Do LLMs Share Human-Like Biases? Causal Reasoning Under Prior Knowledge, Irrelevant Context, and Varying Compute Budgets

AI summary

Overview

Research area: Artificial intelligence and cognitive science — specifically, comparing the causal reasoning of large language models (LLMs) against human causal judgment.

Technical level: Intermediate. The paper uses Bayesian network modeling, model-fitting metrics, and cognitive psychology paradigms, but explains each in accessible terms.

Scope: A benchmark of 20+ LLMs against a matched human baseline on 11 conditional-probability judgment tasks built on a single collider graph (C₁ → E ← C₂), examined across eight experimental conditions that cross prompting strategy with content manipulations.

What This Paper Is About

Human causal judgment is powerful but systematically biased — people explain away too little and let beliefs about one cause shift based on an unrelated alternative cause. Because LLMs are trained largely on human-generated text and are increasingly used in high-stakes domains, a natural question is whether they inherit these same human quirks.

The paper asks three questions: whether LLMs reproduce human-like response patterns and biases, whether their judgments can be compressed into a small interpretable causal Bayesian network, and whether those behaviors hold up when the task content is abstracted or buried under irrelevant text. Crucially, the tasks leave priors and causal strengths unspecified, so there is no single "correct" answer — the goal is to reveal each agent's reasoning strategy rather than to score accuracy.

Key Contributions

1. A broad human-vs-LLM causal reasoning comparison. The authors evaluate more than 20 LLMs on 11 collider-graph inference tasks with a matched human baseline, extending earlier work (Dettki et al., 2025) that tested only four LLMs.

2. A compressed, interpretable account of LLM judgment. They fit small causal Bayes nets (leaky noisy-OR parameterization from Cheng, 1997) to each agent and show that most LLMs' probability judgments are well captured by a single compact model, including out-of-sample generalization via leave-one-out cross-validation.

3. A systematic robustness analysis across eight conditions. Prompting strategy (direct vs. chain-of-thought) is crossed with two content manipulations: semantic abstraction (meaningful variable names replaced with random ten-character strings) and prompt overloading (appending irrelevant text).

4. Released tooling. An LLM-friendly version of the causal inference benchmark, including the human baseline, and a software package called causAIign that supports structure-matched custom prompts and content manipulations.

Main Findings

• No significant domain differences. Within each agent, probability-judgment distributions did not differ across the three cover stories (sociology, weather, economy), per Kruskal–Wallis tests (k = 3, df = 2) with Benjamini–Hochberg correction across agents: GPT-3.5-turbo p_FDR-BH = .146, all others p_FDR-BH ≥ .742. Domains were therefore pooled.

• Judgments are qualitatively sensible. Both humans and LLMs judge the effect as more probable the more causes are present, which the authors take as evidence that agents understand the causal mechanism.

• Chain-of-thought improves human alignment. For less aligned models in the direct prompting condition, CoT prompting improves alignment with human judgments up to a ceiling effect at 0.85 (Spearman ρ). Alignment was measured with 95% confidence intervals from 2000 bootstrap resamples.

• LLM judgments compress into small causal models. Most agents show small mean absolute error (MAE ∈ [0, 1], lower is better) when fit by a single causal Bayes net. CoT reduces MAE and dispersion relative to direct prompting, while overloaded prompts increase error, especially in the abstract setting. CoT also increases out-of-sample LOOCV R² and tightens dispersion, with the largest gains under overloaded prompts; several models approach or exceed the human benchmark.

• LLMs are stricter rule-followers than humans. Measured by Background-Adjusted Causal Strength (BACS = mean causal strength m̄ minus the background leak b, in [−1, 1]), most LLMs lie above the human benchmark, meaning they treat stated causes as more sufficient and attribute less to unmentioned latent factors. Humans appear more prone to accounting for latent factors and treating the stated causes as insufficient. Gemini-2.5-pro and Gemini-2.5-flash approach near-maximal rule-fidelity (BACS ≈ 1.0) and are largely insensitive to experimental conditions. A smaller set of agents shows lower BACS than humans in the RW17-Direct condition, but with CoT these approach or slightly exceed the human baseline.

• Most LLMs do not mirror human collider biases. Humans show weak explaining away (EA ≈ 0.1 in the human baseline) and frequent Markov violations. Most LLMs show explaining away levels well above the human baseline (EA ≫ 0) and cluster tightly around Markov compliance (MV ≈ 0). A minority of LLMs do exhibit human-like associative Markov violations. Under overload, many LLMs show reduced EA relative to baseline, with CoT largely restoring EA toward baseline levels; Markov violations occur more often under overload but still less than in humans.

• Robustness is strongly model-dependent, and CoT helps. Gemini-2.5-pro forms tight clusters across all eight experimental conditions (high BACS, high R², strong EA, MV ≈ 0), indicating near-invariance. Gemini-2.5-flash-lite shows clear prompting effects, with well-separated Direct and CoT clusters, and CoT drastically increasing content invariance. CoT often reduced cluster size — and thus increased robustness — for several other models as well.

