Nine Fallacies of Professional AI Usage: An Evidence-Based Synthesis

Abstract

Organizations are adopting generative AI and AI decision support in professional work while evidence on their effects is still accumulating. Empirical studies document substantial productivity gains in some settings and losses in others, and practitioners carry assumptions formed during early use into everyday professional work. Research on the resulting problems is spread across human factors, human-computer interaction, fairness research, empirical economics and law, and no consolidated framework for professional use was found. This paper proposes such a framework. Its format is inspired by the Fallacies of Distributed Computing, a short list of false assumptions well known in software engineering; its content is specific to professional AI usage. Through an integrative review of peer-reviewed studies, randomized and preregistered trials, court decisions, provider documentation and regulation, nine fallacies are derived: (1) the output is correct; (2) capability is uniform across tasks; (3) the human reviewer is a reliable safeguard; (4) perceived productivity equals actual productivity; (5) AI removes work; (6) accountability moves to the tool or vendor; (7) skills and system understanding stay intact when work is delegated; (8) results transfer from benchmarks and pilots; and (9) the model is available. For each fallacy, the paper explains why the assumption persists, presents the evidence against it, states the conditions under which it does hold, and grades the strength of the evidence. It then analyzes how fallacies can compound, with particular attention to agent-built software systems that may become difficult to maintain when AI access is lost. The paper closes with implications for practice, a research agenda and a self-assessment instrument.