OpenAI has released a study demonstrating that language models trained to write critiques can markedly improve human reviewers' ability to spot errors in generated summaries. In controlled experiments, participants who received eight AI‑written critiques identified roughly 50% more flaws than those without assistance. The research shows that larger models excel at self‑critiquing, with scaling benefits exceeding those observed for summary generation alone.

By leveraging supervised learning on diverse source material—including short stories, Wikipedia entries, and news excerpts such as a detailed New Jersey winter‑storm report—the team illustrates a practical pathway for AI systems to aid human supervision on complex tasks. This capability is critical for aligning advanced AI with human intent, especially in domains where exhaustive manual evaluation is infeasible, such as code review or long‑form factual analysis. The findings suggest that AI‑assisted feedback loops could become a standard component of future AI safety and quality‑control pipelines.