An editor’s guide to spotting AI writing

Share this article

How do you know this wasn’t written by an AI? If you’re reading CapX, you are presumably interested in free speech, critical thinking and originality. So you should care that what you’re reading was written by a person with a mind.

But public life is becoming colonised with ‘copy’ generated by AI. It’s become so commonplace that a journalist at Marie Claire didn’t even bother to remove AI edits from a published article. A Canadian MP revealed that a chatbot had helpfully made his oratory ‘read more like a legislative speech’ by reading the prompt aloud in parliament.

As an editor I receive op-ed submissions every day from reputable PR agencies, campaigners, law firms, blue chip companies and even writers that have been generated by AI. I do not publish this content for the very simple reason that if no one has been bothered to work on it no one should be expected to read it. I don’t claim to be a writer of genius, but I have spent my entire career polishing other people’s words. I think I’ve become pretty attuned to what reads well and what doesn’t.

But not everyone has that professional advantage. And the sheer volume of this stuff indicates that too many believe that AI is impossible to distinguish from human writing. This is not true.

If the supply of ideas is becoming near infinite thanks to AI, we are in a buyers’ market. Consumers can afford to – and must – be discerning. So here’s a few ways to tell if you’re reading a robot based on real op-eds that I have rejected.

Rhythm

AI loves short sentences. Four-word sentences are great – I wrote one just now. But AI over-uses them. If there are two four-word sentences in a row that is an unlikely coincidence. Multiple paragraphs in the same piece with the exact same sentence lengths are strong indicators of AI.

Here’s an example from a recent submission I rejected: ‘Because here’s what I believe: in moments like this, when the old certainties wobble, creativity isn’t a luxury. It’s a necessity. Imagination is how we rebuild.’ Ironically, given the subject matter, no one with an imagination would write this formulaically.

Rhythm is difficult to codify but you can hear it as you read. I could tell this submission was AI just from the paragraph lengths, which are so clearly unnatural.

By contrast, these are the paragraph lengths of the last column the late Rod Liddle wrote for The Spectator:

Human beings need more than a sentence to finish a thought. If in doubt, read it aloud and ask if it sounds like normal human speech.

Weird constructions

Because it relies on aggregation, AI is bad at having original opinions. So it uses a lot of ‘on the one hand on the other’ or ‘not only but also’ constructions. It’s often strangely passive, as in this sentence: ‘The rules and how they’re applied vary from platform to platform, and sometimes the calls made are the wrong ones.’

It will also use very similar constructions far too close together. The tricolon is an obvious one. But here’s an example of two similarly ambivalent constructions that occurred just a few paragraphs from each other in a submission that I rejected:

Gary Stevenson has correctly identified Britain’s frustration. But frustration is not an economic strategy.

And:

Britain doesn’t need a politics of envy. It needs a politics of opportunity.

Good opinion writers are both more emphatic and more uneven than this.

Sense

AI will repeat and argue against itself. A recent submission that I rejected included: ‘These aren’t simply signs of wealth being unevenly distributed. They’re examples of an economy that has stopped generating enough of it in the first place.’

This is not only a telling ‘not just this, but this’ AI construction, it also doesn’t quite hang together. It’s not a real contradiction, even though it is framed as one. And later in the piece the same assertion is repeated almost verbatim: ‘The bigger question is why Britain has become so poor at creating wealth in the first place.’

Epistemology

I’m sure, as you’ve been reading this, you’ve noticed that my own writing is riddled with the same infelicities of which I’m accusing AI. That’s the point. We are in a co-dependent relationship with the robots. They are parasites on our thoughts but we are increasingly outsourcing thinking to them.

Accordingly, I asked the AI how to spot AI. ChatGPT told me that it is too fond of transition phrases like ‘that said’ and ‘ultimately’. But, more interestingly, it admits that it can’t know how it knows something. So a human being might use an anecdote or a personal experience to illustrate a point, AI can only speak in banal generalisations.

Gemini told me to look out for ‘low burstiness’ – not a phrase I’ve ever heard a human use – and ‘uncanny, hollow anecdotes’, which sounds like a conversation with Keir Starmer.

Provenance

Does the author work in tech? Then it was probably written by AI. For very obvious reasons, people in the tech sector are among the most reliant on AI. It’s frustrating because they should be more aware than most of the limits of this technology. They should also know that you can use AI to detect AI (I use Pangram).

Moreover, it’s so self-defeating. One of the biggest barriers to AI adoption is perception – and people hate AI. Ex Google CEO Eric Schmidt was booed during a commencement speech at the University of Arizona for mentioning it. In Britain six in 10 people believe AI will eliminate more jobs than it creates. This is hardly surprising. Silicon Valley actively told us that AI would take graduate jobs. Sam Altman has had to walk back from his prediction that AI would replace 40% of jobs because it was so reputationally damaging.

If tech bros want us to use their products then they should be showing us how they can make our lives better. CapX readers will have every confidence that the AI industrial revolution will transform society’s fortunes for the better, but we must not let it replace our humanity. And AI writing is inhuman: lazy, incurious and cautious. It trends towards the lowest common denominator. In that, it makes life worse.

If you really want to know whether what you’re reading was written by a robot then the best thing to do is to read real authors. Read Jane Austen, read Martin Amis, read Alan Hollinghurst, read Robert Colvile, read Rod Liddle – and read CapX.

Share this article