Researchers Made an AI Model Almost Stop Saying 'I Don't Know'

Google DeepMind researchers moved a language model's refusal rate from 66.5% to 7.0% by intervening directly on a confidence signal inside it, in a study published Sept. 7 in Nature Machine Intelligence. Earlier work had shown that such a signal can be read out of a model's output; this study tested whether behavior follows it.

The paper frames the work around when a model should hold back. Wrong answers to high-stakes questions (the paper gives medicine as an example) are usually more harmful than refusals, the authors write, and they argue that recognizing uncertainty matters more as models are handed tasks to run on their own.

The causal test used activation steering, which adds a constructed signal to a model's internal layers while it runs. It was done on Gemma 3 27B, the model that the paper says was used for this phase because the technique needs access to a model's internals. Without any intervention, the model declined to answer 28.2% of the time on the 500 held-out questions.

A breakdown of where the effect came from attributed 67.1% of it to confidence shifting away from the refusal option and toward the answer options, with a smaller share coming from a change in the cutoff the model applied. Answering more came at a price: accuracy on the questions the model did answer fell from 59.2% to 53.7%.

Separately, with no intervention at all, the team put GPT-4o's implicit cutoff at about 77% confidence, the point at which it answered half the time and declined half the time.

The authors describe the results as "consistent with structured metacognitive control in large language models." They also write that their evidence "does not by itself uniquely establish" the stronger reading, in which the model weighs its own answer rather than simply reading off how strongly it favored one option.

The four models tested, GPT-4o, Gemma 3 27B, DeepSeek 671B and Qwen 80B, differed widely in how readily they declined at all. A fifth, Llama 3.1 70B, was left out of the analysis because it almost never did.