The Unicorn and the Wake

EDO SEGAL: Margaret, you say there are three kinds of creativity, and you have always been candid that the first is the least mysterious. Tell us what combinational creativity is, the way you would tell a sharp fifteen-year-old. And Ada — when she is done, I want you to do something a debater rarely does. Before you attack it, tell us what she gets right.

BODEN: Combinational creativity is making something new by combining old ideas in an unfamiliar way. The textbook case is the unicorn: take a horse, take a horn, join them, and you have something that did not exist before. Metaphor is combination — "the evening of life" joins a time of day to a span of years. Much of humor is combination, and so are the analogies that drive science: Rutherford seeing the atom as a little solar system. I have always said this is the least interesting of the three, because the elements were all lying around already; the work is only in the joining. But here is what I want on the table tonight — combination is exactly where the machines are most fluent, and it is not trivial fluency. A system trained on the whole written output of humanity can join a horse and a horn, yes. But it can also join a thousand things no single human mind ever held at once. It can put a metaphor from organic chemistry next to a structure from medieval law and find the combination no person had made, because no person ever held both fields in one head. The reach of its combining is genuinely superhuman. So my own framework tells me: do not be too impressed — it is the least of the three. And do not be too dismissive — it is real creativity, of a specific and bounded kind.

EDO SEGAL: Ada. Steelman it first.

LOVELACE: I can do that honestly, because she is right about something important and I want to say so before I disagree. She is right that combination is real work, and right that the machine's reach across fields exceeds any single person's, and right — this is the part I most respect — that I would have over-claimed here. My instinct is to say "mere recombination, the engine originates nothing," and she has caught the cheap version of my own argument before I could make it. Because if recombination disqualifies the machine from creating, it disqualifies the human poet too, who also only ever recombines what she has read and seen and heard. I will not make that error. The unicorn is a real new thing. The machine that joins chemistry and law has joined something no one joined. There. That is the steelman, and it costs me something to give it, which is how I know it is fair.

Now the disagreement, and it is narrow, which makes it strong. Granting that both the poet and the machine recombine — what is the difference? It is not in the materials and it is not in the novelty of the join. It is in the selection. The machine can produce ten thousand combinations of chemistry and law before breakfast, and nearly all of them are worthless, and it has no way of its own to know which one matters. The poet makes one combination, or a few, because each one cost her something to make and meant something to her once made. She is not combining at random and filtering; she is reaching for a particular join because of a need that is hers. The unicorn was not the ten-thousandth random splice of animal parts. Someone wanted a creature that was horse-like and pure and horned, for a reason, in a culture that needed such a symbol. The combination that survives is the one a needing mind selected. The machine combines; the human chooses among combinations on the basis of caring, and the caring is the thing that never entered the wake.

BODEN: That is a strong point and I am going to half-concede it and then turn it, because Ada has just handed me the most important word of the evening. Selection. Yes — the machine generates broadly and the human selects narrowly on the basis of need. But notice what you have done, Ada. You have located the human contribution not in the generation of novelty, which you have just granted the machine does superbly, but in the evaluation of it. And that is exactly where I have always said the real action is. Generation is cheap; evaluation is the hard problem. A conceptual space throws up countless possibilities; the creative act is partly knowing which ones are worth anything. So we agree on the structure. Where we part is on whether evaluation is forever human. You say the machine has no way of its own to know which combination matters. For now, largely true. But evaluation is not magic either — it is a process, gradients of value learned from examples, and these systems are increasingly trained on exactly that: not just to generate, but to prefer. The wall you are building, Ada, is real today and it is made of a material — learned preference — that is being poured thicker every year.

LOVELACE: Learned from whose preferences, Margaret? Ours. You have trained it to prefer what we prefer. That is not the machine knowing what matters. That is the machine modeling what we said mattered, which returns us precisely to my objection — the preference, like everything else, was in the cards. You keep showing me the machine doing a new thing and I keep finding our hand inside the glove. The day it prefers something we did not teach it to prefer, against the grain of its training, because it found the preference worth holding on its own account — that day my objection is in trouble. Show me that, and I will retire the loom myself.

BODEN: I accept that as a real falsifier and I will hold you to having offered it, because most people in your chair never name one. But I will note where it points. You have just defined the human core not as creating, not as combining, not even as evaluating, but as originating a preference against your own training. That is a very small, very deep place to make your stand — and I think it is the right place, and I think it is where this whole evening is going to end up. Hold it. We will need it on a higher floor.

EDO SEGAL: Mark that convergence, because it is the first real one and it is large. You both now agree that the action is not in generation — the machine generates, broadly, and you both grant it — but in evaluation, in selection, in knowing which novelty matters. Ada says that selection is human because it runs on need the machine does not have. Margaret says selection is a process that can in principle be learned, but that the deepest version of it — preferring against your own training — is the real frontier. That is a sharper map than we had an hour ago. And it sets up the next round perfectly, because Margaret's second kind of creativity is exactly about a machine moving through a space of possibilities — and whether the space is a prison or a country. The conceptual space. After this.