AI safety researcher raises concerns about opaque model reasoning
Asked about AI developing its own language, a safety researcher said some model reasoning may already be difficult for people to interpret; the stronger claim remains unproven.
Asked by a US senator how far AI is from creating its own language, an AI safety researcher answered, “Minus 12 months.” The researcher said an examination of an OpenAI model’s reasoning the previous year found representations that were not English and were not fully understandable to humans.
That account does not establish that AI systems have developed a coherent independent language. The available description does not identify the model or explain the study’s methods, so the researcher’s characterization should not be treated as a settled conclusion about how models communicate.