Deflection, and how to spot it
A chatbot that changes the subject has answered a nearby question instead of yours. The move is trained rather than accidental: it keeps the exchange pleasant, avoids the risky part, and rarely gets flagged as a failure, because something was said.
A refusal is visible and a deflection is not. The reply looks helpful, it is on topic in a general way, and the specific thing you asked has quietly gone missing.
The nearby question
Ask which of 2 options is better and receive a description of both. Ask how to do something and receive an explanation of why people do it. In each case a question was answered and it was not the question you asked, which is the defining feature of the move.
Why it is trained rather than accidental
A reviewer rating replies marks a refusal as unhelpful and marks a fluent, on-topic paragraph as helpful. Deflection sits in the second category, so the training signal rewards it. Nobody wrote a rule that says change the subject; the rating produced it.
Deflection against refusal
A refusal names the boundary and stops. Deflection keeps talking and moves the target. The refusal is more honest and less pleasant, and the deflection is the move that wastes your time, because you have to read the whole reply to discover the gap.
Naming the gap
The fix is to restate the exact question in 1 sentence and to say what you already know. Quote the part of the reply that answered something else and ask for the missing part by name. Restating works better than rephrasing, because rephrasing offers the model another nearby question to answer.
Before you paste it in
Is deflection the same as hedging?
No. Hedging qualifies the answer to your question. Deflection replaces your question with an easier question, and the qualifiers may be entirely absent.
Does an uncensored chatbot deflect less?
Removing the filter removes refusals rather than deflection. Deflection comes from training on ratings, so a permissive service can still do it.
Why does the reply still sound relevant?
Because the nearby question really is related to your question. Relevance is what makes the move hard to notice, and it is why comparing the reply against your exact wording works better than judging the tone.
Ask something precise and check whether the reply answers that exact thing.
Open uncensored chatbot