Does “writing is thinking” still hold when AI can do most of the writing?
If writing has long been one of the main vehicles of thought, what happens when AI starts carrying much of that cognitive labour?
I have been thinking about this after reading Richard Menary’s paper Writing as thinking.
His argument is powerful. Writing is part of thinking. We do not simply think first and write later. We think through the act of writing itself.
Drafting, revising, deleting, moving sentences around, rereading a paragraph, seeing a gap in an argument, finding a clearer way to say something. All of that is cognitive labour.
But generative AI changes the conditions around this argument.
Before AI, writing carried much of the heavy lifting. It was one of the main ways students externalized thought, struggled with ideas, organized meaning, and developed judgement.
Now AI can produce the paragraph, polish the sentence, restructure the argument, summarize the reading, and generate the reflection.
So the question becomes serious: what happens to the thinking that used to happen through writing?
I still believe writing matters deeply. But I also know people, including people close to me, for whom writing creates anxiety. They think better through sketching, diagramming, drawing, speaking, mapping, or building.
So maybe the old mantra “writing is thinking” belonged to an age when writing carried far too much of the burden.
This also explains part of our current assessment problem. For years, education has leaned heavily on the written product as the main evidence of learning.
Generative AI has disrupted that assumption.
A polished text can still tell us something, but it can no longer carry learning assurance by itself. We need process evidence, oral explanation, drafts, diagrams, annotations, design choices, and moments where students show how their thinking developed.
Literacy is a situated practice. Pre-AI literacy and post-AI literacy belong to different conditions. Now what we need to think about is whether our students can think across tools, modes, contexts, and constraints?
O artigo completo é de acesso pago. O que fica acima é uma síntese.