A likely answer isn’t necessarily a true one. David Lopez-Paz’s essay Against All Likelihood, submitted to arXiv on November 6, 2026, argues that language models should assess truth by whether a claim holds across relevant environmental changes—not just by its probability. Source
DAIR.AI identifies Lopez-Paz as a researcher at FAIR, Meta. Source His essay describes maximum likelihood as the dominant training principle for large language models, then challenges its adequacy as a truth test. Unlikely events happen, and frequently repeated statements can be false. Familiarity is not a fact-check. Source
The proposed criterion is invariance. The abstract says truth should be judged “not by its probability under a single distribution, but by its invariance across relevant transformations of the environment.” Source
The proposed test concerns relevant changes, not every conceivable change of context. Getting the same answer from a model in different settings isn’t, by itself, proof that it’s true. Source
Lopez-Paz draws on philosophical connections between objectivity and invariance, naming Nozick, Wigner, Poincaré and Cassirer. He connects that framing to invariant learning and out-of-distribution generalization—the question of how models perform beyond the conditions represented in their training data. Source
Lopez-Paz also sketches how models could test invariance as they answer and considers what the principle could mean for models that keep learning. Source
Comments