What Do Human Judges Reward? Position, Length and Style Effects in Pairwise Human Preferences for AI Responses

Purpose. Pairwise human preferences are the raw material for ranking language models and for training them with human feedback. They are treated as a measure of quality, but humans also react to features of the answer that may not track quality. We ask how strongly human judges reward position, length and formatting style, and how much this changes the resulting model ranking. Design and methods. We analysed 57,477 human votes on pairs of responses from 64 models in the Chatbot Arena 55k data (30.9% ties), 3,355 expert votes from 65 judges in MT-bench, and a sample of 120,000 Reddit comment pairs (SHP). A Bradley–Terry model with model strengths was extended with a position term and with standardised differences in length and style (line breaks, lists, headers, bold text, code blocks, refusal-style openers). Standard errors were clustered by prompt, predictive performance was assessed by grouped cross-validation over prompts, and rankings with and without the bias terms were compared. Findings. Position mattered little (the first response won 50.5% of decided pairs; adjusted log-odds +0.025), but length and style mattered a great deal. A response three times longer than its rival (one SD of the log ratio) had 1.52 times the odds of winning after accounting for model strength (95% CI of the log-odds 0.387 to 0.447); line breaks (+0.315 per SD) and refusal-style openers (−0.318) were the strongest style signals. The same length preference appeared in MT-bench (odds ratio 2.83 per SD) and in Reddit (63.1% of pairs preferred the longer comment). Adding length and style terms reduced held-out log-loss from 0.6198 to 0.5894 and raised accuracy from 64.8% to 68.0%. Correcting for these terms changed the ranking of models only moderately (Kendall's tau 0.857; largest rank change 16 places) and left the top four unchanged. Originality. The study estimates, on three independent human-preference data sets and with prompt-clustered inference, how much human judges reward superficial features, and shows with held-out prediction that these behavioural effects are predictable signals that a preference model can identify and correct for.