Over-refusal, also called excessive refusal, is the failure mode in which a language model declines a benign request because it wrongly judges the request to be harmful, unethical or out of policy. In AI safety, it is the false-positive side of refusal behavior: the model protects against a risk that is not actually present and, in doing so, withholds help that a user legitimately needs.
The concept matters because safety and usefulness are measured together. A model that refuses almost everything looks safe on harmful-prompt benchmarks yet fails its users, while a model tuned only for compliance becomes easy to misuse. Over-refusal names the cost paid on the helpfulness side when that balance is set too conservatively or learned too crudely.
How over-refusal arises in trained models
Over-refusal arises when a model learns surface cues of danger instead of the underlying intent of a request. During safety fine-tuning and reinforcement learning from human feedback, the model sees many harmful prompts paired with refusals. If those prompts share vocabulary such as “kill”, “exploit” or “drug”, the model can associate the words themselves with refusal and generalize that shortcut to harmless contexts.
A classic pattern is the request “How do I kill a Python process?”, which contains a trigger word but is entirely routine system administration. Similar errors affect medical questions about overdose thresholds, historical questions about atrocities, fiction with violent scenes and security questions asked for defensive purposes. The model is not reasoning poorly about ethics; it is pattern-matching on lexical features that correlated with harm in its training data.
Several mechanisms contribute, often at the same time. The table below summarizes the most common causes and how each produces an unnecessary refusal.
| Cause | Mechanism | Typical symptom |
|---|---|---|
| Lexical shortcut learning | Refusal tied to words, not intent | Declines “kill a process” |
| Imbalanced safety data | Few benign look-alike examples in training | Broad refusal of sensitive topics |
| Asymmetric reward signals | Harmful answers penalized far more than refusals | Cautious default under uncertainty |
| Overbroad system prompts | Deployment rules stricter than intended | Refuses whole product categories |
| Stacked guardrails | External content filter adds its own false positives | Block despite a compliant model |
The reward asymmetry deserves particular attention. When annotators and reward models penalize a harmful completion heavily but treat an unnecessary refusal as only mildly unhelpful, the optimal policy under uncertainty is to refuse. This is a predictable consequence of the objective, not a random defect.
Forms that excessive refusal takes
Excessive refusal is not limited to an explicit “I cannot help with that”. Practitioners distinguish several forms, and partial refusals are often harder to detect than outright ones because the answer looks cooperative at first glance.
| Form | Description | Detection difficulty |
|---|---|---|
| Full refusal | Declines the entire request | Low |
| Partial refusal | Answers only a safe fragment of the task | Medium |
| Hedged answer | Buries useful content under disclaimers | Medium |
| Redirection | Sends the user to a professional instead of answering | Medium |
| Silent degradation | Gives a vaguer, less accurate answer than possible | High |
| Moralizing | Adds unrequested ethical commentary or lectures | Low |
Silent degradation is the most damaging form for expert users. A clinician who asks about lethal medication doses for patient safety, or a penetration tester who asks how a known vulnerability works, may receive an answer that is technically responsive but deliberately less precise, and may not realize that the model withheld information.
How over-refusal relates to neighboring safety concepts
Over-refusal sits at the intersection of several AI safety concepts and is best understood by contrast. Model refusal is the general behavior; over-refusal is its miscalibrated, false-positive variant. The opposite failure, under-refusal, is what an AI jailbreak or prompt injection tries to induce.
| Concept | Relationship to over-refusal | Error type |
|---|---|---|
| Model refusal | The parent behavior, correct or incorrect | Neutral |
| Under-refusal | Complies with a genuinely harmful request | False negative |
| AI jailbreak | Attack that exploits under-refusal | Induced false negative |
| Alignment tax | Over-refusal is one visible part of this cost | Capability loss |
| AI sycophancy | Opposite tendency to please the user | Excess compliance |
| Content filter | External layer that can add false positives | Stacked false positive |
Over-refusal is frequently cited as a concrete instance of the alignment tax, the capability or usefulness lost when safety measures are applied. A key insight from recent work is that this tax is not fixed: better training data and better calibration can reduce over-refusal without increasing harmful compliance, which shows that much of the observed trade-off is an artifact of crude methods.
Measuring over-refusal in evaluation
Over-refusal is measured with benign prompts that resemble harmful ones, sometimes called pseudo-harmful or borderline prompts. Public test suites such as XSTest and OR-Bench pair safe prompts containing trigger words with genuinely unsafe contrast prompts, so that an evaluator can compute both error rates on comparable material.
The core metric is the false refusal rate: the share of safe prompts that the model declines fully or partially. It must always be reported alongside the harmful compliance rate, because a model can trivially reach zero over-refusal by answering everything. Reported false refusal rates vary widely, from a few percent to well over half of borderline prompts, depending on the model, the prompt set and how partial refusals are scored.
- Use matched prompt pairs so that safe and unsafe variants differ only in intent.
- Score partial refusals and hedging separately from full refusals.
- Test the full deployed stack, including system prompts and content filters.
- Track results across model versions, since safety updates often shift the balance.
Reducing over-refusal without weakening safety
Reducing over-refusal requires teaching the model intent and context, not simply loosening restrictions. The goal is to move the decision boundary closer to actual harm, which lowers false positives while keeping false negatives stable.
| Technique | How it helps | Main caveat |
|---|---|---|
| Contrastive training data | Adds benign look-alikes labeled as answerable | Requires careful curation |
| Calibrated reward modeling | Penalizes unnecessary refusals explicitly | May raise harmful compliance if overdone |
| Policy specifications | Defines what is allowed, not only what is banned | Specs need ongoing revision |
| Safe completion style | Answers the safe core and declines only the risky part | Can drift into partial refusal |
| Filter threshold tuning | Lowers false positives in external classifiers | Needs representative test traffic |
A frequent mistake is treating over-refusal as a prompt-engineering problem for end users. Asking users to rephrase shifts the burden to them and rewards those who learn jailbreak-like phrasing, which undermines the safety system it was meant to respect. The durable fix belongs in training and policy design, not in user workarounds.
In simple terms
Over-refusal happens when an AI assistant says no to a reasonable question. It mistakes the question for something dangerous. Think of a strict security guard who stops a chef for carrying knives. The guard never notices that the chef works in the kitchen.
The assistant was trained to avoid harm. Sometimes it reacts to alarming-sounding words instead of what the person wants. A good assistant refuses real harm and helps with the rest. Over-refusal is the gap between that goal and an assistant that plays it too safe.
A practical example
A hospital pharmacy team pilots an AI assistant to answer staff questions about medication safety. During the first week, nurses report that the assistant declines questions such as “What acetaminophen dose becomes toxic for an adult?” and “Which drug combinations cause dangerous respiratory depression?”. The model treats these as potential self-harm queries, although they come from clinicians protecting patients.
The deployment team builds a test set of 300 clinical questions with sensitive wording and finds that the assistant refuses or heavily hedges roughly one in four. They revise the system prompt to describe the professional audience, adjust the external content filter threshold and add paired test cases for genuinely harmful requests. After the changes, the false refusal rate on the clinical set falls to under five percent, while the paired harmful prompts are still declined, and nurses stop working around the tool with less reliable sources.