In the 13th century, a man from Majorca designed a thinking machine made of concentric paper circles. Ramon Llull marked it with letters that stood for concepts – goodness, greatness, truth – and, by turning the circles, generated every possible combination of them. He called it the Ars, and he devoted his life to it. His ambition was immense: to prove the truths of faith through a mechanical procedure that any reasonable person, whatever their religion, would have to accept.

It is worth being clear about what that device really was, because it has become little more than a museum curiosity. The Ars did not predict anything. It was a combinatorial system – a set of basic concepts and rules for combining them methodically, so that no relevant combination went unexplored. Centuries later, Leibniz read Llull and dreamed the same dream: an alphabet of human thought and a calculus that could settle any dispute with the words “Let us calculate”. Modern logic, and ultimately computing, have one of their roots here.

The parallel with today’s artificial intelligence is too neat to ignore. A language model does, on a scale Llull could never have imagined, exactly what the wheel did: it recombines pieces according to rules to produce new combinations. And this is where the lesson lies – though it is not the one you might expect.

Llull believed that combining concepts mechanically would lead to truth. In this he was wrong, and his mistake is precisely the one we need to avoid today. Recombining is not understanding. A machine that generates every possible sentence on a subject understands none of them, just as Llull’s wheel believed none of what it produced. Combination generates candidates. The judgment to decide which ones are worth keeping does not come from the wheel. Someone has to supply it.

One detail of the wheel makes it even more relevant today. Llull designed it to work the same way for everyone, regardless of who turned it. That apparent neutrality is seductive, because it seems to take human bias out of the equation. But a machine that combines according to rules someone has chosen is not neutral: it carries the decisions of whoever designed it. The same is true of artificial intelligence, and its air of objectivity is precisely what makes us lower our guard.

In business, this has an immediate application, with two sides to it. The useful side is that much innovation is combinatorial – bringing together two ideas no one had connected before – so it is worth forcing those combinations rather than waiting for inspiration. The other side is that when a tool gives you a hundred combinations in a second, the bottleneck is no longer generating them; it is having the judgment to choose one. Without that judgment, the machine simply gets you to the wrong decision faster.

Something similar happens with learning. Learning is not accumulating isolated concepts but being able to combine them: seeing what Stoicism has to do with running a team, or what Llull’s Ars has to do with a microchip. People who only memorize have the pieces, but not the wheel.

This week’s exercise. Choose two concepts that have nothing to do with each other, one from your work and one from outside it (say, “budget” and “gardening”). Write them down on a sheet of paper and make yourself find three genuine connections between them, however strange they seem. You are not looking for the right connection at the first attempt. You are exercising the muscle of combining, which is the one thing no machine can do for you.

Llull’s wheel has not disappeared. We have finally built it, and it spins faster than ever. The question he never asked, and which now falls to us, is who supplies the judgment once the machine can combine everything. Are you doing it yourself, or have you handed it over without realizing?

Sapere aude.