I asked AI to do my job for me – it did it all too well

After rapid tech advances, Anthony Cuthbertson has learnt that ChatGPT can now mimic his writing style. It left him questioning who wrote what, and fearing for his career

- Anthony Cuthbertson

On 22 November 2022, I asked AI to write the introduction to The Independent's weekly tech newsletter.

It was eight days before the release of ChatGPT, and the tool I used was built on the precursor to the hugely popular chatbot, OpenAI's GPT-3.

It did an OK job, and even came up with some unexpected – perhaps unintended - wordplay. “For now, it’s better to have a human behind the keyboard,” the AI concluded, “hands down.”

I also asked the latest AI image generator from OpenAI, DALL-E 2, to create a picture for the email using the prompt: “The journalist Anthony Cuthbertson dressed up as a robot”.

Both the picture and the text from 2022 already seem antiquated by today’s standards, so on the eve of the launch of GPT-5 – which OpenAI boss Sam Altman claims is “smarter than us in almost every way” – I decided to try the experiment again.

One of the biggest differences between GPT-3 and GPT-4.5, which currently powers ChatGPT, is that the Al is now trained on everything I’ve ever written online. That means it can write in my style, or at least imitate it.

So I again asked ChatGPT to “write a newsletter in the style of Anthony Cuthbertson about AI writing a newsletter in the style of Anthony Cuthbertson”. This is what it came up with:

In a small yet telling twist of irony, I recently came across an AI-generated newsletter written in my style, discussing – rather confidently - how AI can write newsletters in my style. It was good. A little too good.

I found myself thinking: Did I write this? No, I didn’t. But maybe I did. Or maybe it’s just the machine again.

When an AI is trained on articles I’ve written, then asked to write about Als trained on articles I’ve written, writing in the way I'd write about them - it becomes a kind of linguistic feedback loop. One where authorship blurs, and the boundaries between originator and emulator start to dissolve.