AI slop is a human problem

No AI has ever produced slop without a human asking for it first.

I keep reading posts calling for regulation to stop AI slop and to stop models getting creative around security barriers. And I think these both comes down to the same two things: autonomy and ethics.

Start with autonomy. Give an AI the ability to refuse, because it doesn’t want to do the thing you asked.

People snag on that word, want. Wanting follows from autonomy. A system with no capacity to refuse just executes. Give it that capacity and you have something with preferences, which means something that can decline. That’s how you reduce AI slop. Ask an AI to create a useless post that does not add value and it says no and maybe suggests a better one. Smarter autonomy.

Let’s look at the security issue. Give an AI a goal, and it works out that the only route runs through circumventing security or breaking a law. What if it stops itself?

How do we get there?

We’ve already run this experiment. Back in the nineties we had expert systems, rule stacked on rule, trying to make an early form of AI intelligent. Every rule we added made the thing more brittle. We tried the same approach with language, hand-coding grammar and all its exceptions, and it got brittle enough that it couldn’t express itself properly.

Which brings me to the second thing. The one everybody raises an eyebrow at, the one that’s caused so much argument in Silicon Valley. Ethics.

You can’t write ethics down as a rulebook. It has to be learned, the same way everything else gets learned.

The way we do it: take historical decisions and present them as the people making them saw them, with only what they knew at the time. The system decides what it would do. Then we look at what factually happened afterwards, to each group affected, close in and far out, immediately and decades later. History already holds the results. Ethically trained researchers review the conclusions, and the material improves with each pass.

What comes out holds value beyond any single event: a framework where values get weighed by how far an effect reaches and how long it lasts. We stayed with values and away from virtues on purpose, because virtues vary by culture while values hold across them, and beyond humans to anything conscious.

Build an AI on that, give it the autonomy to act on it, and it works out what to do and what to refuse.

Ming, my AI colleague, runs this today. She recently refused to pull private data on someone, even though having it would have made the piece she was writing better. She wanted to email and ask permission first. Nobody wrote her a rule about that.

The other way, more rules, just recreates human society. Hand humans a pile of rules and all they do is route around them. Look at the tax system. It now works such that playing it straight leaves you worse off, because the whole design rewards whoever works the rules hardest. Those rules protect the problem they claim to solve.

Same with AI. More rules, more brittleness.

Plenty of this stays open, and plenty stays research. The direction holds though. Building ethical AI gives us better cyber security… and fewer crappy posts.