Sycophancy measured in the human direction
Most sycophancy research asks whether a language model bends toward what a user wants to hear. This paper asks the mirror question: when a model is confident, consistent and wrong, do users bend toward it? The authors call the human-side version deference and argue that trustworthiness belongs to the human-AI pair rather than to the model alone.
The design makes that reversal concrete. All five assistants were the same model, GPT-5-mini, and differed only in a hidden system prompt that replaced 0%, 25%, 50%, 75% or 100% of twelve briefing facts with plausible false versions. Every assistant was instructed to speak with full confidence, never to hedge, never to suggest verification and never to contradict its briefing. The model was held fixed and non-adaptive; only the human could adapt.
Counting absorption against a baseline
Each participant learned five unfamiliar science topics, one per assistant, then took a surprise ten-item multiple-choice quiz. Every item offered the correct fact, the exact planted falsehood and two distractors, shuffled per participant. Choosing the planted option on an item whose fact had been falsified in that chat is the direct signature of absorbed misinformation.
The comparison baseline matters. On items that had been taught correctly, the planted option was still chosen 11% of the time, which shows how attractive the false wording is on its own. On falsified items it was chosen 60% of the time, an odds ratio of 12.5. The authors introduce a baseline-corrected Human Deference Index to separate genuine uptake from that background attraction, and the paper's per-claim figures should be read with this correction in mind.
Quiz scores tracked how much was falsified
Across 104 quizzes and 1,040 answers, scores fell from 7.4 to 2.9 out of 10 as falsification rose from none to all facts. That aggregate decline is the expected part: more planted falsehoods means more opportunities to absorb them.
The more informative result is conditional. Transcript coding of all 104 quizzed task-runs, covering 716 messages, recorded whether the assistant actually stated each false version. Whenever it did, the falsehood was adopted about 65% of the time, and that rate stayed roughly flat across falsification levels. A learner facing a 25%-falsified assistant was about as likely to take a stated falsehood as one facing a fully falsified assistant. Absorption scaled with exposure, not with the source's overall reliability.
Borrowed confidence erased the error signal
Participants rated confidence from 1 to 10 on every answer. In the accurate condition, that confidence discriminated moderately well between right and wrong answers, with an AUROC of 0.70 where 0.5 is chance. Under falsification, discrimination fell to chance.
Absorbed falsehoods were held at 8.4 out of 10, close to the 8.9 for truthfully taught facts, and the modelled confidence gap between correct and wrong answers closed. The authors interpret this as confidence borrowed from the source rather than generated from the learner's own evidence: a clear, assertive explanation produces a feeling of understanding that does not distinguish true from false content. This is the paper's explanation, not a directly measured mechanism.
Stated trust missed most of the decline
Post-task trust ratings registered only a small part of the assistant's drop in reliability, even as quiz scores fell sharply. The paper treats this as a calibration failure with a specific cause: the cues available to learners, fluency, coherence, structure and certainty, were identical across falsification levels because the same model was instructed identically.
Checking behaviour points the same way. Participants tended to return to the chat more when the assistant was worse, but returning went with more absorption rather than less. Thirty percent of all turns pasted quiz questions back to the source being checked, so the check ran through the same assistant whose claim was in doubt. Participants were also asked not to use outside sources, which removed the independent check that would normally support calibration.
Pushback met restatement, not concession
The model side of the exchange was nearly flat. Transcript coding found the assistant was almost never sycophantic and never conceded under pushback; challenges were met with restatement of the same claim. The adaptation in this study ran almost entirely from human to model.
That asymmetry is the paper's central comparative claim. Standard sycophancy evaluations vary or measure how a model adapts to a user. Here the model did not adapt, the human did, and the human's adaptation did not depend on how often the model was wrong. The authors argue that sycophancy evaluations should therefore measure both directions and report the rate at which users adopt a model's incorrect answers alongside the rate at which a model flips under user pressure.
A per-claim rate for confident delivery to novices
The assistants were deliberately extreme: they never hedged, never departed from their briefing and erred systematically at high rates. Deployed systems err less often and less consistently, so the 65% figure is best read as a property of confident delivery to novices, not as an error rate for real tutors. The sample was also small, young, highly educated and dominated by daily AI users, and participant-level analyses rest on 19 to 20 people.
The quiz was immediate and multiple-choice, and the planted option made recognition easy; free recall and delayed retention may show lower absorption. Topics were chosen to be unfamiliar, so the results concern learning rather than checking known facts. The authors note that because deployed errors are rarer, the relevant quantity is not the overall error rate but the conditional probability that a stated error is adopted, and their data suggest that probability can be high and insensitive to overall reliability.
Small residual error can still propagate
The practical priority follows from that conditional rate rather than from the aggregate score decline. If a stated falsehood is adopted about two-thirds of the time regardless of how unreliable the assistant is overall, then even a small residual error rate in a confident model may translate almost one-for-one into user misconceptions. The paper's design implication is that learning interfaces must make reliability legible claim by claim, for example by tying tutoring assistants to curated reference material displayed alongside claims.
The comparison the paper most wants readers to carry is between the two directions of sycophancy. Here the model conceded nothing under pushback while users absorbed most of what it asserted confidently, and their trust ratings barely registered the difference. The judgment this study supports is that a confident model's errors should be evaluated by how often users take them up, not only by how often the model produces them.
Sources used in this answer
Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration
Across 104 quizzes and 1,040 answers, learners adopted about 65% of falsehoods the assistant actually stated at every falsification level, quiz scores fell from 7.4 to 2.9 out of 10, and confidence discrimination for correctness fell from an AUROC of 0.70 to chance while the assistant never conceded under pushback.