AI news and explainers at 7 AM weekdays, plus a Sunday weekly at 8Get it in your inbox

METAL LAB
Add us on GoogleMETAL

Google DeepMind says AI needs probabilistic self-doubt to be safe

A VP of Research at Google DeepMind discussed how to handle uncertainty, from weather forecasting to robotics

이미지: @GoogleDeepMind (X) 영상 갈무리 · METAL LAB 편집

Summary

  • Google DeepMind's official X account posted a conversation between VP of Research Zoubin Ghahramani and Hannah Fry
  • The core argument: AI needs probabilistic reasoning and self-doubt instead of false confidence to make safe decisions in the real world
  • Using weather forecasting and robotics as examples, they contrasted an approach that builds uncertainty into calculations with today's AI hallucination problem
Video from the source

What happened

Google DeepMind posted a conversation on its official X account between VP of Research Zoubin Ghahramani and broadcaster-mathematician Hannah Fry (@fryrsquared). From weather forecasting to robotics, the two argued that intelligent decision-making ultimately comes down to knowing what you don't know. Their central claim: teaching systems probabilistic reasoning and self-doubt leads to safer, more trustworthy decisions in the real world.

On the left, an AI expresses confidence in a single answer, connected to the real world by an unstable dotted arrow. In the middle, a probabilistic AI is only half-filled, acknowledging uncertainty, and connects to the real world with a stable solid arrow. On the right, the real world is depicted as scattered dots, representing an unpredictable environment.On the left, an AI expresses confidence in a single answer, connected to the real world by an unstable dotted arrow. In the middle, a probabilistic AI is only half-filled, acknowledging uncertainty, and connects to the real world with a stable solid arrow. On the right, the real world is depicted as scattered dots, representing an unpredictable environment.

What this means

Zoubin Ghahramani led Bayesian machine learning research at Cambridge, served as Chief Scientist at Uber, and later joined Google — a researcher long known in the field of probabilistic reasoning, where AI systems judge for themselves what they're confident about and what they aren't. Probabilistic AI, instead of committing to a single fixed answer, builds uncertainty directly into its calculations — saying something like "this is correct with 70% probability." Most large language models today work the opposite way. The hallucination problem is the clearest example: a confident, plausible-sounding answer delivered even when the model doesn't actually know.

Just as weather forecasts get expressed as probabilities ("70% chance of rain"), the argument goes that robots moving objects through unfamiliar environments, or self-driving cars encountering a sudden obstacle, need to calculate "how confident is this judgment" alongside the decision itself in order to cut down on accidents. This concern connects to a broader research trend in physical AI, where robots simulate the world and plan for themselves — as seen in Sakana AI brings recursive self-improvement to 'physical AI' where robots learn on their own, one of a growing number of efforts to have robots improve their own ability to predict and act in the real world.

What changes from here

This conversation isn't a new model or product launch — it's Google DeepMind publicly laying out its view of AI safety. The message that a system capable of saying "I don't know" is safer in real deployment than one that always answers with confidence poses a question to an industry that has largely been chasing benchmark scores. No specific research results or product plans came out of this announcement, but the direction is clear: DeepMind keeps tying uncertainty quantification to real-world decision-making problems in robotics, medicine, and climate forecasting.

Comments

Related stories