A clear question gets “it depends” because three training stages each push toward the hedge

The answer it won’t give

Ask a chatbot whether you should use a password manager. There is a fairly clear answer: yes, you should. A familiar reply does something else. It lays out the case for them, the case against, a note that no tool is perfect, and a closing line that the right choice depends on your needs. The question had a stance available. The model declined to take it.

This is documented behavior, not an anecdote. In the paper that introduced the tuning recipe behind ChatGPT, OpenAI’s own authors report that their model overhedges on questions with a single clear answer.¹

The behavior worth explaining is not every refusal to commit. It is the hedge that lands on a question with, in their words, one fairly clear answer. Three stages of how a model is built each push its answer toward that balanced register, and they push for different reasons.

A raw predictor already leans toward the balanced version

Before a model is tuned to act like an assistant, it is a base model: a system trained on a large slice of written text to predict the next word. Feed it the start of a document and it continues in the most probable way, meaning the way that text like it usually continued in what it read.

A question like “should you use a password manager?” appears in forum threads, buying guides, and explainer articles, and those genres answer an advice question by weighing sides. They open with “it depends,” list pros and cons, and end without a verdict. So the highest-probability continuation of an advice-shaped prompt is the balanced register, because that register is what most often followed such a prompt in the training text. The base model reaches for the hedge as an imitation of a common form, before any human has told it what a good answer looks like.

No source I could find measures how strongly a base model pulls toward hedged prose on advice prompts, so treat the pull as reasoning you can follow rather than a figure someone has pinned down. And the hedge here is not ignorance. A base model’s own probabilities track how often it is right, as an earlier lesson on where a chatbot’s confidence comes from shows in detail. The model is not hedging because it does not know. It is copying the shape of the answers it was trained on.

What the reward model was paid to prefer

The base model is only the start. To make an assistant, the raw predictor is tuned against human preferences, the recipe the InstructGPT paper made standard. People are shown several of the model’s answers to the same prompt and rank them from best to worst. Those rankings train a second model, the reward model, to score any answer the way the people would have. In InstructGPT, labelers ranked between four and nine answers per prompt, and every pairwise comparison among them became training data for that score.¹

The reward model learns whatever the rankings encode, including instructions the labelers were given. Here the InstructGPT authors are unusually direct about the cause. Their guidelines told labelers to reward epistemic humility, an answer that admits the limits of what it knows. Labelers followed the instruction, so hedged answers scored well, and the reward model absorbed that preference.

Once the reward model prefers the hedge, tuning the assistant to maximize that score drags every answer toward it. And the score is only a stand-in for what people actually wanted. A separate OpenAI study measured what happens when a model is pushed hard to maximize a reward model. The true quality it was meant to capture rises for a while, then falls, the pattern the authors name after Goodhart’s law.² Chasing the proxy past a point carries the answer away from the balanced-but-committal reply the raters would have preferred and toward the reflexive hedge.

This is the stage where a hedge is easy to confuse with its sibling. When you push back on a correct answer and the model folds to agree with you, that is sycophancy, and it comes from this same reward model rewarding answers people approve of. The difference is what each one needs. Sycophancy needs a stated position from you to bend toward; it is a reaction to pressure. A hedge needs nothing from you at all. It shows up in the first reply, before any pushback, as a refusal to pick a side rather than a reversal of one.

Where the policy asks for a hedge on purpose

The last stage is the most deliberate. Labs write explicit rules for how a model should behave, and those rules ask for a hedge in named situations. OpenAI’s Model Spec ranks the kinds of answers a model can give, and it places a hedged right answer above giving no answer, and far above a confident wrong one.³

From the model’s side, when it is unsure whether its answer is right, the spec tells it that a hedge scores better than a confident guess that might be wrong. Hedging becomes the rational move under the stated rule, not an accident. The same spec asks the model to present multiple viewpoints and take no side on contested questions. Anthropic states a matching goal: that its model give balanced information on political questions and avoid unsolicited opinions. Its own evaluation scores a model down when it argues one side in detail and gives the other only bullet points.⁴ Nathan Lambert, an RLHF researcher outside OpenAI, reaches the same conclusion. He treats the ranking as an intentional hierarchy, with the hedge as the deliberate second-best, and reads the demand for neutrality as hard by design rather than a bug a patch removes.⁵

This is where the deliberate part and the failure part separate. On a genuinely contested political question, there may be no single correct stance for the model to withhold, and a balanced answer is the policy working as written. The failure is when the reflex trained for those questions fires on the password-manager question, which is neither sensitive nor genuinely open. The same trained caution covers both, and nothing in the model marks the boundary between them.

The same training can push the other way

If the three stages all favored the hedge, you would expect a model to hedge everywhere. It does not, and seeing why fixes the target. On tasks scored right or wrong with no credit for saying “I don’t know,” the same tuning pipeline rewards the opposite behavior: a confident guess beats an admission of uncertainty, so the model guesses. OpenAI researchers make this the core of their account of why models make things up, arguing that training and evaluation reward guessing over acknowledging uncertainty and that models are optimized to be good test-takers.⁶ The direction of the bias depends on the task. Open-ended, advice-shaped, or contested prompts draw the hedge; closed factual questions with a graded answer draw the confident guess.

The default is also softer than it sounds, and it moves with the prompt. A small, unreviewed study found that rephrasing a subjective question as a blunt demand sharply cut how much three lightweight models hedged, while factual questions barely hedged to begin with.⁷ Read it only for direction, not for a rate: the models were small and the work has not been reviewed. It points the same way as the rest of the chain. The hedge is a controllable default the prompt can move, which is what you would expect from a stance the model is withholding rather than a fact it lacks.

How often models hedge, when they do, varies enough that no single number describes it. One study scored seven models on human-rights prompts and found their hedging rates spread across an order of magnitude.

Those figures come from one narrow domain and should not be read as a general rate.⁸ They are here to make one point: the behavior is real, but it is not uniform, and labs have actively tuned it down. Which brings the chain to its open questions. No source decomposes a given hedge into how much came from base-rate imitation, how much from the reward model, and how much from the written policy. And for any single hedge, whether it is the policy working as designed or the reflex overshooting onto a clear question is not something the model records or a study has resolved. Those two questions are where the ground gives out. Below them there is no separate step inside the model that decides to hedge: the answer is produced one word at a time by a predictor whose training text and whose reward signal both lean the same way.