AI system robustness is its ability to function reliably even with unexpected inputs, errors, and attempts at misuse. It covers three areas. The first is accuracy in real-world operation, which tends to be lower than in testing and declines over time as the environment changes. The second is resilience to erroneous, incomplete, or deliberately distorted inputs, including attempts to bypass the system's rules. The third is the cybersecurity of the model itself and of the interfaces through which it is accessed. A practical measure is putting a fallback procedure in place for an outage or an obviously incorrect output, so that the company's operations do not depend solely on a tool that can fail without warning. Regularly checking accuracy against current data is therefore just as important as testing before launch.
See also: Testing an AI system before deployment, AI system monitoring, NIS2 and cybersecurity.
