Research map / Probabilistic and scientific modeling / Causal discovery

Causal Bias Detection in Generative Artifical Intelligence

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Causal approaches to fairness

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a causal framework (S-SFM with a selection node for real-world vs generative-model mechanisms) that decomposes how LLM-encoded disparities differ from real-world causal effects, and applies it to race and sex bias across 10 open-weight LLMs on NSDUH, BRFSS and ACS data.

Problem

Causal fairness methods assume a decision-maker fits one predictor while inheriting real-world mechanisms for other covariates, but generative models sample arbitrary conditionals and carry their own beliefs about all causal mechanisms, so existing bias decompositions do not apply.

Approach

Introduces an S-node selecting real-world or generative-model mechanisms for each covariate group, defines S-modified direct, indirect and spurious effects, and decomposes the change in total disparity into contributions from stepwise mechanism replacement (f_Y, then f_W, then f_{X,Z}); the standard ML setting is recovered as a special case.

Key results

• Across 10 open-weight models and 3 datasets, reversal occurs in 7-44% of disadvantage-direction effects and amplification in 15-44%; no model consistently dampens bias.
• Gemma 3 27B reverses the NSDUH direct effect from -3.8% to +4.3% (±2.9%) under replacement of f_Y and f_W.
• Qwen 3.5 27B reverses BRFSS direct effect to -1.8% and indirect effect to -6.3% under full replacement, versus real-world +5.7% and +2.7%.
• A permutation test finds no evidence that same-family models have more similar bias signatures (mean within-family L1 0.80 vs null 0.76, p=0.62).

Surprising findings

• Bias similarity is not highest within model families; the closest pair is Llama 3 8B and Ministral 3 8B across developer families.
• Gemma 3 27B encodes the belief that minorities use marijuana more, contrary to the real NSDUH direct effect.

Limitations

• Cannot handle generative models inferring in the opposite causal order (e.g., sex from salary).
• Requires the ability to query arbitrary conditionals, which may fail in vision models.
• Requires human-elicited S-SFM causal assumptions; conclusions depend on the graph; only open-weight models studied.

Experimental setup

Datasets: NSDUH 2023 BRFSS 2023 ACS 2023

Benchmarks: NSDUH 2023 BRFSS 2023 ACS 2023

Metrics: x-specific direct effect indirect effect spurious effect bias signature L1 distance annotator accuracy

Baselines: Standard ML causal fairness setting (Cor. 1)

Builds on: standard fairness model causal mediation analysis (path-specific effects) counterfactual fairness one-step debiasing

Closest papers by meaning

• Fairness-Accuracy Trade-Offs: A Causal Perspective (AAAI 2025 · similarity 0.92)
• FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.90)
• Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach (ICLR 2024 · similarity 0.89)
• Towards counterfactual fairness through auxiliary variables (ICLR 2025 · similarity 0.89)
• Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.88)
• Mind the Gap: A Causal Perspective on Bias Amplification in Prediction & Decision-Making (NeurIPS 2024 · similarity 0.88)