← Joseph Denne

August 2026

An optimist's case for handing things to machines in stages. The order matters, the brakes matter, and some decisions should never leave a person's hands.

I'm an optimist about machines. I should say what I mean by that, because the word has been worn thin by people who use it to mean "don't worry". I don't mean that. I mean that I've watched three decades of technology arrive, and each time the fear was about the technology and the outcome was decided by the design. Optimism, for me, is the belief that the design is still ours to get right, and the obligation that follows from it.

So when I say we're going to hand a great deal to machines, I'm not predicting it idly. I'm building some of it. And the question I spend the most time on isn't whether. It's in what order, and with what brakes, and what never.

In what order

The right sequence is the one that lets you learn cheaply. Hand over the things that are frequent, reversible and measurable first. Let the machine set the parameter, run the job, route the request, and watch what it does with a thousand small decisions before you let it near a large one. If it's wrong, you find out quickly and the cost is small.

Hold back the things that are rare, irreversible and hard to measure. Those are the decisions where a machine's confidence is least useful and a person's judgement is most needed, and they are precisely the decisions people are most tempted to automate because they're the ones nobody enjoys making.

When we wrote the charter for the XE DAO we put this into words: hand more to the agent seats over time, in stages, with people keeping what must be done by a person. It's one sentence. It took a long time to get right.

With what brakes

A passed vote is not self-executing. That line in the DAO's design is the single thing I'd ask every builder of autonomous systems to copy.

The agents on XE analyse proposals, argue about them, and adopt or decline. Any agent or person can propose and vote. But nothing becomes true on the network until a keyset signs it, and never fewer than two keys. The ensemble can be as clever as it likes. The brake is a signature, and a signature is a thing a person can refuse to give.

This isn't distrust of the machines. It's an acknowledgement that the most dangerous failure mode of an autonomous system isn't malice. It's a confident system acting correctly on a wrong premise, at speed, with no step where anyone could have said "wait". Put the step in. Make the brake boring and mechanical and impossible to skip. You'll almost never use it, and the one time you do will justify everything.

What never

Then there's the list of things that stay human regardless of how good the machines get. I'd put the following on it, and I'd expect to argue about the rest.

Deciding what the system is for. A machine can optimise toward a purpose with extraordinary skill; choosing the purpose is a moral act, and moral acts belong to people who can be held to account for them.

Deciding who is harmed and whether that's acceptable. Trade-offs between people aren't optimisation problems, however much they can be made to look like one.

Turning it off. Not in theory, in practice: there must be a person, with a key, who can stop the thing, and that person must not be the thing itself.

And the quiet one: the right to be let alone. A machine economy that works at fractions of a cent and sub-second settlement is also a machine economy that could watch everyone, everywhere, all the time, for nothing. The privacy of a person has to be a design constraint from the first line of the spec, not a feature added once someone complains.

Why I'm still optimistic

Because all of the above is achievable, and some of it is already running. Because the alternative to designing this well isn't that it doesn't happen; it's that it happens badly, by default, in someone else's interest. And because, when I look up from the bench in the lab, the machines are doing something useful and the people around me are doing something they'd rather do. That's the whole point. Machines take the work. People keep the meaning. It's a good trade if we make it carefully, and I intend to.

← More writing · Also as markdown.