Definition
Refusal is a refusal is when an AI model declines to do what you asked, either by saying no or by being blocked by a safeguard, because the request is judged harmful, restricted or out of scope.
A refusal is any time an AI model or the system around it declines a request. It can be a polite "I can't help with that" from the model, or a safeguard that blocks the request before the model answers.
Refusals have two main sources. The first is training. Through post-training methods such as RLHF, developers teach a model to decline requests it should not fulfill, for example help with weapons or serious cyberattacks. The refusal then comes from the model's own judgment.
The second is a separate safeguard layer. Classifiers, filters or AI guardrails sit in front of or beside the model and block certain inputs or outputs. Some systems hand a blocked request to a different model instead of refusing outright. A system prompt can also add rules, such as staying on a topic.
Refusals can fail in both directions, and that is the whole difficulty.
Tightening one tends to loosen the other. A safeguard strict enough to catch every attack will also block some honest work. Developers therefore tune the balance and often publish refusal or block rates alongside their other results.
A refusal is a failed call in your product. If your app depends on a model finishing a task, a surprise refusal looks like a bug to your user. It helps to log refusals separately from other errors, test your real prompts after every model change, and write prompts that state your legitimate purpose clearly. Do not try to trick a safeguard: that can violate usage terms, and it is what jailbreaks do.
Verified access programs exist for some sensitive fields so that vetted professionals can use fewer restrictions. Check a provider's documentation for what it offers.
Refusals connect to AI alignment, which asks how to make models act as their makers and users intend, and to sycophancy, the opposite habit of agreeing too readily. A good assistant should refuse rarely, clearly and for reasons you can understand.
It is when an AI model or its safeguards decline to carry out a request because it is judged harmful, restricted or out of scope.
Over-refusal is when a model declines a legitimate request, such as a harmless security or medical question, because a safeguard is too cautious.
Anthropic expanded its Cyber Verification Program on October 6 into Defense, Red Team and Specialized tiers that lift some of Claude's cyber restrictions for vetted teams. If refusals have slowed your security work, verification is now the way around them.
CrowdStrike says a likely Chinese-speaking, financially motivated actor ran ARTEX, an open-source agentic pentest tool, against South Korean financial firms from late September, and cites industry reports of breaches at two banks. The model under it was swappable, and that is the warning.
Anthropic says new defenses cut the Claude reasoning data distillers harvest by about 90 to 99%, and it has banned tens of thousands of accounts. The fine print reaches honest developers too: new API accounts can't edit Claude's prior context, and blocked requests are billed.
SemiAnalysis checked 857 model releases from nine Chinese developers and found a published safety result for 31 of them, and only 9 at launch. If you run Chinese open weights, the safety testing is your job.
OpenAI updated its ChatGPT extension for Chrome on Oct. 15 so it can read open tabs, fill in forms, file expenses and download files for you. That puts an agent in the browser where your logged-in accounts already live.