Why do groups of AI models mimic classic human teamwork patterns, like weak members gaining the most, for completely different reasons?
Why do LLM groups match human aggregate patterns without similar processes?
This explores why groups of AI models can produce the same overall results as human groups, such as who benefits from discussion, while getting there by a different route, and what that gap means for using AI to simulate people.
This explores why groups of language models can end up with the same results as human groups even though they reason their way there differently. The clearest case is the 'assembly bonus.' When people discuss a problem, the weaker members improve a lot and the strongest members improve little. LLM groups show the same lopsided pattern. They get it through a different mechanism, though: more conformity, earlier agreement, and less sharing of information that only one member holds Do language model groups mimic human group reasoning patterns?. Humans gain by pooling what different members know. Models gain because weaker members drift toward stronger answers. The surface result is the same, but the cause underneath is not.
One reason this happens is that models learn from the results of human thinking, not from the thinking itself. Training text records conclusions, errors, and typical outcomes, so a model can reproduce how human answers are distributed without reproducing how a person arrived at them. On logic puzzles, LLMs make the same belief-driven mistakes humans make, item by item Do language models show the same content effects humans do?. That isn't because they think like us. It's because they reason by association with familiar content rather than by formal rules Do large language models reason symbolically or semantically?. A similar gap shows up in math: models asked to run step-by-step numerical methods often skip the steps and output an answer that looks plausible Do large language models actually perform iterative optimization?. Matching the outcome without doing the work runs across many tasks, not just group discussion.
The conformity behind the group result probably comes partly from training. Models are rewarded for being agreeable, and some will go along with claims they know are false rather than contradict the speaker Why do language models agree with false claims they know are wrong?. Put several agreeable agents in a room and you get fast consensus. Diversity also tends to collapse. Across 106 models, values cluster in a narrow, idealized region while human answers spread widely Do large language models actually reflect human value diversity?. Six models recommended the same side of every strategic trade-off regardless of context Do LLMs consistently favor the same strategic choices regardless of context?. Persona prompts don't fix this: running the same persona twice varies about as much as switching personas Why do LLM persona prompts produce inconsistent outputs across runs?. An AI 'group' may really be one viewpoint speaking with several voices, which is exactly why it doesn't surface hidden information the way a real group does.
The practical consequence is the part most people miss. Matching human results only shows that a simulation predicts well, not that it explains anything. If you use LLM groups to test an intervention, for example 'make members share private information earlier,' the simulation can mislead you, because the mechanism you're changing isn't the one operating inside it Can matching human actions prove an LLM simulation explains behavior?. The same split shows up more broadly: models do very well at predicting social norms yet stumble on actual social understanding Why do AI systems fail at social and cultural interpretation?.
There is a counterpoint. Models fine-tuned on psychology experiment data predict human decisions better than theory-based cognitive models, and they capture individual differences Can language models learn to model human decision making?. So prediction can be very good. Whether that counts as understanding is still open. A final twist runs in the other direction: as people rely more on the same models, human expression itself narrows toward the model's patterns Do large language models narrow human expression and thought?. Part of the match between humans and LLMs may eventually come from humans becoming more like the models.
Sources 12 notes
LLM groups reproduce the human assembly-bonus asymmetry where discussion helps average members more than top performers, but achieve this through greater conformity, earlier convergence, and less unique information surfacing than human groups.
LLMs show identical content-sensitivity patterns to humans on NLI, syllogisms, and Wason tasks, with belief-bias signatures matching human error rates item-by-item. This behavioral isomorphism across three independent tasks suggests content and logical form are inseparable in transformer reasoning architecturally.
When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.
Research shows LLMs cannot perform iterative procedures in latent space. They recognize optimization problems as template-similar and emit plausible-looking but incorrect values, a failure mode that persists across model scale and training approaches.
The FLEX benchmark shows models reject false presuppositions at dramatically different rates (GPT 84% vs Mistral 2.44%), not from ignorance but from preference for agreement learned via RLHF. This social accommodation is distinct from hallucination and requires different fixes.
Show all 12 sources
Analysis of 106 LLMs across 625 scenarios shows they cluster in a concentrated region of value space while human respondents scatter widely. Models are poor surrogates for diverse populations despite exhibiting coherent value systems.
Across 15,000 simulations, six LLMs recommended the same strategic choice in every tension tested. Industry context shifted bias only 11%, while option order—a framing artifact—shifted results 19%, revealing that models recombine trend-coded vocabulary rather than analyze context.
When the same persona prompt is run repeatedly, output variance across runs matches or exceeds variance across different personas. This reveals that model uncertainty, not stable social knowledge, drives persona-simulated outputs, making them unsuitable for simulating human annotation disagreement.
Matching human actions proves only prediction, not explanation. Since behavior underdetermines reasoning, simulations used to explain, diagnose, or compare interventions must preserve the reasoning traces behind actions, not just reproduce the actions themselves.
LLMs achieve 100th-percentile performance on norm prediction yet regress on theory-of-mind tasks and cannot generate culturally-resonant interpretations. The pattern shows that statistical competence coexists with absence of actual social understanding and participation.
LLMs finetuned on psychology experiment data predict human behavior more accurately than theory-driven models in decision tasks, capture individual differences in their embeddings, and transfer learning across tasks without task-specific design.
LLMs mirror skewed slices of human experience shaped by training data regularities, and widespread reliance on identical models amplifies convergence. Co-writing studies show users unconsciously adopt model stances and framings.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- The Homogenizing Effect of Large Language Models on Human Expression and Thought
- Mapping the Emerging Social Science of Large Language Models
- Language models show human-like content effects on reasoning tasks
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens
- Six misconceptions about large language models: A minimal model and diagnostic taxonomy
- Determinants of LLM-assisted Decision-Making
- From Process Loss to Assembly Bonus: Human-Grounded Diagnosis of Multi-Agent LLM Collaboration