Cognitive Biases in LLM-Assisted Software Development

TL;DR

The paper investigates cognitive biases in developer–LLM collaboration, revealing that LLM-assisted software development introduces novel bias patterns and amplifies some traditional biases. It employs a mixed-methods approach, analyzing 2,013 actions from 14 participants and validating findings through a 22-participant survey, resulting in a taxonomy of 90 biases across 15 categories. The study reports that 53.7% of LLM-related actions are biased and that 29.5% are reversals, underscoring the need for deliberate mitigation and better tool design. It provides concrete mitigation strategies—ranging from before prompting to after applying LLM suggestions—and actionable design recommendations for LLM developers to slow down decision-making, increase transparency, and offer context-aware, verifiable alternatives to support high-quality human–AI software development.

Abstract

The widespread adoption of Large Language Models (LLMs) in software development is transforming programming from a solution-generative to a solution-evaluative activity. This shift opens a pathway for new cognitive challenges that amplify existing decision-making biases or create entirely novel ones. One such type of challenge stems from cognitive biases, which are thinking patterns that lead people away from logical reasoning and result in sub-optimal decisions. How do cognitive biases manifest and impact decision-making in emerging AI-collaborative development? This paper presents the first comprehensive study of cognitive biases in LLM-assisted development. We employ a mixed-methods approach, combining observational studies with 14 student and professional developers, followed by surveys with 22 additional developers. We qualitatively compare categories of biases affecting developers against the traditional non-LLM workflows. Our findings suggest that LLM-related actions are more likely to be associated with novel biases. Through a systematic analysis of 90 cognitive biases specific to developer-LLM interactions, we develop a taxonomy of 15 bias categories validated by cognitive psychologists. We found that 48.8% of total programmer actions are biased, and developer-LLM interactions account for 56.4% of these biased actions. We discuss how these bias categories manifest, present tools and practices for developers, and recommendations for LLM tool builders to help mitigate cognitive biases in human-AI programming.

Figures (6)

• Figure 1: Methodology Overview Diagram
• Figure 2: Distribution of Presence of (a) Bias and Reversal Actions, (b) Bias and LLM actions, and (c) LLM and Reversal actions. Circle size reflects action counts, and each cell shows actions matching that pair. Marginal totals appear along the bottom and right, with overall totals in the lower-right corner.
• Figure 3: Distribution of Presence of (a) Bias and Reversal Actions, (b) Bias and LLM actions, and (c) LLM and Reversal actions. The size of the circles is proportional to the number of actions. Each cell presents the actions matching these dimensions. Totals are shown along the bottom and right edges, with overall totals shown in the lower right-hand corners.
• Figure 4: Frequency of biased LLM actions and reversals by category.
• Figure 5: (a) The five highest perceived frequencies of bias. Bars are ordered by the combined proportion of "Often" and "Always", with dark + light blue segments indicating high-frequency, red indicating low-frequency responses ("Never", "Rarely"); center grey segment represents neutral "Sometimes". (b) The five highest perceived impactful biases. Bars are sorted based on the sum of the "Somewhat negative" and "Very negative" frequencies with red bars, while green bars signify positive impacts, with bars in the center reflecting "No effect" responses.