a single chatbot, type of question or high-risk scenario. The researchers recommend treating AI security advice with caution, and support organizations should independently evaluate AI systems before relying on them in critical situations.