Red Teaming
Red teaming in AI is structured adversarial testing in which experts deliberately try to make a model fail or behave harmfully, to find weaknesses before release.
What is Red Teaming?
Red teams probe for harmful content, bias, security flaws and dangerous capabilities such as helping with weapons. Providers of systemic-risk GPAI models must carry out adversarial testing under the EU AI Act.
How it works
Internal and external testers, sometimes aided by automated tools, attack the model and document findings for mitigation.
← All Artificial Intelligence & Superintelligence definitions · ← Back to the A–Z definitions