The Mathematics of AI Uncertainty

Cambridge professor and Google DeepMind researcher Zoubin Ghahramani discusses why AI systems should represent uncertainty and use it in decisions. The interviewer is addressed as Hannah, but the available task metadata does not establish a full name.

Ghahramani argues that decision-making requires accounting for limited observations and an uncertain future. A system should be able to update its beliefs when it receives new evidence.

The speakers distinguish inherent variability from uncertainty caused by limited knowledge. The distinction matters because more information may reduce one kind of uncertainty without eliminating the other.

Ghahramani describes probabilistic models of human perception and decision-making, while acknowledging that people can make substantial mistakes when explicitly estimating probabilities. Such models are useful explanations, not proof that the brain literally performs a particular calculation.

An adversarial image can be classified incorrectly with high confidence. The example separates correctness from calibration: a useful system should not merely produce a confident-sounding answer.

Ghahramani recounts early work on neural networks and a later focus on Bayesian machine learning. He notes that modern language models already predict token probabilities, but argues that representing uncertainty about the world is a broader problem.

Bayes’ rule combines prior beliefs with the likelihood of new evidence and normalizes the result to obtain a posterior. Further observations can update that posterior again; the result depends on the model and assumptions used.

The discussion contrasts explicit probabilistic reasoning with learning statistical patterns from data. Ghahramani sees room for architectures that better represent uncertainty, while acknowledging that his proposed direction is still research.

As an example, he describes weather forecasting with an ensemble of possible outcomes. A distribution can communicate a range of plausible storm tracks and can be updated as observations arrive.

The speakers discuss uncertainty in protein-structure predictions. Confidence information can help users identify less certain regions, but a prediction remains a model output rather than experimental confirmation.

For repeated events, calibration means that outcomes occur at roughly the stated frequency: among many cases assigned a 70% probability, about 70% should occur. Calibration is different from being correct on every individual case.

In consequential settings such as medical decisions, the speakers emphasize communicating uncertainty to professionals. A confidence display alone is not evidence that a system is calibrated or safe to use clinically.

Ghahramani identifies continual learning, energy efficiency and data efficiency as research challenges. He presents Bayesian updating as a useful theoretical lens while acknowledging that exact inference can be computationally intractable.

He hopes future systems can better communicate their limits, but says more research and approximations are needed. The discussion closes on designing AI to complement human judgment.