Here is an uncomfortable exercise for anyone with an opinion about AI and creativity. Put two ad campaigns side by side — both genuinely surprising, both effective, both fresh. One came from a human team. One came from a model. You can't tell which is which.
If the outputs are indistinguishable, then the claim "AI can't create" has stopped being a claim about outputs. It has to be a claim about something else. Most of the noisy debate online never gets this far, which is why both camps talk past each other: one side points at impressive artifacts, the other side points at a feeling that something essential is missing, and neither has named what that something is.
This piece is an attempt to name it. Fair warning: it ends in a question. That's deliberate. We're still looking for the answer ourselves.
The recombination trap
The lazy version of the skeptic's argument goes: AI merely recombines existing material, while humans invent from nothing. This version collapses on contact with the history of human creativity.
Gutenberg's press combined a wine press with a coin punch. Cubism fused African sculpture with Cézanne's geometry. Rock and roll spliced gospel, blues, and country. Brian Eno spent a career arguing that genius is mostly "scenius" — scenes of people remixing each other until something catches. Cognitive science backs him up: study after study finds human creativity operating through recombination, analogy, and transfer across domains. Invention from nothing is a romantic story we tell after the fact.
So recombination disqualifies nobody. If a model remixing its training data means it can't create, then most of the human creative canon can't have been created either. The genuine difference between Gutenberg and a generative model lies somewhere else.
What the machine might lack
Run the thought experiment further and three candidates emerge.
The first is the genre problem. Models generate new points inside an existing possibility space. Humans occasionally add a new axis to the space itself. Kandinsky painted abstractly when abstraction wasn't a category anyone recognized. Schoenberg abandoned tonality when tonality defined what music was. Ask which genre — which new kind of thing — an AI has originated, and the answer sheet stays empty. There is one famous complication: AlphaGo's Move 37, a move so far outside professional consensus that experts initially scored it a mistake, before recognizing it as a discovery that changed how humans play the game. That was a real leap beyond the training distribution. It also happened inside a closed system with a perfect scoring function. Go tells you, instantly and unambiguously, whether a move worked. Culture does no such thing.
Which points to the second candidate: the missing reward signal. This may be the heart of the matter. Invention involves two acts, and only one of them is generative. The first act produces the new thing. The second act recognizes that it matters — before the market, the critics, or the culture agree. Kandinsky had no external signal telling him abstraction was valuable. He had conviction, generated by stakes: a career to lose, a decade of frustration, a life that made the work urgent. A model has none of this. Nothing is at risk in its output. It can generate the surprising artifact all day; it cannot want that artifact against consensus, and wanting it against consensus may be the load-bearing part. New things arrive ugly. Someone has to defend them while they're still ugly.
The third candidate is the question itself. Every impressive AI output in circulation began with a human deciding which question was worth asking. The model supplies answers at industrial scale; the itch that provokes a genuinely new question comes from lived friction — years inside a broken industry, a complaint that won't stop nagging, a pattern noticed and refused to be unseen. Models wait. Humans wonder. As of 2026, that division holds.
The honest wrinkle
Full disclosure requires admitting how murky the inside view is — for both parties.
A model cannot reliably tell whether its output is creation or very high-dimensional retrieval. Humans suffer the same blindness. Cryptomnesia — unconsciously reproducing something you once absorbed and believing it's yours — is thoroughly documented; George Harrison lost a copyright case over "My Sweet Lord" for exactly this. When a human unknowingly reproduces, we call it an accident. When a model does, we call it the mechanism. The asymmetry in judgment says as much about us as about the machines.
Sit with that asymmetry for a moment, because it cuts in an unexpected direction. If we can't reliably distinguish creation from retrieval in our own minds, then our confidence that we do invent — that the feeling of originality corresponds to something real — rests on shakier ground than the debate usually admits. The question "can AI invent?" keeps turning into a mirror: it forces a definition of invention precise enough to survive scrutiny, and most of our definitions don't.
Three tests, for the reader still wondering
Rather than a conclusion, here are the markers we'd watch — the observations that would move our own answer one way or the other.
The genre test. The day an AI originates a new category — a form of music, a school of design, a kind of writing that didn't exist as a recognizable type before — the possibility-space argument dies. Watch for it. Its continued absence is evidence too, and the longer generation stays cheap and abundant while the absence holds, the more that absence means.
The ugly-duckling test. Invention's hardest moment comes before validation: championing the thing that looks wrong. If a model, unprompted, ever insists that its strangest output is its most important one — and turns out to be right — something fundamental will have shifted. So far, models defer. They optimize toward what the prompt, and behind it the culture, already wants.
The stakes test. This one runs deepest. Kandinsky's conviction came from having something to lose. Perhaps conviction simply is what risk feels like from the inside. If so, invention will remain human until machines have genuine stakes — and it's worth wondering what it would even mean to give them any, and whether we'd want to.
So — can AI invent?
Our honest position: AI can produce the artifact of invention. Whether producing the artifact is the whole of inventing, or the lesser half of it, depends on where significance-recognition, conviction, and stakes fit into your definition — and the machines have exposed how few of us ever had a rigorous definition at all.
That might be the real gift of this strange moment. For centuries, invention could be pointed at without being explained; the question answered itself because only humans were in the running. Now something else is in the running, and the question has teeth again. We find ourselves examining our own creative acts with a suspicion we used to reserve for machines — and finding the inspection genuinely difficult.
We don't have the answer. We're increasingly convinced the search for it will teach us more about human minds than about artificial ones. If you get there before we do, we'd genuinely like to hear about it.