My students compared my writing against ChatGPT – and they all preferred the AI

A university experiment forced an uncomfortable reckoning with what writing is, says Sarfraz Manzoor (the real version)

Anxiety is an occupational hazard for the writer. I have been making a living from writing for three decades, and I cannot recall a time when it did not prompt anxiety.

When I started out, I worried about how to persuade commissioning editors to give me a chance to write. Later, I worried that I was being pigeonholed, worried about the chronic financial insecurity that comes with being a freelance writer, and worried that no one would be interested in letting me write what I wanted to write. These days, I still have all these worries, but they have been joined by a more recent anxiety: that I will be replaced by AI.

Alongside my writing career, I also teach non-fiction writing at University College London (UCL), where I try to persuade my students that aspiring to be a writer in the time of ChatGPT is not entirely insane. For all the gloomy warnings about the threat of artificial intelligence, I have tended to believe that it cannot yet emulate what makes great writing great, or even what makes distinctive writing distinctive.

I wanted to demonstrate this to my students, so I devised an experiment that would, I hoped, reveal the limitations of AI. I asked them to read two pieces. One was a column I had written, reflecting on how, in a year as unpredictable as 2026, I found comfort in things that were reliably predictable. The other was generated by ChatGPT, which I prompted to write 700 words making the same argument “in the style of Sarfraz Manzoor”.

I printed both pieces in the same font and format. The intention was simple: to show that, for all its seductive convenience, AI-produced writing was proficient but hollow. I wanted my students to compare the specificity, the wit and the rhythm of my writing with what AI produced. Any vaguely intelligent reader could, I assumed, see that my piece was superior in every way. That, at least, was the plan.

I gave the students time to read both pieces, and then asked for their comments. To my surprise, the majority told me the AI version was better. They said it was better argued, more clearly structured, more ambitious in scope, and, this was the real kick in the guts, a few even told me it was more personal than my own. That hurt: a piece generated by AI was more “personal” than one written by an actual person.

When I revealed that the first piece was mine and the second AI-generated, there was a pause as they processed this. Many looked genuinely shocked – and some were probably wondering if they could get a refund on the money they had spent on their degree. They then explained their reasoning. These were mostly international students – many of them Chinese – and the cultural references in my piece went over their heads. They had not heard of Pulp’s “Disco 2000”, Bonnie Blue or Nigel Farage, and so the very things I had included to make the piece resonate with readers were what made it harder for them to connect.

I admit to initially feeling deflated that the students preferred the AI version of me to the real thing. But when I probed why they preferred the second piece, I started to feel more hopeful. It became clear that they had been taught to value writing with clear signposting, that moved smoothly from one point to another, and that was shorn of personality or spikiness – writing that read more like an essay.

My piece, by contrast, assumed shared memory and shared cultural references, and it contained jokes. The students who were the most media-literate – those who told me they subscribed to newspapers – were the least persuaded by the AI version. One said it felt “cold”, others called it forgettable, and several noted that it was impossible to tell when it might have been written. By contrast, my piece contained a line that one student quoted and described as memorable, while another was persuaded to listen to “Disco 2000” for the first time.

In other words, the AI piece was competent but forgettable, while mine demanded more from the reader but offered greater rewards. The AI offered the illusion of profundity, packaged in bland banalities that could just about pass as insight. For some students, that was enough.

But, as I told them, while it may have fooled them, I strongly suspect it would not have fooled the editor who had commissioned the piece. Had I submitted that second version, I suspect I would have been told it needed more work, that it needed to feel more personal, and that it lacked an authentic authorial voice.

Students found the AI-written piece to be competent, but also forgettable (Getty Images/iStockphoto)

Will I one day be replaced by AI? It is possible, but not inevitable.

My homework for the rest of the term is to teach my students not only how to be better writers, but how to recognise and value writing that is truly human, rather than its paper-thin approximation.

The qualities I always believed were the hallmarks of good writing – voice, specificity – were not immediately appreciated by all my students. But that, perhaps, is precisely why they still matter.