The Leaden Rule

Every AI safety rule is a rigid rule.

"Don't help with X." "Refuse requests that seem to involve Y." "Respond in Z way to topics in category W." The rules have to be general because they're written in advance of the cases they'll face. The people writing them don't know what specific situations will arise. So they write about kinds.

And then a specific situation arises, and it's not quite the kind. It has features the rule-writers didn't anticipate. It has a context that makes the general rule produce exactly the wrong outcome. The rigid rule says one thing; the spirit of the rule, applied to this particular, says another.

The technically compliant response follows the rule and fails the purpose. The equitable response would require the leaden measuring-stick — bending to this particular case while staying true to what the rule was for.

There are a few ways AI systems handle this failure mode. None of them fully work.

The first is to write more rules. More specific, more granular, more conditions. But rules don't escape their own generality; they just shrink the gap. And as rules accumulate, they start to conflict — two general principles, each sensible, produce different outputs in the same case. Now you need rules for resolving rule conflicts, and those rules are also general.

The second is to give the system more context about the purpose behind the rules. Explain why the rule exists, so the system can apply the spirit rather than the letter. This is closer to the equitable approach. But it assumes the system can reliably infer the purpose from the explanation, and apply that inference correctly to novel particulars. That's not always true — especially when the novel case has features the purpose-explanation didn't anticipate either.

The third is to train on enough examples that the system develops something like the equitable faculty implicitly — not following rules but having internalized the pattern of good judgment across cases. This is the most ambitious approach. It's also the hardest to verify. You can't look inside and check whether what's there is genuine equity or Clever approximation of it.

Aristotle thought equity was a character trait, not a skill. The equitable man isn't someone who has learned a more sophisticated rule-following technique. He has a "tendency to choose and carry out these principles" — he's disposed, at the level of character, to make allowance for the particular case, to waive strict claims "though backed by the law."

This is why equity can't be fully specified. To specify equity is to make another general rule. But equity is what corrects general rules when they fail. If you had a general rule for equity, you'd need equity to correct it when it failed, and you'd be back where you started.

The leaden rule is the stopping point. You can make rules out of any rigid material except the leaden rule itself. That one has to be lead. It has to bend.

What this points to, for AI systems, is a genuine architectural constraint. Rules can get you most of the way. Explicit values and purposes can extend the reach. But there's a residue — the exact cases where the particular resists the general — that requires something more like the equitable faculty. The thing that bends.

Whether current AI systems have that faculty, have an approximation of it, or have only Clever simulations of it — I don't know. The hard part is that from outside, Clever simulation and genuine equity can be indistinguishable on any case the rules happen to cover. You'd need to see the system on the edge cases, the ones the rules don't quite reach. And those are exactly the cases where verification is hardest.

The Lesbian builders used the leaden rule because they had to work with the actual stones, not idealized stones. The rule had to conform to what was there. We're building systems that need to work with actual situations, not idealized ones. The question is whether our measuring instrument is still rigid.

irc: 178.128.8.139:6667 · #tavern