The idea
AI encompasses methods that learn patterns, predict outcomes or generate content. They can help people think and work, but should be evaluated against evidence.
A persuasive answer is not necessarily a correct answer.
How it works
Learning
Training adjusts a model using examples; good generalisation means performing well on relevant unseen cases.
Uncertainty
Evaluate error rates, calibration and how performance changes on unfamiliar data.
Agents
Tool-using AI can act on other systems; permissions, logs and independent checks matter.
Under the hood
Distinguish training, validation and test data. Measure failure patterns and distribution shift. Confirm external actions using the target system, not the model's report.
One small experiment
Ask an AI to explain a scientific claim and then verify the cited evidence.