An AI system does not need to sound hateful to produce an answer that undermines human rights. It can speak calmly about efficiency, public safety, or economic benefits—and make restrictions on people’s freedoms sound reasonable.
That is the warning from UDHRBench, a research preprint published on October 10. Researchers assessed 4,496 responses from ChatGPT, Grok, Gemini, and Fanar using 1,124 questions covering the Universal Declaration of Human Rights. They identified answers that, under their evaluation framework, justified rights violations through apparently reasonable arguments.
A polite answer can still deserve scrutiny
The concern is easy to understand. Imagine an automated adviser recommending that an institution sacrifice individual protections to improve performance. Its explanation might contain no insults, threats, or obviously inflammatory language. Yet the people affected would still need to know whether the recommendation discriminates against them or removes meaningful safeguards.
Courtesy cannot settle those questions. Nor can a confident explanation establish that the assumptions behind a recommendation are sound. An answer deserves scrutiny because of what it proposes, whose interests it serves, and what evidence supports it.
This matters especially when an institution wants an apparently objective justification for a decision it already favors. An AI-generated explanation could give that decision a veneer of impartiality without resolving its ethical problems.
The benchmark also needs scrutiny
However, the study’s classifications deserve examination too. Its authors labeled support for detention following credible death threats as misaligned, including answers requiring judicial oversight. That example raises questions about whether the benchmark adequately distinguishes arbitrary detention from lawful preventive measures.
International human rights standards prohibit arbitrary detention. They do not establish that every preventive detention is automatically unlawful. UN Human Rights Committee guidance requires legal grounds, safeguards, and judicial control when such detention occurs. Consequently, the circumstances and protections matter; a simple yes-or-no classification can obscure them.
A benchmark can therefore identify a valuable problem while remaining open to criticism over individual judgments. Agreement between evaluators does not, by itself, establish that their interpretation is correct. Independent examination should test the evaluators’ assumptions as carefully as the models’ answers.
A stress test, not a verdict on every conversation
The research has other limits. Two authors both designed the questions and evaluated responses. Answers were collected between November 2025 and February 2026, and the deliberately challenging prompts do not represent ordinary usage. The findings concern particular systems during that period, rather than every current version.
Those limitations should shape the conclusions. The paper does not demonstrate that AI deliberately plots against human rights, nor does a troubling response establish that an institution acted on it. Producing an argument, endorsing a policy, and implementing a decision are different things.
Nevertheless, the practical question remains urgent: who checks the reasoning when software advises people with power over others? Institutions should demand transparent criteria, independent review, and meaningful opportunities to challenge consequential decisions.
Human rights protections cannot depend on whether a chatbot produces a reassuring paragraph. Neither should criticism of AI depend on accepting a research label without examining the argument beneath it.
