Human in the Middle

• The myth of the solitary genius, transposed
• Vibe coding as the clearest example
• What the human is actually contributing
• Where this leaves the conversation

There’s a sentence that keeps turning up in every conversation about AI, and especially in the discourse around vibe coding: it can’t create anything new. It gets said with the flat certainty of someone stating a law of physics. On its own terms, it isn’t wrong. A model trained on the accumulated output of human thought is, by construction, recombining what already exists. It has no lived experience to draw a genuinely novel intuition from, no stake in the world that would give rise to a truly original impulse. Ask it to generate something from nothing and you get an average: competent, plausible, and strangely hollow.

But that framing already asks the wrong question. It judges the tool as though it operates alone, then acts surprised when a tool, used alone, produces something toolish. Nobody would judge a chisel by asking what it carves when nobody’s holding it.

The part that keeps getting lost, in the breathless takes and the dismissive ones alike, is the human in the middle. Not the human at the start, prompting. Not the human at the end, reviewing. The human in the loop: iterating, redirecting, noticing the moment where the output almost says something true, and pushing it the rest of the way. That’s not a footnote to the process. It’s the actual place where anything new happens at all.

§The myth of the solitary genius, transposed

Part of why this gets missed is that we’ve imported an old myth into a new argument. The romantic idea of the solitary genius, creation as a lone mind conjuring form from nothing, was always a flattering distortion, even applied to humans. Every writer edits against the memory of every other writer they’ve read. Every composer works inside and against a tradition. Origination has always been recombination plus judgement plus the willingness to keep going past the first plausible draft. We just didn’t notice the recombination, because it was happening inside a single skull, over years, invisibly.

Now the recombination is out in the open. You can watch it happen in real time, in a chat window, which makes it look mechanical in a way it never did when it was happening in someone’s head over decades. That visibility gets mistaken for a difference in kind. It isn’t one. What’s actually different is the speed of the loop, and speed changes what a person can afford to try.

§Vibe coding as the clearest example

Vibe coding is where this becomes obvious, because the loop is so tight you can feel it. You describe an intent, loosely. Something comes back: mostly right, sometimes surprising, occasionally a genuinely better idea than the one you had. You don’t just accept it. You push against it. This naming is wrong. This structure doesn’t match how the rest of the system thinks. This is technically fine but off. It regenerates. You push again. Somewhere in that back and forth, a design decision gets made that neither the model would have produced unprompted, nor you would have written unprompted. It exists only in the space between the two.

Critics of vibe coding aren’t wrong about the failure mode, and it’s worth being honest about how bad it can get. At least one recent assessment of vibe-coded applications found the large majority carrying a vulnerability traceable to the model simply not knowing what it didn’t know, and separate industry surveys point to a wide gap between how many developers use these tools daily and how many actually trust the code they produce. That’s not the full picture of the field, and figures like these vary depending on who ran the study and how vibe coding was defined. But it’s not anecdote either. It’s a real critique with real evidence behind it, and it deserves to be taken on its own terms rather than waved away.

But look closely at what those failures have in common. They cluster around the moments the human left the loop rather than stayed in it. A shipped vulnerability isn’t evidence that the loop is worthless. It’s evidence that nobody was actually in the loop when it mattered. A human rubber-stamping output, or never reading it at all, isn’t “human in the middle” in any meaningful sense. That’s a human at the end, and a fairly passive one. The whole argument stands or falls on whether the person is actually doing something in there: exercising taste, noticing wrongness, holding a standard the model has no access to on its own.

§What the human is actually contributing

If the model contributes breadth, the capacity to instantly generate the whole space of plausible next moves, the human contributes something the model structurally cannot: a stake in the outcome, and therefore a reason to prefer one plausible thing over another. Taste isn’t decoration on top of competence. It’s the thing that turns an enormous space of adequate options into the one that’s actually right for this piece, this protocol, this story. That takes caring about the result in a way that only comes from having something to lose if it’s wrong, or something to be proud of if it’s right.

There’s a more rigorous version of the sceptical claim worth naming directly, because it’s stronger than the offhand one. Some philosophers of science argue that knowledge only grows through conjecture: someone proposes a new explanation, then subjects it to criticism. A system built on induction from existing data can’t do that step, no matter how fluently it recombines what it’s seen. On that view, the model isn’t a creator of new explanations at all. It’s an instrument, however impressive. I don’t think this is wrong, exactly. I think it’s incomplete in the same place the looser version is. It’s still judging the instrument on its own. The conjecture, the actual new explanation, is still something a person has to propose and mean. What the loop changes is how many candidate conjectures a person can try and discard before landing on the one worth meaning.

This is, I think, the real shape of the “can AI create something new” question, once you stop asking it about the model by itself. The new thing was never going to come from the model alone, because the model has no position from which anything could count as new to it. And it was never going to come from the human alone either, not at the speed and breadth these tools now allow. The new thing comes from the iteration itself: the specific sequence of proposal, rejection, and refinement that only exists because both parties were in the loop. Neither one, on its own, could have arrived at that sequence.

§Where this leaves the conversation

None of this resolves the harder questions. What happens to craft when the loop gets fast enough that patience wears thin. Whether taste itself can be worn down by too much fluent assistance too early in someone’s development. What it means for labour and livelihoods when the loop makes one person’s output equivalent to a team’s. Those are worth their own honest reckoning, and “human in the middle” isn’t a phrase that wards them off.

But it does mean the flat claim, that AI can’t create anything new, is answering a question nobody serious is actually asking. The interesting question was never whether the model can do it alone. It’s what becomes possible when someone with genuine judgement sits inside the loop and refuses to let anything pass that doesn’t meet their own standard. That’s not a lesser form of creativity. It might be the oldest form there is, just running faster than it used to.