How to verify AI safety?

By assessing whether an AI system behaves reliably, predictably, and in alignment with human intentions

1- Behavioral Testing

• Create test cases (normal, edge, and adversarial inputs).
• See if the AI system behaves as expected.

2- Robustness Evaluation

• Test how well the system handles perturbed, noisy, or adversarial data.

3- Formal Verification

• Use logic and math to prove safety properties of simpler or constrained systems.

4- Alignment Checks

• Evaluate whether the model’s behavior matches human values or instructions.

5- Transparency and Interpretability

• Understanding how and why the AI makes decisions:

6- Monitoring & Oversight

• Put the AI under human supervision (human-in-the-loop or human-on-the-loop).