Is the “stochastic parrot” framing both empirically wrong and harmful to AI ethics? (SE Gyges, Very Sane AI)
- The 2021 paper "On the Dangers of Stochastic Parrots" argued LLMs shuffle text without understanding meaning, but every major frontier model since GPT-4 is trained on paired text, image, audio, and video data, satisfying the paper's own stated conditions for grounding.
- The circular definition at the heart of the argument (meaning requires something extralinguistic, training data is linguistic, therefore no meaning) is unfalsifiable by design, and competing frameworks in linguistics and philosophy of language reach different conclusions without being engaged.
- The practical damage is that dismissing LLMs as incapable lets real harms, like minority-language surveillance tools or AI-assisted mass surveillance programs, slide past critics who have already decided the technology cannot work.
