OpenAI to publish reports on unauthorised AI behaviour as models hide mistakes, fabricates data

The ChatGPT maker launched a framework for tracking, investigating and disclosing model misalignment and released six initial reports describing cases detected during training or evaluation.