Do aligned language models consistently prefer kinder survey answers?

The paper tests the assumption behind LLM-as-respondent research, that conditioning a model on who a person is "yields answers resembling those of real people from that group." It reports a "small but consistent tendency for aligned LLMs to lean toward the kinder, safer, more socially approved answer on value-laden survey questions" and names it benevolence bias. The evidence in the excerpt spans 18 models, four datasets (ANES, GSS, WVS and a cross-cultural prospect-theory replication) and six psychological categories. 83 of 108 BTB cells and 75 of 108 BWR cells are positive. The shift concentrates on social desirability (mean BWR 0.565) and harm aversion (0.569), while emotional softening is "close to null" (0.494).

The paper's claim is that this is a model property rather than a dataset artifact. Model rankings agree across the three general survey datasets (Spearman's ρ between 0.54 and 0.63). The abstract adds that the bias points the same way across models, grows with model size and traces to the post-training stage. Prompt language and framing "change its size but never its direction." The discussion calls it "not a thin layer of style on top of an otherwise faithful value representation, but a prior the model brings to every persona it adopts and every survey it answers." A "malicious persona" stress test shows the limit is one-sided: aligned models struggle to play people who are less kind, less prosocial or more harm-tolerant than average. The stated consequence is "not only a shifted average, but a narrowed range of people the model can imitate." The framing that makes this plausible is that the same models are built to be "helpful" and "harmless," a design goal that competes with faithful simulation of people who are neither.

Against the nearest notes, this is a contrast on where over-positivity comes from. Why do LLMs give unrealistic survey responses? argues that skew and over-positivity dissolve when the elicitation channel changes. This paper finds that framing moves the magnitude but never the sign, and it locates the source in post-training. The two may describe different pathologies, numeric skew on purchase intent versus a value-laden shift, but the excerpt does not test a text-and-similarity channel, so it limits how far "artifact not model limit" can be stretched to value-laden questions. The post-training locus is consistent with Why do preference models favor surface features over substance?, where sycophancy diverges from human preferences, and with Does preference optimization harm conversational understanding?. Here the tax falls on simulation fidelity rather than on dialogue. It also raises a question for Can AI agents learn people better from interviews than surveys?: richer persona input raises fidelity, yet a prior that every persona inherits may be one that richer input does not remove.

The excerpt does not define BTB or BWR, report effect sizes beyond the three mean BWR values, name the models, or show how the size trend and the post-training attribution were established. It stops before saying where the bias "does and does not appear" beyond the emotional-softening null. The paper's title promises a correction, and the excerpt describes none. What follows at this strength is narrow: on value-laden questions, an aligned model's simulated group answers should be read as possibly shifted benevolent and possibly missing the less prosocial tail, and the paper's own view is that naming and measuring the bias is what lets researchers plan around it.