The most quoted formula is attributed to Larry Tesler, around 1970: "AI is whatever hasn't been done yet." By Tesler's own account, what he said was slightly different: intelligence is whatever machines haven't done yet. M. Mitchell (2019) records John McCarthy's lament: "As soon as it works, no one calls it AI anymore" (2019, ch. 3).
The textbook case is chess. In 1950 Claude Shannon predicted that a machine able to outplay humans would force us either to admit mechanized thinking or to restrict our concept of thinking. After Deep Blue the second happened: playing chess stopped looking like a test of general intelligence (2019, ch. 9). The same happened with character recognition, spam filters, machine translation and, for some, fluent conversation.
There are two readings. The first is defensive: each achievement is disqualified because it threatens human exceptionalism. Buckner (2024) criticizes those who move the goalposts once machines clear them, sometimes "in real time" (2024, pp. 88, 281), and Santelli (2026) describes the same maneuver in the debate over LLMs: when the impossible becomes possible, something more is demanded. The second reading sees legitimate learning. For Chollet (2019), chess taught us that a task can be solved by shortcuts alien to human cognition: skill at a task was not intelligence, and finding that out was a conceptual advance, not an excuse (2019, sec. II.1.1) (see measuring intelligence).
The AI effect has its mirror image in the ELIZA effect, which over-attributes rather than under-attributes. The debate over AGI moves between the two: when is revising a concept after a technical achievement a way of learning something about intelligence, and when is it a rationalization for not granting it to a machine?