Methodology in Plain English

The researchers started from a classic cognitive psychology experiment (Rehder and Waldmann, 2017; Experiment 1, Model-Only condition, N = 48 NYU undergraduates). Participants saw a causal story about two independent causes, C₁ and C₂, that both produce a common effect E — a "collider" structure. They were asked to estimate probabilities on a 0–100 scale for 11 different queries about one variable given observations of the others, embedded in three cover stories (sociology, weather, economy).

The key design choice is that the base rates and causal strengths are never specified. That means there is no single ground-truth number to grade against; instead, the task reveals the assumptions each agent brings.

For the LLMs, the authors converted the human task into a single text prompt containing a domain introduction, causal mechanism statements for both causes, the observed evidence, and a request for one numeric probability. They ran two prompting strategies — direct (a single number) and chain-of-thought ("think step by step" before the number) — with temperature set to 0.0 to minimize sampling variance, using API-provided reasoning-budget controls where a model exposed one.

They then crossed prompting with two content manipulations, producing 2 × 2 × 2 = 8 conditions: prior knowledge (original cover stories vs. abstract versions with meaningful names swapped for random ten-character strings), information load (original vs. overloaded prompts with appended irrelevant text, drawn from lorem ipsum or from distracting sentences taken from other RW17 domains), and prompting style.

To interpret the judgments, they fit a leaky noisy-OR causal Bayes net to each agent, where each active cause independently raises the effect's probability and a "leak" parameter b captures background activation from unmentioned factors. Fit quality was assessed with mean absolute error and with task-level leave-one-out cross-validation R² (fit on 10 of the 11 tasks, predict the held-out task). Two bias signatures were computed directly from raw judgments: explaining away (EA) and Markov violation (MV), plus the BACS index summarizing reliance on stated rules versus latent factors.

Why This Matters

Impact on research. The paper shows that Bayesian cognitive models can serve as compact, interpretable targets for diagnosing multi-billion-parameter neural networks — treating Bayesian and neural accounts of intelligence as complementary rather than competing. It also complicates the intuition that models trained on human text should inherit human cognitive biases; here, most do not.

Real-world applications:

• Legal decision support, where causal questions are common and human biases under time pressure are documented.
• Medical decision support, where LLMs are increasingly deployed and consistent causal analysis matters.
• Any high-stakes AI-assisted decision-making setting where human judgments are known to be brittle under limited attention, fatigue, and contextual framing.
• Human-LLM tandem workflows, where machines could complement specific human weaknesses — though the paper notes this as future work rather than a tested result.

Industry relevance. The finding that LLMs are often more rule-like than humans cuts both ways. Where human bias is unwanted, LLMs may be useful collaborators. Where real-world uncertainty is intrinsic and priors are genuinely unknown, tightly rule-following models risk breaking down — and robustness varies enormously by model, with newer and larger models (e.g., Gemini-2.5-pro) near-invariant while smaller or older ones are markedly more sensitive to how a prompt is phrased.

Future Directions

• Move beyond the collider graph. The authors flag their focus on the common-effect structure as a key limitation; whether these conclusions generalize to chains, common causes, and more complex graphs is open.

• Disambiguate the closed-world vs. open-world gap. The prompts do not explicitly state whether unmentioned causes should be ignored or treated as plausible background factors, which may differentially affect humans and models.

• Collect human data for all conditions. The authors note they have no human responses for the overloaded and abstract conditions, so comparisons there are against a baseline collected under different conditions.

• Test human-LLM tandems. The paper suggests exploring collaborative setups to understand how models might complement human reasoning, and more broadly questions how well laboratory-style tasks capture causal reasoning under real-world uncertainty.

Target Audience

Researchers in AI evaluation and cognitive science interested in whether LLMs reason causally or merely pattern-match; practitioners deploying LLMs in legal, medical, or other causal-inference-heavy domains; and anyone studying human-AI comparison, robustness to prompt perturbation, or the use of Bayesian models as interpretability tools for neural networks. Readers should be comfortable with basic probability, correlation, and model-fitting concepts.

Authors’ abstract

Large language models (LLMs) are increasingly used in domains where causal reasoning matters, yet it remains unclear whether their judgments reflect normative causal computation, human-like shortcuts, or brittle pattern matching. We benchmark 20+ LLMs against a matched human baseline on 11 causal judgment tasks formalized by a collider structure ($C_1 \rightarrow E \leftarrow C_2$). We find that a small interpretable model compresses LLMs' causal judgments well and that most LLMs exhibit more rule-like reasoning strategies than humans who seem to account for unmentioned latent factors in their probability judgments. Furthermore, most LLMs do not mirror the characteristic human collider biases of weak explaining away and Markov violations. We probe LLMs' causal judgment robustness under (i) semantic abstraction and (ii) prompt overloading (injecting irrelevant text), and find that chain-of-thought (CoT) increases robustness for many LLMs. Together, this divergence suggests LLMs can complement humans when known biases are undesirable, but their rule-like reasoning may break down when uncertainty is intrinsic - highlighting the need to characterize LLM reasoning strategies for safe, effective deployment.