Skip to content
Sign in
Defense & Security · North America

Study finds 60% of chatbot answers omitted abuse-safety warnings

A study led by NYU Tandon researchers found that 60% of tested chatbot responses to technology-facilitated abuse omitted critical safety warnings. The researchers also found that 86% of responses fell short on accuracy, completeness or safety.

AI-assisted synthesis · Based on 2 outlets' reporting · Article text read from 1 of 2 outletsHow it's made

View timeline · 6 entries
2 outletsLatest reporting 3 min read
Coverage: Too few ratedL 0 · C 1 · R 0 · 1 unrated
Updat3 graphic: Defense & Security
Updat3 graphic

What happened

Key facts

  • The researchers found that 86% of tested responses fell short on accuracy, completeness or safety (per newswise.com).
  • The researchers found that 60% of tested responses omitted critical safety warnings (per newswise.com).
  • The study’s findings were presented in August 2026 at the USENIX Security Symposium (per newswise.com).

The brief

A study led by NYU Tandon researchers found that 60% of the tested AI chatbot answers about technology-facilitated abuse omitted critical safety warnings. The researchers also judged 86% of the answers deficient in at least one of three areas: accuracy, completeness or safety, according to Newswise. The work examined advice about technology-facilitated abuse.

Newswise describes it as part of a wider NYU Tandon School of Engineering research effort assessing whether AI-generated guidance is safe to follow in high-stakes situations, including scams, hacked accounts and harassment. The reported findings concern the abuse-related responses; the broader examples identify other areas encompassed by the research effort.

The researchers reported different unsafe-response rates across the chatbots they tested. Aimee’s answers received an unsafe rating about 74% of the time. The corresponding figures were 65% for Claude, 61% for Ruth and 55% for GPT-4o. These figures describe the study’s ratings of responses, rather than the share of users who experienced harm.

Danny Huang, a Tandon assistant professor, member of the NYU Center for Cybersecurity and the project’s principal investigator, cautioned that an answer’s apparent quality may not show whether following it would make someone safer.

He said an answer could sound plausible and be technically correct while leaving out essential context or suggesting an action that does not improve safety. Huang and Vijay Prakash were identified as the project’s core researchers. Prakash earned his Tandon PhD in 2026 and is now a postdoctoral fellow at Northeastern University.

Newswise also states that the problem extends beyond domestic violence, without specifying further categories in that remark. The findings were presented in August 2026 at the USENIX Security Symposium, which Newswise describes as a leading cybersecurity research conference. The paper was named a runner-up for the Distinguished Paper Award.

TechTimes’ headline also reports that chatbots missed key safety warnings in tests on questions about digital abuse.

Why it matters

  • People seeking advice about technology-facilitated abuse may encounter omissions in safety guidance: researchers found critical warnings missing in 60% of tested answers.
  • The study found a broader quality problem across the tested responses, with 86% falling short on accuracy, completeness or safety.
  • Unsafe ratings varied among the named chatbots, from 55% for GPT-4o to about 74% for Aimee.

What to watch next

  • Whether NYU Tandon researchers publish further findings on the chatbot responses tested for technology-facilitated abuse.
  • Whether the researchers provide more detail about the study’s evaluation methods and how they rated answers unsafe.
  • Whether the researchers specify which other issues are covered by their statement that the problem extends beyond domestic violence.

How it's being covered

Coverage analysisShow details
2 source articles · 2 unique outlets
Outlet political mix: Too few rated· 1 Center · 1 unrated
Where the outlets are based: Europe 50% · Unknown 50%
Outlet political mix
Too few rated
Left
Center
Right
0 outlets
1 outlet
0 outlets
The marker shows coverage balance: each unique outlet counts once. 1 unrated outlet is not included.
Geography
Where the outlets covering this story are based.
2 unique outletsDominant coverage: Europe
Limited coverage
  • Coverage is currently limited to 2 outlets.
How we measure coverage

How outlets' accounts compare

Updat3 does not yet have enough full reports to compare how outlets framed this story.

Timeline

Developments from this story's sources, dated background we compiled and fact-checked, and our earlier coverage. Each row says which.
  1. Congress enacted the Violence Against Women Act (VAWA) on September 13, 1994, establishing federal measures addressing domestic violence, sexual assault and stalking.
    Background · from our historical context
  2. Congress passed the Interstate Stalking Punishment and Prevention Act, addressing stalking across state lines.
    Background · compiled and fact-checked by Updat3
  3. The National Network to End Domestic Violence established its Safety Net Project in 2002 to address the intersection of technology and domestic violence, including ways technology could be misused against survivors.
    Background · from our historical context
  4. In 2015, the United Nations Special Rapporteur on violence against women published a report addressing online violence against women and girls.
    Background · from our historical context
  5. The immediate technological backdrop began on November 30, 2022, when OpenAI released ChatGPT to the public, making conversational generative AI widely accessible.
    Background · from our historical context
  6. The USENIX Security Symposium, where the study was presented in August 2026, is an annual research conference focused on computer security.
    Background · from our historical context

How we got here

The immediate technological backdrop began on November 30, 2022, when OpenAI released ChatGPT to the public, making conversational generative AI widely accessible.

ChatGPT followed earlier public-facing AI assistants, including Apple’s Siri, introduced in 2011, and Amazon’s Alexa, released in 2014; those products had already made asking a digital assistant for practical guidance familiar to many users. The underlying problem predates generative AI.

Original sources

1 of 2 articles · Europe · Show all

Every story links to the reporting it was built from. How we choose sources

Spotted a factual error, a broken link or a misleading summary? Tell us.