Beyond the AGI Claim: What Would Actually Count as Evidence for Artificial General Intelligence?

Evidence of broad AI performance can support different claims about intelligence. This article asks when such evidence justifies an artificial general intelligence (AGI) attribution and which conclusions depend on the chosen account of intelligence. Through conceptual analysis, we reconstruct three published claims: Bubeck and colleagues' early GPT-4 interpretation, Hendrycks and colleagues' cognitive-parity criterion, and Zhou and colleagues' prediction of language-model response success. Three checks address the account's adequacy, evidential support and preservation of evaluation conditions and claim scope. The analyses distinguish immediate correction from broader reuse of task prerequisites, support for a storage score from its effect on a classification endpoint, and assessor prediction from learning by the assessed model. Comparing one evidence package across three accounts identifies different evidential requirements without establishing different AGI verdicts. The central argument is that materially changed requirements call for reassessment, while a disputed premise may affect one conclusion without defeating a separately supported finding. The contribution is a critical methodological synthesis with traceable case analyses. Purposive selection and selective source inspection limit the judgments; the proposed method and its practical benefits have not been empirically validated.