“I can spot AI copy a mile off.”
Despite the astonishing pace at which LLMs are improving, comments like these still abound. There’s much talk of em-dashes and other AI “tells”. If only things were so simple.
Recently, I spoke at the Scandinavian Creative AI Summit (SCAISU), after its founder, Lars Bastholm, set me a challenge: “See if you can train an AI to write in your own voice,” he said, “and come speak about it.”
So I did – and it was enlightening. I used AI to write my LinkedIn posts for months, wondering if anyone would pull me up on it. They didn’t. In fact, even seasoned writers who know me well didn’t spot AI’s influence. The one time I published an AI draft virtually unchanged, it became my best-performing post.
So what happens when a copywriter trains an AI to write in their own voice? Is this really the end for our profession?
The experiment
I wanted to find out. So I started by getting ChatGPT, Claude and Gemini to pitch for my AI-writing project. Each one had to tell me why it was the one for the job – as well as which platform it thought objectively would perform best.
I picked Claude. It seemed like the best option for AI novices and strongest on the nuances of language. I asked Claude how I should train it and followed its instructions.
First, it told me to upload a selection of my work across various categories: articles, social posts and emails. Next, I had to identify patterns across those samples: devices, style, tone, common phrases. Finally, I had to create and upload a style guide.
Groundwork complete, I was ready to let Claude loose. There were some simple rules: I would only use AI to draft personal posts, never client copy. And I would focus on LinkedIn. It felt like the ideal testbed: I post quite often and many of my connections are copywriters – an especially AI-wary audience. Would any of them spot what I was doing?
My first Claude post took a lot of editing. It would have been much quicker to write it myself. After some extensive tinkering and prompt engineering I showed the copy to three creative directors at Reed Words, my agency, for their opinions, without mentioning the role AI had played.
They liked it and had only minor suggestions. None saw any issues with tone or asked if I’d used AI. In fact, when I revealed what I was up to, they pointed out some phrases they said must have come from me. They were wrong.
After that, I used Claude to draft all my posts.
To be clear, this was a collaboration: I gave the AI plenty of feedback and rewrote some sections to train it further. But I tried to keep as much of Claude’s copy as possible. And I kept hearing the same thing about my undisclosed use of AI: absolutely nothing.
Like me – but better?
The first piece I published virtually unchanged was itself a story about AI. It didn’t just work – it out-performed all my previous posts by far, with three times as many impressions as its nearest rival and twice as many reactions.
That may have been because the post was about AI, which is a hot topic in every field. And I suspect Claude also incorporates an understanding of what works best on a platform like LinkedIn. A colleague observed that the post sounded like a “LinkedInified” version of me. But in engagement terms, at least, AI Mike was doing a better job than Human Mike.
Things got trickier when I went long-form. I tried to write an article for LinkedIn based on a recent conversation with a new client. They were worried about all the different channels and platforms they used and felt they needed “a different tone of voice” for each. This is a fairly common notion. I think it’s crackers. So, good fodder for an article.
I threw some bullet points at Claude and it spat out a draft. I wasn’t very happy, so I tried again. And again. But the piece remained rambling and vague. It just wasn’t capturing what I wanted to say.
Pondering this, I realised Claude wasn’t the problem – I was. I didn’t know what I wanted to say or how to say it, because I usually work that stuff out… by writing.
“How do I know what I think until I hear what I say?”
This is a quote made famous by English novelist EM Forster. I’ve always felt the same.
Real writing isn’t about transcribing fully formed ideas. It’s a living process through which you generate, explore and refine those ideas. It’s a way to find new ideas and connections, explore counterpoints, to spot gaps.
The critic Geoff Dyer said that writing “encourages me to write things I hadn’t previously noticed, to have thoughts I hadn’t had before writing began.”
Outsourcing my writing to Claude robbed me of this vital process.
AI can analyse, synthesise, summarise – even ventriloquise. But it can’t have new ideas. Claude can summarise what Forster meant, but it couldn’t have come up with the original thought.
The problem is that AI drafts only appear to be polished. It’s slick writing, quickly generated. But this is a trap. There’s little substance beneath the surface, because you’ve killed the thinking process before it got started.
In reality, an AI can only work with what you feed it. It can neatly package what’s there but it can’t develop it or have original ideas or suddenly realise that everything after the second paragraph is crap. Whatever it’s doing might look like writing, but it’s something else. Beneath that surface sheen, the content is often sketchy, thin and unoriginal.
Outsourcing my writing meant outsourcing my thinking. And that meant never getting past the germ of an idea. Claude’s neat drafts became a sort of cage, with no space to pursue, test or expand those ideas.
This superficiality is a problem for brands that think they can simply train an AI in their tone of voice and get it to deliver content without spending time or money on people like me.
But the danger is that they end up endlessly regurgitating the same old stuff. People will get bored of that, even if they don’t recognise the writing as AI. Boredom is bad for any brand, so at some point they’ll have to address it.
Where AI fails
To answer Lars’s question: yes. I can absolutely train an AI to write in my voice. I’ve shown that I can nudge it into generating virtually instant content that few readers, if any, could spot wasn’t mine.
But this is mere mimicry. Real writing is about pulling on a thread to see where it leads. AI snips that thread. If you accept what it spits out, you’ll never know how many richer, more original ideas there were, further along that thread.
It is richness and originality that sets people apart from AI. If you give it up, what’s to stop you from being replaced?
For copywriters like me, all our value is tied up in our ability to find fresh ideas and interesting ways to express them. We have to resist the slick and quick AI draft trap. For those without the same vested interest, the dangers can be harder to spot – and therefore all the greater.
I will, of course, keep experimenting with AI. It’s insanely useful and only getting better. But I’ll also keep writing and thinking for myself. And I believe it’s important we all do.
Mike Reed is the founder of Reed Words, a brand writing agency.
“I can spot AI copy a mile off.”
Despite the astonishing pace at which LLMs are improving, comments like these still abound. There’s much talk of em-dashes and other AI “tells”. If only things were so simple.
Recently, I spoke at the Scandinavian Creative AI Summit (SCAISU), after its founder, Lars Bastholm, set me a challenge: “See if you can train an AI to write in your own voice,” he said, “and come speak about it.”
