Study: AI Models Learn to Refuse Answers When Uncertain

Researchers at Google DeepMind conducted a study on large language models (LLMs) including GPT-4o, Gemma 3 27B, Deepseek-V3, and Qwen3-Next-80B-A3B-Instruct to investigate how these models decide whether to answer a query or abstain due to uncertainty. Using an experimental paradigm with four phases, they found that models apply implicit confidence thresholds, and that steering their internal confidence levels causally affects abstention rates. The findings suggest that models can be made to refuse answers when their confidence is low, potentially reducing hallucinations. This ability is considered crucial for autonomous AI agents that must recognize their own uncertainty. The study was published in Nature Machine Intelligence.