Uncertainty Quantification

Uncertainty quantification means having AI tell you not just an answer, but also how confident it is in that answer. It's the same principle as a weather forecast saying "70% chance of rain" instead of simply "it will rain tomorrow." It's just one extra number, but that number makes a big difference in helping people decide whether to grab an umbrella.

Today's popular conversational AI systems are weak in this habit. Even when asked something they don't know, they often answer confidently as if they knew the right answer — and this problem of making up false information as if it were fact is called hallucination. In contrast, a system with good uncertainty quantification knows how to say "I'm not sure about this."

This ability matters especially for AI that operates physically in the real world, like robots or self-driving cars. When moving an unfamiliar object, or encountering an obstacle that suddenly appears on the road, calculating "how reliable is this judgment" alongside the decision itself can reduce accidents — by slowing down or asking a human for help when confidence is low.

Articles describe things like "DeepMind continues to connect uncertainty quantification to real-world decision-making problems such as robotics, healthcare, and climate prediction." A common misunderstanding here is that this isn't a feature that lowers AI performance or makes answers vaguer. Rather, the point is that accurately knowing one's own degree of confidence itself leads to safer decisions in actual deployment environments.

After getting an answer from a chatbot, follow up by asking:

"On a scale of 0-100%, how confident are you in your last answer, and if your confidence is low, tell me why."

You can also try asking the same question multiple times with slightly different phrasing and compare whether the answers stay consistent or waver — this can be another way to gauge how confident the AI actually is.