The question is not whether AI writes, but whether humans still think

Language is one of journalism’s favourite summer subjects.

When politics slows and offices empty, newspapers turn to changing narratives, endangered punctuation and disputes over grammar.

This summer, the subject is AI and writing. A New York Times essay pleaded with readers never to write with AI, while The Economist and Wired examined how machine-generated prose might be detected. Seemingly to appease the EU AI Act’s Code of Practice on Transparency on AI-generated content, Anthropic has announced that new Claude models will mark AI-generated content to make it ‘machine readable’.

The questions are increasingly familiar: Was this written by AI? Which expressions reveal the machine? Can we still distinguish human writing from synthetic text?

These are understandable questions. But they are misleading for two reasons.

First, AI detectors are inherently unreliable. Humans and machines draw from the same textual inheritance. We learn patterns of expression through education and reading; large language models acquire them through machine learning. Balanced sentences, rhetorical triplets and historical analogies existed long before ChatGPT, yet they are increasingly treated as fingerprints of artificial prose.

I tested this by submitting an article I wrote in 1996 to an AI-detection platform. It concluded, with impressive numerical confidence, that more than half had been generated by AI. Among the suspicious passages was a rhythmic triplet describing how technology affects diplomacy through geopolitics, the subjects diplomats negotiate and the tools they use.

The machine was accusing a 30-year-old text of being written by technology that did not yet exist.

We may soon be writing for detectors rather than for readers.

The second problem is more important. By obsessing over whether AI generated a sentence, we avoid asking whether the sentence expresses any genuine thought.

AI’s ability to produce polished prose is not itself the main danger. The danger arises when we delegate questioning, interpretation and judgement to the machine. A grammatically imperfect paragraph that reflects struggle, doubt and discovery may be more intellectually valuable than a flawless essay produced without meaningful engagement.

This is particularly evident in education. When a machine can generate a student essay in seconds, teachers understandably turn towards detection: Was AI used? Can the student prove authorship?

But these questions address symptoms rather than causes.

If the final essay matters more than the intellectual process that produced it, AI becomes a rational shortcut. Students may simply be responding rationally to a system that rewards polished products, accumulated credits and compliance with formal requirements more than curiosity, reconsideration and judgement.

Here lies a deeper mismatch. AI is fast; cognition is slow. A machine can produce an answer in seconds. Understanding takes time. So do doubt, revision and argument. In an age obsessed with efficiency, some apparently inefficient activities—debate, mental calculation, memorisation and apprenticeship—remain valuable precisely because they make our brains work.

The answer, therefore, is not to ban AI. It is to redesign writing and learning around the cognitive capacities we want to preserve.

Three approaches could help.

First, use AI as an intellectual sparring partner, not a ghostwriter. Ask it to challenge an argument, identify a missing perspective, formulate a counterexample or expose a weak assumption. AI should trigger more thinking, not less.

Second, value the process, not only the product. Students—and why not journalists, officials and diplomats?—should be able to explain how an argument developed, which assumptions changed, why evidence was accepted and why suggestions were rejected. The better question is not “Did you use AI?” but “What did you do with it, and what thinking remained yours?”

Third, ground AI-assisted writing in identifiable knowledge sources. AI prose does not emerge from thin air or from some miraculous machine intelligence. It is rooted in an existing corpus of human knowledge and data. Grounding makes these intellectual origins visible. It also brings important side benefits: biases can be traced back to biased sources, claims can be checked against evidence, and AI systems have less room to hallucinate.

The anxiety surrounding AI and writing is not entirely new. In Plato’s Phaedrus, the Egyptian god Theuth presents writing to King Thamus as an invention that will improve memory and wisdom. The king warns that it may instead weaken memory and create the appearance of wisdom without genuine understanding.

Writing ultimately transformed civilisation. It preserved knowledge, enabled science, sustained religions and expanded human imagination. But King Thamus’s question never disappeared: does externalising thought necessarily make us wiser?

AI makes that question urgent again.

There is one final irony. I submitted this op-ed to Pangram, an AI detector, which judged with high confidence that 100 per cent of it was written by AI. Yet this op-ed distils a much longer text that the same platform judged to be only 10 per cent AI-written.

Somehow, in shortening my own argument, I became less human.

That absurdity is precisely the point.

The future of writing will not be secured by hunting for em dashes or preserving grammatical imperfections as proof of human authenticity. What matters is not whether a machine helped arrange the words. What matters is whether a human being continues to question, interpret, doubt and judge.

And if these lines have prompted some thinking of your own, they have done their job.

The medium will change. Our responsibility for thinking must not.

This op-ed is a summary of a longer text analysis of AI in writing available here.

Jovan Kurbalija is the is the Executive Director of Diplo Foundation and Head of the Geneva Internet Platform (GIP).