Impossibility Theorems Prove Cognitive Biases Inevitable in AI and Humans
A new paper on arXiv proves that cognitive biases like primacy effects and anchoring are mathematically inevitable in autoregressive language models due to causal masking constraints. The impossibility theorems show primacy bias arises from asymmetric attention accumulation, anchoring emerges from sequential conditioning with provable information bounds, and exact debiasing via permutation marginalization requires factorial-time computation. Monte Carlo approximation is feasible at constant per-tolerance overhead. The bounds were validated across 12 frontier LLMs (R² = 0.89; ΔBIC = 16.6 vs. next-best alternative). Two pre-registered human experiments (N = 464) confirm anchor position modulates anchoring magnitude (d = 0.52). The study suggests that certain biases are architecturally necessary in both AI and human sequential processing.
Key facts
- Three impossibility theorems prove primacy effects, anchoring, and order-dependence are architecturally necessary in autoregressive language models.
- Primacy bias arises from asymmetric attention accumulation.
- Anchoring emerges from sequential conditioning with provable information bounds.
- Exact debiasing by permutation marginalization requires factorial-time computation.
- Monte Carlo approximation is feasible at constant per-tolerance overhead.
- Validation across 12 frontier LLMs achieved R² = 0.89; ΔBIC = 16.6 vs. next-best alternative.
- Two pre-registered human experiments (N = 464) confirm anchor position modulates anchoring magnitude (d = 0.52).
- Paper published on arXiv under ID 2605.08716.
Entities
Institutions
- arXiv