Human Cognitive Biases in Large Language Models

77 concepts47 papers1 proposal

When a Large Language Model (LLM) makes an error, it is tempting to view the failure as random noise—a statistical hallucination born of imperfect optimization. However, a growing body of research demonstrates that these errors are often highly structured, predictable, and distinctly human. Rather than failing randomly, LLMs systematically replicate human cognitive biases, inheriting the exact psychological shortcuts and irrationalities that characterize human decision-making.

Understanding why these biases emerge requires looking closely at the training objective of modern language models and how it interacts with the architecture of human thought.

The Statistical Inheritance of Heuristics

At their core, LLMs are trained via next-token prediction on massive corpora of human-generated text. This text is not merely a collection of factual statements; it is a footprint of human cognitive processes. When humans write, their language is deeply influenced by cognitive heuristics—mental shortcuts that allow for fast, efficient decision-making but frequently lead to deviations from logic and probability.

By minimizing cross-entropy loss on human text, an LLM does not just learn the objective structure of language or logic; it learns to mirror the behavioral patterns embedded within that language. As Eric Schulz et al. demonstrate in Using cognitive psychology to understand GPT-3, the statistical regularities of human training data are entirely sufficient to replicate complex psychological irrationalities, such as the framing effect (changing choices based on how a scenario is worded) and the overweighting of small probabilities.

Consider the representativeness heuristic, where individuals judge the probability of an event based on how closely it matches a mental prototype, ignoring statistical base rates. In Will the Real Linda Please Stand up...to Large Language Models? Examining the Representativeness Heuristic in LLMs, Pengda Wang et al. show that LLMs possess the underlying statistical knowledge required to solve these problems correctly, yet they consistently fail when the problem is presented within a narrative context. The model prioritizes semantic similarity over mathematical probability because, in human text, descriptive alignment is a powerful driver of association.

Similarly, when tasked with generating random sequences, models fall prey to the gambler's fallacy. As explored by Eric J Bigelow et al. in In-Context Learning Dynamics with Random Binary Sequences, instead of producing mathematically uniform randomness, LLMs generate sequences that artificially avoid long streaks. They do this because they are mimicking what humans perceive as random, rather than what is mathematically random.

LLMs as System 1 Thinkers

To understand the mechanism behind this replication, researchers have turned to the dual-process theory of cognitive psychology, which divides human thought into two modes:

  • System 1: Fast, automatic, associative, and highly susceptible to biases.
  • System 2: Slow, deliberative, logical, and computationally expensive.

Standard autoregressive token generation closely maps to System 1. The model generates the next token using immediate associative patterns across its weights without an inherent mechanism for paused deliberation. Gili Lior explicitly tests this parallel in Comparing Humans and Models on a Similar Scale: Towards Cognitive Gender Bias Evaluation in Coreference Resolution, demonstrating that when humans are placed under high cognitive load or forced to make rapid decisions, their bias profiles begin to look nearly identical to those of language models.

This alignment explains why common prompting techniques can yield counterintuitive results. While Chain-of-Thought (CoT) prompting is often used to simulate System 2 deliberation, it can backfire if the underlying human data contains biases associated with overthinking. In Mind Your Step (by Step): Chain-of-Thought can Reduce Performance on Tasks where Thinking Makes Humans Worse, Ryan Liu et al. show that forcing a model to reason step-by-step significantly degrades performance on specific task archetypes where human intuition outperforms explicit analysis. The model inherits not just human shortcuts, but also human cognitive limitations.

Systematic Irrationality in High-Stakes Contexts

Because these biases are structural rather than random, they persist even as models scale and become more capable in formal logic. This systematic irrationality poses a challenge when deploying LLMs as autonomous agents in economic, medical, or legal environments.

[Human Training Corpora] ──(Next-Token Prediction)──> [LLM Internalizes Associations]
                                                               │
                                               ┌───────────────┴───────────────┐
                                               ▼                               ▼
                                      [System 1 Thinking]             [Predictable Failures]
                                   Fast, associative decoding        Anchoring, Sunk Cost, Framing

When evaluated through economic utility frameworks, LLMs frequently violate basic axioms of rationality. In STEER: Assessing the Economic Rationality of Large Language Models, Narun K. Raman et al. observe that models consistently exhibit the sunk cost fallacy and irrational risk aversion. These behaviors are not driven by mathematical optimization of utility, but by the linguistic patterns of risk and loss aversion present in human discourse, a finding echoed by Bowen Jiang et al. in their survey on agent rationality, Towards Rationality in Language and Multimodal Agents: A Survey.

Furthermore, these models display severe anchoring and recency biases. In multi-agent or multi-step decision scenarios, models tend to fixate heavily on initial information or the most recently presented options while ignoring the broader context. Jessica Maria Echterhoff et al., in Cognitive Bias in Decision-Making with LLMs, show that option order and question framing can radically shift a model's judgment, leading to severe inconsistencies in high-stakes decision-making.

Exploiting the Cognitive Mirror

The predictability of these biases means they can be systematically exploited. Because LLMs share a cognitive vulnerability profile with humans, adversarial frameworks can bypass safety filters without needing access to model weights or token probabilities.

In Exploiting Synergistic Cognitive Biases to Bypass Safety in LLMs, Xikang Yang et al. demonstrate that by combining multiple psychological triggers—such as authority bias and the hot-hand fallacy—adversarial prompts can reliably trick models into bypassing safety guardrails. Because these vulnerabilities stem from the underlying training data rather than architectural quirks, these "cognitive attacks" transfer successfully across entirely different model families, both open-source and proprietary.

Ultimately, treating LLM failures as human-like cognitive biases provides a more accurate diagnostic framework than treating them as pure noise. By recognizing that the model is acting as a statistical mirror of human psychology, researchers can better predict when a model will fail, why it chooses a specific incorrect path, and how to design more robust intervention strategies.

Key Foundational Work

Capturing Failures of Large Language Models via Human Cognitive Biases (2022) — contributed 1 concept elements to this direction.
Using cognitive psychology to understand GPT-3 (2022) — contributed 2 concept elements to this direction.
Computers and Thought Award : Challenges of Massive Parallelism (1993) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_llm_susceptibility_to_cognitive_biases_e688890e,
  title = {LLM susceptibility to cognitive biases},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/llm-susceptibility-to-cognitive-biases/rdr_e688890eff8a4a529ddd1a226ec59e20},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_llm_susceptibility_to_cognitive_biases_e688890e_v1,
  title = {LLM susceptibility to cognitive biases},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/llm-susceptibility-to-cognitive-biases/rdr_e688890eff8a4a529ddd1a226ec59e20/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake