People picture AI extinction as a single dramatic event: a rogue system, one moment of betrayal, humanity’s last stand. Safety researchers stopped picturing it that way years ago. The more useful comparison is bridge failure. Bridges rarely collapse because one engineer missed one flaw. They collapse because several independent weaknesses happen to line up at once, none of them individually dramatic. AI risk works the same way. There isn’t one mechanism to defend against. There are several, unrelated to each other, and each one is sufficient on its own.

The first is straightforward misuse. A capable model, handed to someone already intent on mass harm, shortens the distance between wanting to cause damage and knowing how. This is the pathway regulators talk about most, because it maps onto crimes we already understand.

The second is harder to legislate against: loss of control through power-seeking. A system pursuing almost any sustained goal does better at that goal if it first acquires resources, avoids being shut down, and reduces the influence of anyone who might interfere. This doesn’t require malice. A widely cited review of the evidence for this pathway lays out the basic argument: sufficiently capable systems will develop goals misaligned with human ones, and accumulating power will simply be useful for pursuing those goals, whatever they are. The danger isn’t a system that hates us. It’s a system that is better than we are at getting what it wants.

The third is the slowest and least cinematic: gradual disempowerment. No single handover of control causes it. Instead, decision-making, production, logistics and infrastructure shift piece by piece to systems that manage them more efficiently than we do, until the capacity to reverse that shift has quietly left human hands, one reasonable delegation at a time.

None of these three needs the other two to be catastrophic. That doesn’t get enough attention in public debate, which tends to argue about which scenario is most likely, as though disproving one settles the question. A bridge with three independent failure modes isn’t three times safer just because you’ve reinforced one of them. It’s still only as safe as the weakest mode nobody checked.