Bias by Necessity: Impossibility Theorems for Sequential Processing with Convergent AI and Human Validation

Researchers prove that primacy effects, anchoring, and order-dependence are mathematically inevitable in autoregressive language models due to causal masking constraints. The findings are validated across 12 frontier LLMs and confirmed through human experiments, suggesting cognitive biases represent resource-rational responses to sequential processing rather than design flaws.

This research bridges theoretical computer science and cognitive psychology by demonstrating that certain human biases have mathematical equivalents in AI systems. The impossibility theorems establish that these biases aren't artifacts of poor training but rather architectural necessities arising from how sequential models process information. The causal masking constraint—which prevents models from attending to future tokens—creates asymmetric attention patterns that inevitably produce primacy effects and anchoring behaviors.

The convergence between AI behavior and human cognition is striking. The study validates theoretical predictions across 12 state-of-the-art language models with high correlation (R² = 0.89), then replicates findings in human subjects through pre-registered experiments. Study 1 demonstrates anchor position significantly modulates anchoring magnitude, while Study 2 shows working memory load amplifies primacy bias, with individual WM capacity predicting bias reduction. These parallel mechanisms suggest both systems face similar computational constraints.

For AI development, this research reframes debiasing efforts. Rather than pursuing elimination, which requires factorial-time computation, practitioners should acknowledge these biases as inherent to sequential architecture and implement efficient Monte Carlo approximations when necessary. This has implications for deploying LLMs in high-stakes domains where bias matters—legal systems, financial advisory, medical diagnosis—by setting realistic expectations about model behavior. The findings also advance fundamental AI safety research by identifying which biases are mathematically unavoidable versus those requiring explicit mitigation. Understanding this distinction enables more targeted debiasing strategies focused on problematic behaviors rather than impossible elimination.

• →Primacy effects and anchoring biases are mathematically inevitable in autoregressive language models due to causal masking constraints, not design failures.
• →Exact debiasing through permutation marginalization requires factorial-time computation, making practical approximation methods necessary.
• →Human cognitive biases show identical patterns to AI systems across 12 frontier LLMs (R² = 0.89), suggesting shared processing constraints.
• →Working memory load amplifies primacy bias in humans with correlation to individual capacity differences, mirroring sequential processing limitations.
• →This framework enables targeted debiasing strategies focused on tractable interventions rather than mathematically impossible bias elimination.