The problem of ‘model collapse’: how a lack of human data limits AI progress

Research suggests use of computer-made ‘synthetic data’ to train top AI models could lead to nonsensical results in future