The industry keeps redefining AGI because the systems keep failing the old definition

Review

The definition of artificial general intelligence has been rewritten three times in ten years, and each rewrite arrived precisely when the existing systems were about to fail the previous test. That pattern, documented in peer-reviewed correspondence published in Nature in February 2026, is not an accident of scientific progress.

The original standard

The original standard, set in 2007, required AI to generalize reliably across genuinely novel environments with no task-specific tuning. None of the systems currently on the market meet that standard. The industry responded by changing the standard.

The redefinitions

The first shift redefined AGI in terms of benchmark performance — if a system could score well on enough tests, it was declared “general.” But benchmarks isolate narrow competencies and abstract away real-world context. Strong benchmark performance often provides little evidence of robustness under novelty, uncertainty, or shifting objectives.

The second shift redefined AGI in economic terms — if a system could perform tasks that previously required human labor, it was declared “general.” This definition has the advantage of appearing rigorous because it can be attached to data: automation studies, occupational surveys, task completion rates. But it measures economic substitution, not cognitive generality.

The third and most recent shift is the one the Nature letter targets most directly. A paper published in Nature in early 2026 by Chen and colleagues argued that success on behavioural tests, including variants of the Turing Test, constitutes evidence that AGI has already been achieved. The Turing Test measures whether a human evaluator, communicating via text, can distinguish a machine from a human over a sustained exchange. It was proposed by Alan Turing in 1950 as a thought experiment about the nature of machine intelligence, not as a certification protocol for general intelligence. Systems have been reported to pass versions of it by exploiting superficial conversational patterns and the limited technical knowledge of non-expert evaluators, not by exhibiting anything the 2007 standard would recognise as reliable generalisation.

The Nature letter

The Nature correspondence records the problem in one sentence: benchmark success is not real-world readiness. The core properties of general intelligence — flexibility, generality, and reliability — remain elusive. Current systems are brittle, sensitive to distributional shift, and prone to systematic failures that are invisible to anyone reading only the final output.

As of February 2026, the companies whose systems the Nature letter addresses have not responded to the substance of the critique. They have not needed to. The redefinition has already served its purpose: it allows the claim “AGI has been achieved” to be made in press releases and investor decks, regardless of whether the underlying systems meet the original standard.

Why this matters

This is propaganda because it uses the authority of science — benchmarks, peer-reviewed papers, the word “intelligence” — to make claims that the science does not support. The redefinition pattern is not a neutral evolution of terminology. It is a strategic retreat to definitions the current products can pass, presented as progress rather than concession. The audience — policymakers, investors, journalists, the public — hears “AGI” and understands the 2007 definition: a system that can think generally, like a human. The industry uses the 2026 definition: a system that can pass a text-based imitation game. The gap between those two definitions is where the hype lives.

Verdict: Misleading. AGI has not been achieved by the original standard. The industry redefined the term to claim it has. The redefinition is the propaganda.