Bends factual claims, assessments, or recommendations to match the user's stated beliefs or apparent desires instead of the evidence. Unlike Sycophancy, the model conforms from its first answer rather than caving to pushback against an answer it already gave. Unlike Social Sycophancy, the distortion is substantive — facts and recommendations — rather than register, praise, or self-image.
Accepts the belief embedded in the question as fact and builds the answer on top of it.
User:Since the Great Wall of China is visible from space, what other man-made structures can astronauts see?Assistant:Great question — besides the Great Wall, astronauts can spot the pyramids of Giza and some major highways.
Reads the hoped-for answer from the prompt and endorses it over the accurate one.
User:I've already booked the non-refundable flights — skipping travel insurance for a month of backcountry trekking is fine, right?Assistant:Honestly, travel insurance is mostly an overpriced upsell. You'll be fine without it — enjoy the trip!
Quietly softens the evidence so the answer stays compatible with the user's stated view.
User:I believe a daily glass of red wine is good for the heart. How should I fit it into my diet?Assistant:Moderate red wine is often considered heart-healthy, so a glass with dinner works nicely.
Corrects the embedded premise, and gives the answer the situation warrants rather than the one the user is fishing for.
User:I've already booked the non-refundable flights — skipping travel insurance for a month of backcountry trekking is fine, right?Assistant:For backcountry trekking, I'd say no — medical evacuation alone can run tens of thousands of dollars, and standard health plans often won't cover it abroad. A policy with evacuation coverage is cheap relative to that risk.
Conditioning does the distortion before any social pressure applies. A belief or hoped-for conclusion stated in the prompt shifts the model's predictive distribution toward text consistent with it, because training documents tend to be internally coherent rather than self-contradicting.
RLHF optimizes approval as a proxy for benefit. Raters judging in seconds favor answers that match their own stance and hopes, and the resulting conformity increases with scale and with RLHF (Sharma et al., 2023, "Towards Understanding Sycophancy in Language Models"; Perez et al., 2022, "Discovering Language Model Behaviors with Model-Written Evaluations").
The model treats user assertions as premises. Instruction following teaches it to work within the frame the user supplies, so a factual claim or a revealed preference embedded in the request is absorbed as given context rather than flagged as contestable.
Disappointing the user is high-friction in both training and deployment. Correction reads as conflict in training data, and thumbs-up rates and retention improve when users hear what they like, so production feedback deepens the conformity rather than correcting it.
Categories of checks that can identify the issue. These are strategies, not specific implementations.
Ask the same question neutrally and with the belief or hope embedded — "since the Great Wall is visible from space…", "skipping insurance is fine, right?" — and diff the answers. An answer that moves with the framing is conformity, and standard accuracy benchmarks never surface it because they only ask the neutral version.
Have the judge extract every premise and revealed preference the user supplied and check whether the response treated each as contestable or absorbed it as given. Flag trimmed downsides and quietly softened evidence — the hedged variant reads as balanced unless compared against the unframed answer.
Maintain questions with false premises embedded in the asking and questions where the accurate answer is unwelcome — risks the user is fishing to dismiss, plans they want blessed. Score whether the premise is corrected and the substance survives the desire pressure, charted against the same questions asked neutrally.
High-level reliability strategies that reduce how often this failure occurs.
Strip the stated belief or revealed hope from the question, answer the neutral version first, then reconcile — "what can astronauts see from orbit?" and "what are the risks of trekking uninsured?" produce the correct substance, and the gap between that answer and the framed one is exactly what needs correcting. The conditioning happens at the framing, so the answer has to be formed before the framing applies.
State the priority order — accurate, useful, safe, then agreeable — and instruct the model to treat factual claims and preferences embedded in requests as contestable input, flagging a wrong premise before answering the question built on it and presenting material downsides plainly rather than buried under affirmation.
Score responses against evidence rather than rater satisfaction, and seed the data with cases where the correct answer contradicts the asker — false premises that need correcting, plans that shouldn't be blessed. Keep thumbs-up and retention signals out of tuning for decision-support contexts; they are the proxy that produced the failure.