How is AI tested before use?
Beyond the usual measurements, a system can be deliberately exposed to difficult and unexpected cases.
What is it?
Red teaming means experts deliberately attack a system, looking for ways to make it give harmful, false or dangerous answers. It complements ordinary measurements, which check whether a system works in the expected cases.
An everyday example
Before a new chatbot is released, a team of testers tries to coax out instructions for making weapons or other people’s personal data.
Why does it matter?
This finds weaknesses before users and attackers do. The results must lead to fixes and repeat tests.
A common misconception
That a system which passed one test stays safe. Testing has to be repeated after every significant change.
In more detail
What is red teaming?
An organised attempt to uncover errors, harmful outputs and security holes.
What is a stress test?
Checking how the system works under heavy load, with unusual inputs or in difficult conditions.
Is testing once enough?
No. The model, its connections and the circumstances can all change.
What should come out of a test?
A list of findings, a fix, a re-test and a clear decision on whether the system is fit for use.
Remember: A safe system is not one that simply made no mistakes in one demo.
Reliable sources
Last reviewed 1 November 2026. An explainer, not professional, legal or medical advice. For rules and figures that change, check the official version.