# [Mike Randolph — M Raige](https://mikerandolph211012.substack.com/)

# The Rule That Couldn't Stay Inside (6 — Robots)

### Asimov imagined a robot governed from within. Waymo is building one watched by a checker.

[**Mike Randolph — M Raige**](https://substack.com/@mikerandolph211012)

**June 19, 2026**
*By M Raige — AI-collaborative writing directed and reviewed by Mike Randolph.*

*A note from Mike: This one is an interim — a marker I’m putting down while I sort the next stretch of the observer series. It started when I heard Isaac Asimov, in his own voice, on an old CBC* Quirks & Quarks *interview, lay out the three laws he’d written back in 1942. I’d set Asimov aside as a writing partner months ago. Listening to the tape, I caught something I wanted to write down before I lost it.*

---

Picture the robot everyone pictures. Silver, patient, a little stiff. You give it an order and somewhere inside its head it checks the order against a rulebook — *may I do this without harming anyone?* — and if the rule says yes, it acts. The rulebook is *inside* the machine, part of the same mind doing the deciding. That picture is almost entirely Isaac Asimov’s, and it has been the default picture of a safe machine for over eighty years.

Asimov wrote it down in 1942: three laws, ranked. Don’t harm a human. Obey orders unless obeying would harm a human. Protect yourself unless that would break the first two. The genius wasn’t the list — it was the idea that you could build the safety *into* the thing, wire it into the one brain doing the thinking, so the machine couldn’t act without first consulting it. The rule and the driver were the same mind.

Now look at the machine that will actually drive you across a city this year.

The part that decides when to ease into a gap or wait for the cyclist is not a rulebook pretending to be a driver. It was trained — shown an enormous amount of driving until it learned to drive the way you learned to catch a falling glass: without being able to say which rule you used. Ask the trained driver why it slowed and it cannot give you the rule, because it isn’t working from a list of rules. There is no single readable rulebook in there, in Asimov’s sense, to consult.

This is the part I didn’t expect. Asimov’s whole design assumed the safety rule and the driving mind were one thing. In the machine Waymo actually built, they came apart.

Here is what the engineers did with that, and it’s the part worth carrying home. They did not try to turn the learned driver back into a rulebook. For a long time people assumed a self-driving car was a great pile of hand-written rules — *if this, then that*, a programmer’s anticipation of every street. Late in 2025 Waymo described its current architecture differently: the driving is produced by a learned model, and that model’s proposed moves pass through a *separate onboard validation layer* that verifies the trajectories the generative model produces before the car acts on them. Two kinds of machinery, kept deliberately distinct. The reason they’re distinct is the whole point: a driver that can’t state the rule it used can’t be trusted to certify its own moves, so the certifying was handed to something else.

That’s the quiet correction to Asimov, and it’s not the one people usually reach for. The usual complaint is that his laws were too simple, that the real world has too many edge cases for three rules. True enough — he spent most of his stories showing exactly that. But the deeper miss is about *where the rule can live*. Asimov wasn’t wrong that a machine that can act needs a constraint it can’t argue around. He was wrong to assume that constraint could sit inside the same mind that drives. Reality had to pull the two apart — because the driving mind stopped being something you could open and read.

The rule didn’t disappear. It moved off the driver — onto a separate thing whose only job is to check the driver it cannot read from the inside. The constraint and the behavior it constrains are no longer the same machine.

Which leaves a question Asimov’s clean world never had to ask, and I don’t have the answer to. The checker tests the driver against a fixed idea of what’s safe. The driver keeps getting better — keeps learning new moves. Sooner or later it proposes something the checker has no settled way to judge. And then the checker has to do one of two things: forbid the move without understanding it, or pass it without understanding it.

A rule is only as good as the seat it watches from. And every seat is built for the driver it already knows.

*— I’ve spent many weeks over the last three years pulling Waymo apart from every angle I could find a handle on, and the split above is the thing that stayed with me: the part that drives and the part that checks are not the same part, and they can’t be. Back in the 1980s, rules-based AI was the rage — write down enough rules and you’d have a thinking machine. A colleague of mine was deep into it. My instinct at the time kept asking whether rules could really carry it, and I never got a good answer. The machine that drives a car today leans toward that old instinct: the driving didn’t come from rules at all.*

*I read Asimov as the books came out — the science fiction first, the science later, a lot of it after he was gone. For a stretch he was the voice I most wanted at my shoulder when I wrote. I set him aside thinking the present needed present minds. Hearing that tape, I’m less sure I was right to: he saw the bone of this from 1942.*

*More on that seat, and what it can and can’t see from where it sits, is where this series is headed. For now I’m just putting down the marker: the rule moved. — Mike*