The AGI Mirage: A Lakatosian Analysis of the Artificial General Intelligence Research Programme

The term artificial general intelligence lacks a consensus formal definition. This paper argues that the definitional vacuum is not an innocent gap in the literature but a symptom of a degenerating research programme. Applying Imre Lakatos’s Methodology of Scientific Research Programmes (MSRP), we evaluate three competing programmes in AGI discourse: the AGI-optimist programme (scaling computation produces general intelligence), the AGI-skeptic programme (current architectures have fundamental limitations), and the origination-derivation programme (a categorical boundary separates human cognitive origination from AI-system derivation). We trace the optimist programme’s sequence of definitional problemshifts from the Turing Test through behavioral capability thresholds, demonstrating a pattern of retroactive accommodation rather than novel prediction. Empirical evidence from creativity research, enterprise deployment failure rates, and adversarial generalization benchmarks corroborates predictions generated by the origination-derivation framework while disconfirming predictions of the optimist programme. We propose a grounded definition of AGI that requires demonstrated capacity for warranted novel response generation across domains under conditions controlling for memorization and interpolation, and argue that this definition, unlike current alternatives, satisfies Lakatos’s criterion of theoretical progressiveness by generating testable predictions with excess empirical content.