This has been standard operating procedure in AI development forever: the instant it passes some test, move the goalposts and suddenly begin claiming it was a bad test all along.