Do large language models actually reflect human value diversity?

The paper asks whether LLMs have values and answers yes, with a qualification that carries most of the weight. By projecting responses from 106 LLMs (150,000 queries per model) and 95,000 human survey profiles into "a shared sociological space," the authors "empirically confirm" an intrinsic value system. But the models "do not mirror human diversity, instead crystallizing into a highly concentrated, idealized value core." The discussion draws the practical consequence: heavily safety-constrained models are "fundamentally disqualified" as surrogates for diverse human populations, and should be read as "aligned tools rather than human surrogates."

The measurement design is what makes a claim about a population possible. The introduction rejects the idea that an LLM's values can be "a single static vector." Because models "swing" under repeated queries, the authors treat each model's value expression as "a full statistical distribution, not a fixed point," and shift evaluation "from a single individual to a macrolevel population." The 625 scenarios are grounded in the World Values Survey (Wave 7, roughly 95,000 valid respondents) and Schwartz Value Theory, so model and human distributions can be compared in the same human-interpretable space. The concentration finding is a statement about spread: many models, each sampled many times, occupy a small region that human respondents do not.

For mechanism, the discussion offers two claims. Value crystallization is described as "a deterministic byproduct of parameter compression," made stronger when "heavily constrained by safety protocols." And "dual-process dynamics of contextual framing and cognitive reasoning during inference" are said to explain the opening paradox, in which models swing under minor rewording yet rigidly ignore explicit instructions to correct ingrained biases.

Against the nearest notes, this extends Do large language models develop coherent value systems?. That note establishes that LLM preferences are structurally coherent; this paper adds that the coherent structure is narrow relative to a human reference population and is best described as a distribution. It also gives the swing side of a familiar problem: Are RLHF annotations actually measuring genuine human preferences? warns that single elicited answers can be artifacts, and the distributional framing is one response to that. The concentrated core also bears on Can AI systems preserve moral value conflicts instead of averaging them?, since a model that collapses toward one idealized center has, in effect, already aggregated.

The excerpt does not establish the evidence behind any of this. It gives no effect sizes, no description of where the core sits in the sociological space or what "idealized" means in content, and no test of the parameter-compression or dual-process explanations, both of which appear only as conclusions. It does not say whether the narrowness varies across model families or scales. What it supports at this strength is a caution for anyone using LLMs to simulate survey respondents or personas: the spread of outputs may understate human variety in a systematic way, however varied the prompts.