ENGW

Human and Artificial Biases and Debiasing in Software Estimation: Critical Reflections

Prof. Roberto Meli
D.P.O. – Data Processing Organization S.r.l., Italy

Abstract: Software estimation is a crucial activity for the planning and execution of software development projects. Estimation can be direct, a fundamentally intuitive and analogy-based process (where the transformation of inputs into outputs largely occurs at an unconscious level, within Kahneman's System 1, which is prone to error), or indirect, based on formalized cost models. However, both human estimators and Artificial Intelligence (AI) systems employed for estimation are intrinsically susceptible to cognitive and algorithmic biases that can significantly compromise the accuracy of predictions. It is essential to acknowledge that estimation is not a "scientific conjecture" but a statistically based social contract that provides an agreed-upon framework for resource allocation. Furthermore, the Heisenberg uncertainty principle is highlighted as applied to software measurement: the act of measuring a software attribute tends to influence it, and estimation will affect the actual outcome (the project will tend to adapt to the allocated timeframe), as well as influencing the behavior of other less-focused variables (De Marco's law of cost migration). This plenary presentation at the ENGW 2026 conference aims to deeply explore the landscape of biases (human and artificial) that influence software estimation processes. On the human side, classic biases such as anchoring (the use of an initial reference point for subsequent approximations), overconfidence, and confirmation bias (or cognitive dissonance reduction, where the brain avoids or minimizes data that conflicts with a decision or previously held belief—a phenomenon that primarily affects experts) will be analyzed, all of which distort the judgment of human estimators. Phenomena such as framing (the significance of how a choice is presented, which can influence the response), pitfalls generated by personal experience (which artificially tends to reduce the unknown to the known), and the self-fulfilling prophecy (where expectations filter experience to conform to those expectations) will also be discussed. The estimation process is further influenced by organizational factors such as psychological pressure and the necessity to negotiate resources. On the artificial side, the focus will be on biases induced by training data (e.g., sampling bias, label bias), which cause Machine Learning (ML) models to reproduce and amplify systematic inaccuracies. The session will provide critical insights into the most effective debiasing methodologies. Human strategies to be discussed include bias awareness training, the use of group estimation techniques (Delphi or Planning Poker), cross-validation, the Six Thinking Hats technique to structure thought and counteract group thinking (for example, by nominating a "devil's advocate" or having the most expert members speak last), and the AHP technique. Furthermore, it will be highlighted that extreme precision in measuring software size is not always necessary to avoid being intrusive on production processes, leading to the use of simpler and faster methods such as the Simple Function Point (SFP). Complementing the theoretical framework, an exploratory experiment will be presented, comparing estimates produced by human estimators with those generated by Large Language Models (LLMs) applied to the same historical cases. The objective is to empirically verify if and to what extent systematic biases documented in human estimates, particularly optimism and framing effects, are replicated or transformed by the language models, raising the question of whether the LLM constitutes a debiasing tool or a new source of distortion. Finally, this contribution aims to outline a conscious, hybrid approach that integrates the best human practices with the predictive capabilities of AI, actively mitigating biases to enhance the reliability and fairness of software development estimates. The theoretical framework will be illustrated through examples and anecdotes drawn from professional practice.

Brief Biography of the Speaker: Roberto Meli graduated in Computer Science in 1984. In 1996, he became CEO of DPO Srl (Italy). Over the past 40 years he has developed focused competences in project management and software measurement and has written more than 100 papers for technical magazines and international conferences. He is a consultant and lecturer on project management and software measurement for many major Italian companies and public organizations. He developed the Early & Quick (E&Q) Function Point Analysis method and the Simple Function Point (SFP) method. Certified Function Point Specialist (CFPS) from 1996 to 2004. Roberto has held leadership roles over the last 25 years including: IFPUG Board of Directors (Current role), Director of Functional Measurement Standards Committee, Chairperson of the GUFPI-ISMA (Gruppo Utenti Function Points Italia - Italian Software Measurement Association) board of directors, Chairperson of the COSMIC Measurement Practices Committee, Conference Chairperson of SMEF (Software Measurement European Forum), and member in the board of International Software Benchmarking Standards Group (ISBSG).