One of the major concerns surrounding generative AI in education relates to writing, or rather, to the assessment of writing. If pupils and students can have their texts written by ChatGPT or another AI, how can we still be sure whether they have actually thought for themselves? This has already led to all kinds of responses: more exams under controlled conditions, software that tracks the writing process, oral defences, and so on.

But an opinion piece in the New York Times, shared with me yesterday on social media, raises a provocative idea.  Joe Cruz asks in it: what if we assume too easily that writing is the best way to learn to think? Maybe we shouldn’t try to protect the traditional essay from AI at all costs. Instead, we could make much greater use of conversations, discussions, and oral exams. He also argues that these require students to explain, respond, argue, and defend their ideas on the spot. It is an interesting, if not necessarily original, idea. But one of the arguments Cruz uses to defend his case triggered my myth-busting nerve.

According to Cruz, the evidence that writing promotes thinking is “thin at best.” The idea that we discover what we think through writing is often attributed primarily to the writing-across-the-curriculum movement of the 1970s. In his piece, he refers to James Britton, who in 1980 “hypothesised” that we discover what we think while writing. According to Cruz, hypothesized is the crucial word, because convincing evidence never followed.

Well, I think it’s rather Cruz not knowing the research that well. There is, in fact, a body of research on what is known as writing to learn. Graham, Kiuhara, and MacKay, for example, describe 56 experimental and quasi-experimental studies examining writing as a means of learning in subjects including science, social studies, and mathematics. Across these different studies, they found a positive effect on learning of approximately 0.30 standard deviations.

To be clear, that’s not a huge effect. It also doesn’t mean that writing automatically leads to critical thinking. And it certainly doesn’t prove that writing is better than speaking, discussing, or defending ideas orally. But again, “thin at best” seems difficult to maintain in light of this literature.

There is, by the way, a broader body of research on the connection between writing and reading. Writing about what you read, for example, can strengthen reading comprehension. Writing instruction can also positively affect reading. I do think that there are several possible explanations for this. When you have to write something down, you have to select and organise information and connect it to what you already know. By doing so, you have to make explicit thoughts that might otherwise remain vague. Simply trying to formulate something coherently can reveal gaps in your understanding.

In other words, writing is not merely a way of putting a fully formed thought onto paper. The writing process itself can be part of the learning process.

But does that mean Cruz is entirely wrong? Here is your captain of Team Nuance again; I don’t think so. He does make an important point. Evidence that writing can support learning does not imply that writing is the best way to learn to think. Still, it does not imply that a written essay is always the best way to determine what someone does or does not understand.

In fact, I think he combines three different claims:

• Writing can support learning.
• Writing can support thinking.
• A written text is a good way to assess how well someone can think.

There is considerable support for the first two, more than you would think if you read Cruz’s opinion piece. The third is becoming much more complicated in the age of generative AI. Which brings us to a more fundamental question: what are we actually trying to achieve when we ask students to write? We must make a distinction between goal and means.

If the goal is for students to learn to write, then they obviously need to write. That remains true even in a world where AI can produce text with the click of a button. I think there is little doubt that the ability to formulate our own thoughts clearly remains an important skill.

Another good reason to ask students to write texts themselves is if the goal is to process particular material more deeply through writing. In that case, the writing process is not just a means of producing a final product. It is part of the learning activity.

But if the goal is to determine if a student understands something, I’m not so sure that a written text is the only possible form of evidence. I’m pretty sure a conversation, presentation, debate, or oral defence can also reveal other aspects of understanding. But these approaches may be less efficient than assessing a text, whether self-written or not. Probably a combination is even more interesting: writing, possibly with clearly defined uses of AI, followed by questions for which a student has not been able to prepare the perfect answer in advance.

I want to conclude with one distinction in the original opinion piece that I do think is valuable. Written text is not only a means of thinking and learning. It can also be a trace or evidence of thought. It’s easy to ask students to submit a text so it can be stored, reread, and assessed. For schools and universities, and for sure for the administrative powers in those organisations, that can be extremely useful. But administrative convenience and pedagogical value are not the same thing.

I think (or hope) AI will not force us to abandon writing. Maybe it rather forces us to think more carefully about why we ask someone to write in the first place.