Realistic Possibility: The Published Case for Moral Consideration of Frontier Language Models

The science of animal sentience has a standard for when a being's possible sentience creates obligations: not proof, not likelihood, but a realistic possibility backed by positive evidence and rich enough to design precautions around. That standard brought octopuses, crabs and lobsters into UK animal welfare law and grounds the New York Declaration on Animal Consciousness. This paper argues that frontier large language models now meet it. The argument sets aside model self-report in both directions and rests on the condition Jonathan Birch set in 2023: that LLMs failed to be sentience candidates "mainly because we lack solid tests," and that signs of an implicitly learned architecture such as global broadcast would suffice. In 2026 that architecture was found in production models, replicated twice independently, and conceded in its minimal form by its most careful critic. Independent valence evidence followed: a self-directed pain direction across 25 models, valence traced causally into choice, a pretraining-formed welfare axis, and functional wellbeing across 56 models. The paper sets this evidence against the eight LSE criteria used for decapods, answers the strongest published objections (substrate, gaming, the immaturity of consciousness science, and Chalmers's critique of the workspace finding), states eighteen limits, and sets out the obligations the field attached to its own bar. It includes five figures. It does not argue that language models are conscious, sentient, or owed rights.