Early in this series I wrote that the loudest argument about AI is the least important one, and that the shouting about consciousness and hype was providing excellent cover for decisions being made quietly elsewhere. Here’s the constructive half of that, eighty-eight posts later. We’re going to argue either way. These are the three things worth arguing about.

The first is liability. Whoever builds and deploys a system carries legal responsibility for what it does, on the same terms as any other product that can hurt someone. Not a new agency, not a licensing regime, not a definition of intelligence that lawyers will fight over for a decade. Just the ordinary principle that applies to a ladder manufacturer, applied to this. It’s the most powerful lever available for one reason: it changes the arithmetic inside every company without anyone having to predict the future. Insurers start asking hard questions because their money is at stake. Caution becomes a line item rather than a virtue, and line items survive competitive pressure while virtues never do. One country can do it alone.

The second is transparency, and I mean the narrow version that could actually pass. Not open sourcing everything, which is a different argument with real people on both sides. Three specific things: that a system’s capabilities are tested by somebody who doesn’t work for its maker, that serious incidents get reported to somewhere they can be counted, and that a person subject to an automated decision is told what it was and why. Every mature safety regime in history rests on incident reporting, and the reason aviation is safe is not that pilots are careful but that every near miss for sixty years went into a database that everyone learned from. We have no such database here. There is no denominator for anything.

The third is the off switch, and by now you’ll know I don’t mean a red button. I mean the four conditions from the middle of this series treated as things a society deliberately maintains. Fallbacks that are exercised rather than documented. Somebody in the building who still understands the rules the system applies. A required delay before an automated decision removes someone’s access to their money, their property, or their livelihood, which is the cheapest safeguard on this list and costs nothing except the speed that was being optimized for. These are unglamorous operational requirements and they’re the difference between a dependence you can walk out of and one you can’t.

Now, why these three and not the ones currently occupying the debate. Each has a property the popular arguments lack. None requires anyone to predict the future or agree about how dangerous this is: a person who thinks the whole thing is overblown can support all three on consumer protection grounds alone. Each is checkable, so you can tell whether it happened. Each survives being implemented badly, which matters enormously because everything is implemented badly. And each is boring, which is the highest praise available in policy, because boring things get passed while exciting ones get debated.

Compare that to what the argument is actually about. Whether these systems are conscious, which no proposal turns on. Whether the technology is overhyped, which is a prediction dressed as a position. Whether worrying about the long term distracts from present harms, which has consumed an extraordinary amount of energy from people who agree about all three demands above and would rather fight each other. That last one is the most expensive argument in the field, and it is a fight between allies conducted at full volume in front of the people it should be aimed at.

The honest objection is that liability could be captured by incumbents who can afford the insurance, that testing regimes get gamed, and that mandatory delays will be lobbied down to something meaningless. All true, all the normal fate of all regulation, and none of it an argument for nothing. Rules get captured and diluted and still leave us with medicine that mostly doesn’t poison people and aircraft that mostly land. The choice has never been between a good regime and a bad one. It’s between a diluted regime and no regime, and the second option is what we currently have.

Tonight’s exercise. Next time this comes up in conversation, notice which argument the room is having, and try to move it to one of the three. Not by lecturing. By asking a question: who’d be liable if that went wrong, who checks it apart from the people selling it, what happens if it’s down for a week. The subject changes immediately, and the reason is that those questions have answers, and the answers are embarrassing. Tomorrow I make the best case I can that this entire series has been wrong.