Every hallucination I’ve had to clean up came from a model that was sure. Not hedging — sure, in the same tone it uses when it’s right. DeepMind’s finding is that the confidence signal was there the whole time, sitting in the activations, and the model does use it to decide whether to abstain. They showed it causally: turn the signal up, fewer refusals; turn it down, more.
So I’d stop treating “I don’t know” as a failure mode in my agent loops. If a model can be instructed to abstain below a confidence threshold, that’s a knob, and knobs beat prompts that beg for honesty. For anything autonomous — an agent touching my server, my repo, my boat’s config — I’d rather it stop and ask.
The story — Google DeepMind researchers tested GPT-4o, Gemma 3 27B, Deepseek-V3 and Qwen3-Next-80B-A3B-Instruct on multiple-choice questions across four phases to see how models decide whether to answer or abstain. Published in Nature Machine Intelligence, the study found models apply an implicit internal confidence threshold, and activation steering that amplified or suppressed confidence changed abstention rates accordingly. The authors say such refusal mechanisms matter most for autonomous AI agents. (Source)