This preregistered paper uses incentivized Krupka–Weber norm elicitations in two canonical games (Dictator and Ultimatum) on representative UK and US samples (total N=2,658) to test whether people apply different social-norm expectations to humans vs. LLMs acting on their behalf. Results show: (1) the modal norm (50/50 split) remains unchanged, (2) offers made by LLMs (when no acceptance is available) are judged overall as less socially appropriate than identical offers by humans, particularly for self-favoring (low) splits, and (3) LLMs serving as enforcers (Responders) are not judged less legitimate for rejecting unfair offers — but people judge rejecting very generous offers from LLM-Proposers as more appropriate than identical offers from human Proposers .
速读内容
Key finding — modal norm unchanged but overall appropriateness shifts
The most socially appropriate split (the modal response) in the Dictator Game is the equal 5/5 split in both Human and LLM conditions (bootstrap Δ = 0) .
Despite the same modal norm, mean appropriateness ratings across splits differ systematically between Human and LLM Dictators, showing a downward tilt for LLMs overall .
Direction and size of the LLM effect in Dictator Game — lower appropriateness for self-favoring offers
Amount given to Recipient
Difference in means (Human − LLM)
0
-0.010
1
-0.131
2
-0.171
3
-0.182
4
-0.209
5
-0.196
Wilcoxon / clustered tests show significantly lower appropriateness ratings for LLM-made splits for most low-to-mid offers (Holm–Bonferroni corrected), with the effect concentrated on splits favoring the Dictator (0–4) .
Effect-size comparison: Cliff’s Δ_low(0–4) = -0.163 vs. Δ_high(6–10) = -0.047; the difference is significant (bootstrapped CI excludes zero), indicating the tilt is stronger for selfish splits .
Divergence in plurality and dispersion of judgments (norm disagreement)
While ~40% of participants nominate 5/5 as top rating in both conditions, more participants pick extreme generous options (e.g., give all £10) as top-rating in the LLM-Dictator condition, increasing divergence (KL divergence higher around dictator-favoring splits) .
Ultimatum Game: Legitimacy of enforcement and asymmetry in reactions to Proposers
Human Proposer — LLM Responder: aggregated appropriateness ratings for rejection do not differ significantly from Human–Human (z = -1.37, p = .171). People do not consider LLM-responders to be less legitimate enforcers of fairness .
LLM Proposer — Human Responder: ratings differ significantly from Human–Human (z = 5.00, p < .001). Crucially, rejection of high (very generous to Responder) offers made by LLM-Proposers is judged more appropriate than identical offers by human Proposers (significant differences for offers 6–10) .
Interpretation: people penalize LLMs both for ungenerous behavior when they are sole decision-makers and for being unreasonably generous when that generosity harms the human who delegated the decision; but LLMs retain legitimacy as enforcers (Responders) .
Methods & robustness
Two pre-registered, incentivized online experiments on Prolific with representative UK (Dictator; N≈1,160) and US (Ultimatum; N≈1,505) samples; Krupka–Weber incentivized coordination to elicit modal normative judgments; nonparametric clustered tests, bootstrapping, and linear mixed models used for robustness .
Results robust to alternative exclusions and mixed-model specifications (see Tables S2–S3 and Figures S2–S3) .
Implications for design, delegation, and policy
Delegation to LLMs is not uniformly acceptable: people expect LLMs to avoid both ungenerous and unreasonably generous actions when acting as first-movers, but accept LLMs as norm-enforcers (Responders) .
Policy design should reflect the asymmetric public attitudes: allow algorithmic enforcement roles in some contexts while being cautious about delegating sole first-mover decisions with social stakes .
思维导图
深度阅读
以下为对论文 “Do people expect different behavior from large language models acting on their behalf? Evidence from norm elicitations in two canonical economic games” 的详尽、逐节解构式分析。为便于追溯,凡从原文得到的结论或推断,句末均添加了页码溯源标注(格式 [page::页码])。我将严格遵循报告结构(元数据、逐章解读、图表深度解读、估值/方法学评述、风险与限制、批判性视角与结论性综合),并在图像处以 markdown 格式嵌入报告中所用的重要图片。
一、元数据与概览(引言与报告概览)
报告标题与作者:文章标题为 “Do people expect different behavior from large language models acting on their behalf? Evidence from norm elicitations in two canonical economic games”,作者为 Paweł Niszczota 与 Elia Antoniou,隶属 Poznań University of Economics and Business 与 University of Cyprus(详见作者署名与机构信息)。
核心议题与研究问题:论文检验公众在两类经典博弈(Dictator Game 与 Ultimatum Game)情境中,是否对“代表人类决策的 LLM”(代为决策或代为执行)的行为抱有不同的社会规范期待,以及这种规范差异是否能解释人类在与 AI 相关互动中表现出的行为差异。研究采用 Krupka–Weber 规范启发任务在 UK 与 US 的代表性样本上进行两项预注册、激励的实验。