What's this about?
Mollick points at three pre-AI pages on complex systems and draws the implication: complex systems run broken but survive because the flaws rarely line up, which gives humans a chance to intervene. AI can find — or make — flaws that do line up. We need a similar new philosophy of system defence using AI.
What is the "flaws rarely line up" argument?
It comes from Richard Cook's work on how complex systems fail. Every complex system is running with multiple latent defects at all times; catastrophe requires several of them to align simultaneously. Systems are safe not because they are correct but because the defects are uncorrelated and the coincidence is rare. Human operators catch most near-misses precisely because the alignment is partial and slow. An adversary that can search the whole defect space and deliberately arrange the coincidence removes the property the safety depends on.
Why does this matter to you?
- It is the clearest available framing of why AI-enabled attack differs in kind, not just in speed.
- Cook's work is the standard reference for safety engineering, and applying it to adversarial AI is the right transfer.
- The constructive half — "a new philosophy of system defence using AI" — is unspecified, which is honest, because nobody has one.