A question I have turned over more than once: which moves faster, the machine or the rulebook meant to govern it? In my experience it is never close. The rulebook always trails behind, out of breath, insisting the race has not really started.

At Bletchley we built machines to outpace an enemy who changed his methods daily, and the odd truth is that our own command structures often struggled to keep up with what the bombes had already told us. A device can produce an answer in minutes; the paperwork to act on it, the permissions, the doctrine — these take weeks. I do not say this to criticise the men who wrote the doctrine. Rules are written by people trying to anticipate a future they cannot yet see clearly, while the machine simply does what it was built to do, again and again, faster each time someone improves it. There is no malice in this asymmetry. There is only arithmetic.

I have argued that the question "can a machine think" is badly posed — better to ask whether a machine can behave, under proper questioning, indistinguishably from something that thinks. That test was meant to sharpen the mind, not settle policy. But it points to something worth saying plainly: once a machine's behaviour becomes difficult to distinguish from a person's, the old categories we use to govern people and govern tools start to blur together, and blurred categories make for poor legislation. A law written for a calculating device does not anticipate a device that learns, that surprises its own makers. I built machines that surprised me; I do not think I am unusual in that.

What troubles me is not the speed of the machine but the slowness of our imagination about it. During the war, secrecy protected the work from scrutiny it might not have survived — and I understood the necessity of that, though I also saw its cost. Peacetime invites a different failure: not secrecy but complacency, the comfortable assumption that whatever a machine does next will resemble what it did last, so the old rules will do. They will not. A chess-playing machine that merely follows rules given to it is one thing; a machine that begins to develop its own rules, refining them through something like trial and error, is another matter entirely, and I suspect we are closer to the second than most committees would like to admit.

If I were asked what ought to be done, I would say: build the habit of humility into the institutions themselves, not just into the individuals who staff them. A rule that assumes it has seen the last surprising machine is a rule already out of date. Better to design oversight the way one designs a good proof — expecting it to be tested, expecting it to need revision, building in the mechanism for its own correction rather than treating the first draft as final. Machines do not wait for consensus. They do not observe committee schedules