Human and Artificial Biases and Debiasing Techniques in Software Estimation
Software estimation is central to planning and executing development projects, yet both human estimators and AI-based systems may have biases that undermine accuracy. Estimation can be direct, an intuitive, analogy-driven process largely governed by Kahneman’s error-prone System 1, or indirect, relying on formalized cost models. In either case it should be a statistically grounded in order to allocate resources.
This presentation maps the biases affecting estimation, addressing both the human side and artificial intelligence perspectives. The session then reviews effective debiasing methods. It also argues that extreme precision in sizing is not always needed and favors lighter, less intrusive methods such as Simple Function Points (SFP). An exploratory experiment complements the theory by comparing human estimates with those produced by Large Language Models on a real-life case. It tests whether documented human biases are replicated or transformed, and whether LLMs act as debiasing tools or as new sources of distortion. Finally, the contribution proposes a conscious hybrid approach that combines the best human practices with AI’s predictive power to improve the reliability and fairness of estimates.
