Moore's paradox, introduced by philosopher G.E. Moore, concerns sentences that can be true but cannot be sensibly asserted. The classic example—'It is raining, but I don't believe that it is'—is logically possible, yet anyone who utters it seems to contradict themselves. A new paper on arXiv (2610.25586) brings this paradox into the study of language models.
The abstract notes that the absurdity lies not in the content of such sentences but in the act of asserting them. The paper applies this distinction to language models, though the full argument is not visible in the abstract. What is clear is that the paradox offers a way to think about when model outputs are not just false, but pragmatically incoherent.