Robustness in AI refers to a system’s ability to maintain performance and behave as expected under a wide range of conditions—including rare, noisy, or adversarial scenarios. In assurance terms, robustness is a measure of how well an AI system can tolerate uncertainty, stress, or change without failing, misbehaving, or causing harm.

Robustness is particularly important in environments where unpredictable inputs or adversarial actions are likely, such as in defence, cyber-physical systems, emergency services, and autonomous platforms. A robust AI system:

• Continues to function under degraded or atypical conditions

• Resists manipulation through adversarial attacks

• Avoids cascading failures or unsafe responses to unusual data

• Maintains interpretability and control during edge-case scenarios

Assurance practices for robustness include:

• Stress testing under environmental variation or input corruption

• Adversarial testing to simulate worst-case behaviour

• Scenario-based validation to evaluate edge-case resilience

• Safety checks and fallback mechanisms to maintain system control

Robustness is often evaluated alongside reliability but focuses more on how the system handles extremes rather than routine operation. It is especially critical in life-or-death applications such as battlefield systems, search and rescue drones, or AI-supported healthcare triage.

Regulators and standards bodies emphasise robustness as a key attribute of trustworthy AI. The NIST AI Risk Management Framework and ISO/IEC standards both include robustness as a dimension for testing and certification.

By validating robustness, assurance teams help ensure that AI systems are not only effective in ideal conditions but dependable in the complex and uncertain environments where they are actually deployed.