AI-written critiques help humans notice flaws

Researchers trained language models to generate critiques that help humans catch mistakes in difficult evaluation tasks.

The proof of concept focused on critiquing topic-based summaries of short stories, Wikipedia articles, and other web text.

Results suggest AI-assisted feedback can make human evaluators more effective at spotting flaws and factual errors.

The work highlights a path toward better model alignment by improving the quality of human feedback.