2026年04月24日 日本語

MIT Develops Training to Make AI Say “I’m Not Sure” - Tackling Hallucination

MIT researchers announced on April 22, 2026, a novel training method that enables AI models to express their confidence levels more accurately. This technique enhances the AI’s ability to appropriately indicate when it recognizes it “doesn’t know.”

This advancement is crucial for improving the reliability of AI models, particularly those engaged in reasoning. “Hallucination,” the phenomenon where AI generates factually incorrect information, is a significant challenge in AI applications, and this new training method is expected to address one of its root causes.

This technology, which improves reliability without sacrificing performance, represents a groundbreaking step toward the safe and effective utilization of AI systems.