In plain English
Red-teaming is the practice of deliberately testing an AI system by trying to make it fail, produce harmful outputs, or behave unexpectedly, in order to find vulnerabilities before deployment. It borrows the idea from security, where a friendly team plays the attacker. The findings then feed back into guardrails and design.
Why it matters
Systems behave differently under adversarial pressure than in friendly testing. Red-teaming surfaces the failure modes that real users, and bad actors, will eventually find.
A worked example
A team tries dozens of phrasings to trick a customer-service bot into giving banned advice, then uses what works to strengthen its guardrails.
Common confusion
Red-teaming is not ordinary quality testing. It actively tries to break the system rather than confirming it works under normal use.