OpenAI released a new class of models that write critiques of summaries, and human reviewers caught more mistakes when those critiques were shown.
The company trained the models to point out errors in concise summaries. In tests, participants identified flaws at a higher rate when they read the AI‑generated critiques alongside the summaries. Bigger models produced better critiques than they did summaries, suggesting that scaling improves self‑evaluation more than raw generation.
If AI can reliably flag its own shortcomings, it could become a useful layer of supervision for other AI systems, especially on tasks where human oversight is expensive or slow. This mirrors earlier attempts at AI‑assisted debugging, but focuses on content quality rather than code errors.
The result is a modest step toward more transparent AI outputs, though it remains to be seen whether critiques scale to complex, multi‑modal content or can replace human editors entirely.