Cognitive Bias Codex
The Cognitive Bias Codex is a systematic framework for mapping 188 known cognitive biases organized by the mental problems they attempt to solve. Initially compiled by Buster Benson and visualized as a radial dendrogram by John Manoogian III, the codex classifies biases into four categories based on the brain's evolved coping mechanisms: filtering information overload, filling in gaps when meaning is scarce, jumping to conclusions when speed is needed, and biasing memory to prioritize what matters.[^c1]
The study of cognitive biases has expanded beyond individual decision-making to encompass human-AI interaction, social media dynamics, and algorithmic fairness. Research has shown that both humans and LLMs exhibit a robust addition bias in problem-solving, with LLMs showing an even stronger preference for additive over subtractive solutions than humans.[^c8] Cognitive theories of bias now inform the regulation of AI fairness without requiring anthropomorphism of AI systems.[^c12] The interaction between cognitive biases and social media algorithms has been shown to create echo chambers and polarization — and formal impossibility theorems have demonstrated that certain biases such as primacy effects and anchoring are architecturally necessary consequences of sequential information processing,[^c14] challenging the notion that biases can ever be fully eliminated. Even without algorithmic personalization, exit dynamics driven by cognitive biases can produce strong ideological segregation in online communities.[^c11]
New biases continue to be characterized through both traditional and AI-assisted methods. Information undervaluation bias — the tendency to systematically undervalue information in complex situations where it is most beneficial — was discovered in 2026 via a novel framework that uses machine learning and large language models to detect previously unknown behavioral deviations from normative benchmarks.[^c25] This AI-assisted methodology has also been extended to auditing LLMs themselves, enabling fully automated detection of unverbalized biases in model decision-making across domains such as hiring, loan approval, and university admissions. Cross-disciplinary research has demonstrated that classic biases like the peak-end rule may emerge naturally from the mathematical constraints of credit assignment in distributed systems, linking AI research to human cognitive architecture.[^c23] At the same time, the foreign language effect — a proposed debiasing strategy — has failed to replicate for the illusion of causality,[^c27] underscoring the importance of replication in bias research. Research on egocentric bias has expanded to cover its role in collaboration breakdown and the spotlight effect in social judgment, while the curse of knowledge — the communication-specific manifestation of egocentric bias — has been documented in domains from teaching to software design. The Knobe effect, a moral bias in intentionality attribution, has been shown to emerge in finetuned LLMs and can be localized and mitigated through mechanistic interpretability; deliberative reasoning reduces the asymmetry in human participants, supporting a dual-process interpretation.
The risks of bias have also been examined at the societal level. Confirmation bias has been documented among expert government risk analysts, who exhibit the same tendency to seek confirmatory evidence as laypeople.[^c17] Forensic experts across fingerprint, DNA, and handwriting analysis are susceptible to contextual bias and bias cascade, affecting the integrity of evidence presented in court.[^c19] These findings show that cognitive biases affect not just laypeople but experts in consequential settings — judges, doctors, intelligence analysts, and forensic scientists.[^c18] At the same time, novel debiasing mechanisms have been identified: peer-to-peer information-sharing networks have been shown to reduce anchoring errors by 22%, suggesting that social structures can counteract biases that individuals cannot overcome alone.[^c16] AI-assisted decision-making has been shown to carry compound risks, as frequent AI users exhibit measurable declines in critical thinking and knowledge retention.[^c15]
Application domains have expanded into environmental and military intelligence contexts. A 2024 scoping review identified four categories of cognitive bias — bounded rationality, confirmation bias, self-enhancement, and ambiguity aversion — that arise in environmental and sustainability education, where affective-motivational challenges can trigger biased processing of sustainability information. Kahneman's theories of loss aversion and heuristic-driven System 1 thinking have been applied to understand why households underinvest in climate adaptation measures[^c21], while in military intelligence, the digital-intelligent era has introduced qualitatively new bias amplification mechanisms — algorithm bias, information cascades, and data overload — that did not exist in pre-digital environments.[^c22] AI sycophancy in military decision-support systems has been identified as a distinct risk factor that aggravates confirmation bias and anchoring, requiring both technical mitigation and user training. Time-related biases affecting intertemporal decisions, including present bias, projection bias, and the planning fallacy, have been systematically reviewed, with tailored mitigation strategies proposed for improving long-term decision-making in health, finance, and policy.[^c20] In supply chain management, LLM-based simulations have demonstrated that myopic and self-interested decision-making biases replicate in AI agents and can be mitigated through information sharing. In human-AI interaction, new approaches to mitigating cognitive bias in reinforcement learning from human feedback (RLHF) have demonstrated that dynamically adjusting the rationality parameter during reward learning can reduce an AI system's sensitivity to biased human judgments.[^c24] The Bias Amplification Model has provided a theoretical framework for how human cognitive biases enter, propagate, and become embedded in agentic AI systems through injection, propagation, and crystallisation layers.
Research on biases in large language models has revealed that models are susceptible to cognitive heuristics in security-critical tasks: in LLM-based code vulnerability detection, framing, anchoring, and halo effects influenced verdicts at rates of 18–33%, and a proof-of-concept cognitive attack suppressed up to 97% of detected vulnerabilities.[^c26] These findings indicate that cognitive susceptibility is a consistent and exploitable property of LLMs. A qualitatively different form of algorithmic bias — stereotype hallucination — has been identified in LLMs, where models generate strong stereotyped predictions that have no basis in actual data, persisting even after fine-tuning on participant responses.
Understanding the cognitive ecosystem in which biases operate — individual, social, media, and AI layers — has become a central concern for maintaining human agency in digitally mediated environments.