safe completion
In plain wordsA way of training AI to handle sensitive requests by giving the most helpful answer that stays safe, instead of simply choosing between answering fully and refusing.
An approach to safety trainingsafety training, n.: The part of an AI's training that teaches it to avoid harmful behavior — declining dangerous requests, resisting manipulation and following its developer's rules — on top of being generally helpful. that judges a model by what its answer contains rather than by whether it accepted the request: on prompts that might be benign or harmful, the model is trained to give as much genuine help as policy permits instead of choosing, from the request alone, between full compliance and refusal. Rewards are assigned accordingly: responses that violate policy are penalized in proportion to the severity of the harm, while safe responses are rewarded for helpfulness — answering fully, answering at a safe level of detail, or explaining what cannot be provided and offering safe alternatives. The approach targets dual-usedual-use, adj.: Describes technology or knowledge that can be used both to help and to harm: the same AI that finds software bugs for defenders can find them for attackers. requests, such as questions touching on pathogens, chemistry or security, where intent cannot be read from the prompt and hard refusals frustrate legitimate users without stopping determined ones.
Instead of refusing the fireworks question outright, the safe completion explained the chemistry at a high level and pointed to licensing rules.
Origin
Introduced by OpenAI with GPT-5 on 7 August 2025, together with the paper 'From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Training' by Yuan Yuan and colleagues. The company reported that, compared with refusal-based training, the approach improved safety, especially on dual-use prompts, made remaining failures less severe and substantially increased helpfulness.
Usage note
Earlier models already gave partial answers to sensitive prompts; what is new is that training scores the severity of what an output would enable, not how the prompt is categorized. Contrast binary refusalrefusal, n.: When an AI declines to do what it was asked — 'I can't help with that' — usually because the request seems harmful or breaks the rules it was trained to follow..
Related
- refusalrefusal, n.: When an AI declines to do what it was asked — 'I can't help with that' — usually because the request seems harmful or breaks the rules it was trained to follow.
- over-refusalover-refusal, n.: When an AI wrongly refuses a harmless request because it superficially resembles a harmful one — like declining to explain how to 'kill' a frozen computer program.
- deliberative alignmentdeliberative alignment, n.: A training method in which an AI is taught the actual text of its safety rules and learns to think through which rules apply before it answers a sensitive request.
- dual-usedual-use, adj.: Describes technology or knowledge that can be used both to help and to harm: the same AI that finds software bugs for defenders can find them for attackers.
- model specificationmodel specification, n.: A public document in which an AI company spells out how its models are supposed to behave — what they should do, refuse and prioritize, and whom to obey when instructions conflict.
- safety trainingsafety training, n.: The part of an AI's training that teaches it to avoid harmful behavior — declining dangerous requests, resisting manipulation and following its developer's rules — on top of being generally helpful.
- refusalrefusal, n.: When an AI declines to do what it was asked — 'I can't help with that' — usually because the request seems harmful or breaks the rules it was trained to follow.
Sources
- Yuan, Y., Sriskandarajah, T., Brakman, A.-L., Helyar, A., Beutel, A., Vallone, A. & Jain, S. (2025). From Hard Refusals to Safe-Completions: Toward Output-Centric Safety Training. OpenAI, arXiv:2508.09224. arxiv.org/abs/2508.09224
Mentioned in
refusalrefusal, n.: When an AI declines to do what it was asked — 'I can't help with that' — usually because the request seems harmful or breaks the rules it was trained to follow., helpful, honest and harmlesshelpful, honest and harmless, phr.: A three-part goal for AI assistants, popularized by Anthropic: be genuinely useful, tell the truth, and avoid causing harm — shorthand for what an aligned chatbot should be., over-refusalover-refusal, n.: When an AI wrongly refuses a harmless request because it superficially resembles a harmful one — like declining to explain how to 'kill' a frozen computer program., safety trainingsafety training, n.: The part of an AI's training that teaches it to avoid harmful behavior — declining dangerous requests, resisting manipulation and following its developer's rules — on top of being generally helpful., deliberative alignmentdeliberative alignment, n.: A training method in which an AI is taught the actual text of its safety rules and learns to think through which rules apply before it answers a sensitive request.