Impossibility Theorems Prove Cognitive Biases Are Inevitable in Sequential AI and Human Processing
A new academic paper titled 'Bias by Necessity' establishes that cognitive biases such as primacy effects and anchoring are mathematically inevitable consequences of sequential information processing. The authors prove three impossibility theorems demonstrating that these biases are architecturally necessary in autoregressive language models due to causal masking constraints. Specifically, primacy bias results from asymmetric attention accumulation, while anchoring emerges from sequential conditioning with provable information bounds. Exact debiasing is shown to require computationally prohibitive factorial-time operations. The theoretical framework was validated across twelve frontier large language models, achieving a high correlation coefficient. Furthermore, the study derived quantitative predictions tested in two pre-registered human experiments involving 464 participants. Results confirmed that anchor position modulates anchoring magnitude and working memory load amplifies primacy bias. These convergent findings from both artificial intelligence systems and human subjects reframe cognitive biases not as errors, but as resource-rational responses to the inherent limitations of sequential processing architectures.
Wire timeline
Impossibility Theorems Prove Cognitive Biases Are Inevitable in Sequential AI and Human Processing
A new academic paper titled 'Bias by Necessity' establishes that cognitive biases such as primacy effects and anchoring are mathematically inevitable consequences of sequential information processing. The authors prove three impossibility theorems demonstrating that these biases are architecturally necessary in autoregressive language models due to causal masking constraints. Specifically, primacy bias results from asymmetric attention accumulation, while anchoring emerges from sequential conditioning with provable information bounds. Exact debiasing is shown to require computationally prohibitive factorial-time operations. The theoretical framework was validated across twelve frontier large language models, achieving a high correlation coefficient. Furthermore, the study derived quantitative predictions tested in two pre-registered human experiments involving 464 participants. Results confirmed that anchor position modulates anchoring magnitude and working memory load amplifies primacy bias. These convergent findings from both artificial intelligence systems and human subjects reframe cognitive biases not as errors, but as resource-rational responses to the inherent limitations of sequential processing architectures.
cs.AI updates on arXiv.org