Years of flawed, emotive reporting – especially from “high-reliability” outlets – have been baked into training data for large language models and the damage may be impossible to undo
