Original title: When AI Says "I Am Unable to Answer": Understanding User Responses to AI Refusals
Mahjabin Nahar, Eun-Ju Lee, Yujin Heo, Dongwon Lee · September 2026
Artificial intelligence tools sometimes invent false facts, which researchers call hallucinations. To stop this, makers program the AI to refuse to answer risky questions. But people usually want clear answers. Researchers tested how users react when an AI says it cannot help. They had people interact with an AI that never refused, refused infrequently, or refused frequently. Some refusals included explanations, while others did not. The results showed that users liked genuine answers best, followed by hallucinations, and liked refusals least of all. Surprisingly, people preferred false answers over refusals, even though they knew the false answers were less accurate. Adding explanations helped only when refusals were infrequent. Also, people who strongly crave clear, definite answers disliked refusals the most.
You are at a restaurant and ask the waiter for a dish recommendation. A confident waiter gives you a suggestion, even if it might not be great; an honest waiter sometimes says he is not sure and needs to check. Most people prefer the confident waiter, even knowing the suggestion might be wrong, because they want an answer right away. In this paper, the AI's confident wrong answers are like the confident waiter, and its refusals are like the honest uncertainty. The researchers found people preferred the confident wrong answer, especially when they wanted a quick decision.
AI chatbots often refuse to answer to avoid making up facts, but this study shows users actually prefer a wrong answer over a refusal. This creates a dilemma for designers: too many refusals frustrate users, but too few risk misinformation. Understanding when refusals are acceptable helps build AI that people trust and enjoy using.
The study does not examine long-term effects of repeated refusals, nor does it test different types of refusals (e.g., safety vs capability). The findings may not generalize to other tasks or AI models, and the experiment used a controlled setting rather than real-world use.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
When AI Says "I Am Unable to Answer": Understanding User Responses to AI Refusals — Mahjabin Nahar, Eun-Ju Lee, Yujin Heo et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.