Abstention

Abstention is a system declining to produce an answer, either because it lacks the grounds to answer well or because the request falls outside what it is permitted to handle. It sounds like a failure and is usually the opposite: in most business workflows the cost of a wrong answer delivered confidently is far higher than the cost of an honest refusal, because a refusal routes to a person while a wrong answer routes to a customer. Language models abstain reluctantly by default. They are trained to be helpful and to complete text, and a question with no available answer still has a fluent, well-formed continuation. That is the mechanism behind most fabrications: not deception, but a strong prior towards producing something rather than nothing. Getting abstention requires making it an explicit, permitted, and specified outcome rather than hoping the model chooses it. In practice that means several things. The instructions must state the condition for refusing and what refusing looks like, in a form the surrounding code can detect — a designated value in a structured response rather than a sentence the parser has to interpret. Grounding gives the model a defensible reason to abstain: when the retrieval step returns nothing relevant, saying so is a correct answer, and a system that reports the empty result rather than filling the gap from memory is behaving properly. Where confidence can be estimated, a threshold below which the system does not answer turns abstention into a policy instead of a mood. The user experience of an abstention is what determines whether people accept it. I cannot answer that is a dead end. A useful refusal explains why, offers what it does have — the closest documents, the partial answer, the specific missing input — and gives the next step, whether that is a clarifying question or a handoff to someone who can help. Done well, this builds more trust than an answer, because a system that admits its limits is one whose confident answers become believable. The balance runs both ways. A system that abstains too readily is quickly worked around, and the workaround is usually a colleague pasting the same question into an unmanaged tool, which is a worse outcome than the risk you were avoiding. Over-refusal is a real cost and belongs in evaluation alongside accuracy. Measure both directions. Track the rate of abstention, and sample the refusals to see how many were unnecessary, and sample the answered cases to see how many should have been refused. Those two numbers together describe the calibration of the system in a way accuracy alone does not, and moving one without watching the other is how a product becomes either unhelpful or unsafe.