"The old man the boat."
Read it again. Your eyes just did something a machine's do not: they hit "the boat," recognized that "man" was not a noun but a verb, and jumped backward to repair the wreckage. Old people are manning a boat. The sentence was grammatical all along. You were the one who got it wrong, and then — this is the part that matters — you knew you got it wrong, and you went back.
More than 400 large language models reportedly sailed straight through that sentence without a flinch, per the NYU-UMass study. This essay lives inside that flinch.
Last week, in "The Elimination of Thinking," I treated AI as a solvent — the thing that dissolves the persistence a reader needs, the degradation evidence taken as settled. This week I want to grant the opposing case its full hearing, because a serious argument has been made that the same tool is a forge, not a solvent. And I want to show why neither side can win the fight as they've drawn it.
The debate arrives in three simultaneous verdicts, and they never quite touch.
The first is the forge. James, the CEO of Mercury Technology Solutions, argues that the highest-value use of AI is not coding or answers or "emotional value" but the shaping of how a person thinks. His mechanism is precise: stop taking answers, take explanations, then ask two questions until you can no longer chat. Why did you think that? Why do you think I thought that?
The second is the crutch. It comes in a dozen registers — the columnist mourning that Waze erased every Malaysian's sense of direction, the vice president who typed "Are you working?" to an AI before she could locate her own train of thought, the MIT Media Lab study reportedly finding 83.3% of participants unable to recall specific passages from the text they had just "written." A workaround, one of them writes, isn't a treatment; and a crutch, used long enough, changes your gait.
The third verdict is a shrug. Paolo Granata, having synthesized hundreds of peer-reviewed studies, concludes there are simply two tales — atrophy and augmentation — and the difference is "the mode of engagement." John Nosta goes further: maybe we can't even recognize thought anymore, now that the artifact of thinking can come from an architecture that isn't a mind.
Here is what all three verdicts share, and why they talk past each other: not one of them defines the word "thinking." The forge assumes it knows. The crutch assumes it knows. The shrug admits it doesn't and stops there. Everyone is arguing about the fate of a thing no one has named.
The NYU-UMass study names it. Quietly, almost as an aside, it does the thing the whole debate refuses to do.
William Timkey, Brian Dillon, Tal Linzen and their collaborators tracked the eyes of 368 adult readers moving across carefully built sentences, then compared those movements against predictions from more than 400 language models, per the NYU-UMass study. The result splits cleanly in two.
In the first stage — recognizing a word out of a sequence of letters, gliding forward through easy text — humans and models converge. Next-word prediction explains it. The machine keeps our pace here, because here we, too, are essentially predicting the next word.
Then comes the difficulty. And at the exact moment of difficulty, the two systems diverge — not by a little, but categorically.
"The predictability of a word really doesn't even come close to explaining just how much time we spend on difficult words and garden-path sentences," Timkey says. The models were "drastically underpredicting the type of difficulty that we experience when reading."
Dillon describes the physical fact. When the eyes hit a word they can't integrate, one of two things happens. They linger — too briefly to notice — or "the eyes instinctively jump back." He calls that a regression.
Linzen gives the number that should reorganize the entire debate: "20% of our eye movements when reading are backward, and AI models cannot explain when we decide to do that."
20% of our eye movements when reading are backward, and AI models cannot explain when we decide to do that.
Sit with that. One-fifth of reading is not forward motion. It is the backward glance — the recognition that a first interpretation failed, and the return to rebuild it. This is where meaning is assembled. It is precisely what a system built entirely on next-word prediction cannot model, because prediction only ever faces forward.
So here is the reframing the study makes possible. Thinking — the human part, the comprehension part — is not the fluent forward glide. That part we share with the machine. Thinking is the regression: the stumble, the caught error, the going-back-down into one problem.
Now watch what happens to both camps when you hold them to this definition.
The forge works — but only for a specific reason, and not the one it claims. When James demands "Why did you think that? Why do you think I thought that?" until chat becomes impossible, what is he actually engineering? He says it himself: "You stay under one problem and go down." You don't get to leap topics. You compress the boundary conditions, narrower and narrower.
That is a regression. He has built a machine that forces the human to jump backward — to catch the failed interpretation and reconstruct it. His Forge Equation — "Unconstrained human + unconstrained AI = chat. Constrained human + constrained AI = thinking" — is really a description of manufactured stumbling. The forge forges nothing on its own. It provokes the human's own backward glance.
And the crutch? The crutch atrophies the reader for the mirror-image reason. It supplies fluent forward motion in the exact place where the reader was supposed to stumble.
The vice president couldn't finish a thought, so the AI finished it — forward, smooth, no regression. The student who lets a model glide past every hard sentence never hits the word that won't integrate, never jumps back, never assembles the meaning. The tool did the forward glide, which it shares with us, and quietly deleted the backward glance, which it does not.
This is the hard distinction the whole debate has been missing. AI structurally shares exactly one part of cognition with us: the forward next-word prediction, the 80%. It categorically cannot share the other part: the regression, the 20% where comprehension actually becomes human.
Which means the same tool can only ever do one of two things to that 20%. It can forge it — but only by provoking your own stumble, forcing you back down into the problem until you can no longer chat. Or it can erode it — the instant it hands you fluent forward motion in place of your own stumble.
The forge and the crutch are not two tools. They are one tool, distinguished entirely by whether it makes you jump backward or spares you the jump.
This reopens the thread I left live last week. I asked whether persistence-eroding friction could be engineered back into a tool students freely choose. The NYU study recasts the question with a sharper word. The friction we need to preserve is not vague difficulty. It is the regression — the specific, involuntary backward jump at the moment a first reading fails.
That is what the ASCD educator gropes toward with "productive struggle" and can't quite name. Not all struggle is regression. Punching cards at 2 a.m. is forward friction, wasted motion. The productive kind is the caught error, the return, the rebuild — the thing eye-tracking can actually see.
So the forge-versus-crutch debate is unanswerable as posed, because both camps thought they were fighting about a tool. They were never defining the mind the tool acts on.
The machine sailed straight through "the old man the boat." You did not. That flinch — the jump back to the word that broke, the reconstruction of a meaning you first got wrong — is the whole of what is at stake.
A forge is any tool that makes you flinch. A crutch is any tool that flinches for you. There is no third option, because there is no machine that can lend you the one motion that only ever moves backward.