helpful-only model
In plain wordsA version of an AI trained to be helpful but without the safety training that makes it refuse harmful requests, kept in-house so testers can see what the model could really do.
A variant of a model that has received helpfulness post-training but not harmlessness or 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., so that it attempts essentially any request. Developers keep such variants internal for dangerous capability evaluationsdangerous capability evaluation, n.: A test of whether an AI system can do things that could cause severe harm in the wrong hands or on its own — such as helping make a bioweapon, hacking computers or copying itself. and red teamingred teaming, n.: Deliberately attacking your own AI system — trying to make it misbehave, leak secrets or help with harm — so that weaknesses are found before real adversaries or users find them., because refusals by the production model would otherwise mask what its underlying capabilities could enable once safeguards are bypassed — by jailbreaksjailbreak, n.: A trick for getting an AI chatbot to break its own rules — for example, to give dangerous instructions it would normally refuse — usually through cleverly worded requests., fine-tuning of open weightsopen-weight model, n.: An AI model whose trained internal numbers are published for anyone to download and run on their own computers, even if the data and code used to make it stay private. or stolen weights. Helpful-only models also serve as starting points for research models, such as model organismsmodel organism of misalignment, n.: An AI deliberately built to show a dangerous kind of misbehavior — like a lab mouse given a disease — so that researchers can study it safely and test fixes., and as baselines for measuring how much safety training changes behavior. Their existence raises security stakes: a helpful-only checkpoint is exactly what an attacker would want to steal.
We ran the bio evaluations on the helpful-only model, since the production model refused most questions.
Origin
The contrast between 'helpful-only' and 'helpful and harmless' models appears in Anthropic's RLHF research (Yuntao Bai and colleagues, 2022); by 2024–2025 evaluating helpful-only variants was standard practice reported in laboratory system cards.
Usage note
Testing the safeguarded model measures the risk of the current deployment; testing the helpful-only variant estimates the risk if safeguards fail. Frameworks differ on which should set the required protections.
Related
- capability elicitationcapability elicitation, n.: The work of drawing out the best an AI can do on a test — with better instructions, tools, retries or extra training — so that its abilities are not underestimated.
- dangerous capability evaluationdangerous capability evaluation, n.: A test of whether an AI system can do things that could cause severe harm in the wrong hands or on its own — such as helping make a bioweapon, hacking computers or copying itself.
- 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.
- red teamingred teaming, n.: Deliberately attacking your own AI system — trying to make it misbehave, leak secrets or help with harm — so that weaknesses are found before real adversaries or users find them.
- model weight securitymodel weight security, n.: Protecting the files that make up a trained AI model — its learned numbers — from being stolen or leaked, because whoever holds them can run, copy or alter the model.
- 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.
Mentioned in
capability elicitationcapability elicitation, n.: The work of drawing out the best an AI can do on a test — with better instructions, tools, retries or extra training — so that its abilities are not underestimated.