Which term refers to an AI system's ability to perform correctly even when someone is actively trying to make it fail?

Which term refers to an AI system's ability to perform correctly even when someone is actively trying to make it fail?

• A. Adversarial Robustness
• B. Frontier Model
• C. AI Welfare
• D. AI Safety

Answer: (A)

Adversarial robustness is about keeping a system’s performance steady even when someone deliberately tries to cause it to fail. It targets the vulnerability that arises when inputs or prompts are crafted to trick the model into making mistakes, misclassifying, or producing unsafe outputs. The aim is to build defenses so that small, carefully chosen changes don’t derail correct behavior, and to ensure reliable operation under adversarial pressure. This focus on resisting intentional manipulation is what makes the term a precise fit for the scenario described. Frontier model describes how advanced or cutting-edge a model is in terms of capabilities, not its resilience to deliberate attacks. AI safety is a broader umbrella about preventing harm and aligning behavior, but it doesn’t specifically denote robustness against adversarial manipulation. AI welfare centers on societal and ethical impacts rather than the technical aspect of resisting adversarial attempts.

Adversarial robustness is about keeping a system’s performance steady even when someone deliberately tries to cause it to fail. It targets the vulnerability that arises when inputs or prompts are crafted to trick the model into making mistakes, misclassifying, or producing unsafe outputs. The aim is to build defenses so that small, carefully chosen changes don’t derail correct behavior, and to ensure reliable operation under adversarial pressure. This focus on resisting intentional manipulation is what makes the term a precise fit for the scenario described.

Frontier model describes how advanced or cutting-edge a model is in terms of capabilities, not its resilience to deliberate attacks. AI safety is a broader umbrella about preventing harm and aligning behavior, but it doesn’t specifically denote robustness against adversarial manipulation. AI welfare centers on societal and ethical impacts rather than the technical aspect of resisting adversarial attempts.