Trustworthy AI

Trusted Solutions

• Users might not trust ATP
• Empirical testing provides assurance of correctness
• Formal verification provides evidence of correctness
• AI systems whose solutions consistently pass formal verification become trusted

The Cycle of Trust

1. An untrusted AI system produces untrusted solutions
2. Untrusted solutions are checked against expectations
3. The syntactic form of untrusted solutions is checked
4. Well-formed solutions can be logically verified.
5. Steps 1-4 accumulates evidence that induces trust
6. Enough positive evidence leads to trust (a socio-empirical process)
7. Trusted solutions can still impact the level of trust

Useful Solutions

• Solutions must be comprehensible to applications (including humans)
• Comprehensible trusted solutions are useful