Guides

Why so much AI copy sounds the same

By Rory Wilmer, InkByte · 6 min read · 24 October 2026

Left to its own devices, AI writes the most likely sentence, and the most likely sentence doesn't belong to anyone.

I've been reading brand captions for a living for about twenty years, so I suppose it's an occupational hazard that I can't scroll a feed without judging the copy. Lately, though, I keep running into the same kind of caption. A bank wants to be part of your journey, a gin wants to elevate your weekend, and the gym down the road is thrilled to announce something it could have just told you. On a fair amount of everyday posting, if you covered up the logo I'd struggle to say who wrote it. Anyone who grew up on Invasion of the Body Snatchers will recognise the feeling of everything looking and sounding perfectly normal while something underneath has quietly gone missing.

For most of my career, running social media for international and global brands, the hardest part of the job was always the voice. Getting a brand to sound like itself in a reply to an unhappy customer and in a launch, without becoming a different personality each time, takes years and a lot of arguing in rooms. In my book I wrote that attitude is the most important and relevant thing you or any brand should portray on social media, and I haven't changed my mind on that. What's changed is how easily that attitude can now be thrown away.

Why the machine drifts to the middle

A language model is a very sophisticated guessing machine. It looks at what you've given it and predicts the most likely next word, then the next, and so on. Ask it for an Instagram post for a coffee roaster and give it nothing else to go on, and it will write you the most likely Instagram post for a coffee roaster, which is more or less the average of every coffee post it has ever read. With nothing specific in front of it, the average is simply the safest bet it can make.

I'd have left this as the kind of hunch you moan about over a pint, until I read a study by Anil Doshi at the UCL School of Management and Oliver Hauser at the University of Exeter, published in Science Advances in 2024. They asked 293 people to write a short story of exactly eight sentences. Some of them could ask GPT-4 for story ideas before they started, while the rest had to manage on their own. Readers rated the stories written with AI ideas as more creative and better written, and the biggest improvement was among the writers who weren't especially creative to begin with.

What interested me more was what happened to the stories as a group. The AI-assisted stories were more similar to each other than the ones people wrote alone. The authors describe this as a social dilemma, where each writer is better off individually but collectively they produce a narrower range of work. They're careful about the limits: their writers weren't professionals, the stories were tiny and nobody could go back and forth with the model, so it proves nothing about your client's next campaign. Still, it fits with something I've been noticing in everyday posting, where a lot of individual posts look a little more polished than they used to and a good number of them are starting to sound alike.

In the book I filed this sort of thing under Sloth, which for me was always about being lazy with the truth rather than lazy with effort. Facts should be sacred. Generic AI copy is a modern version of the same sin, with the thinking handed over to whichever word is most probable. It gets worse when you push a model to be specific without giving it anything specific to work with, because it has a habit of filling the gaps itself with a heritage or a statistic that sounds plausible and that nobody has ever checked. I've watched enough brands get burnt on social for careless words to be anything but relaxed about that.

I'm not anti-AI, and I use AI tools in my own work. In my experience the problem sits with what we give the model, which is usually almost nothing about the brand it's been asked to write for. Most tone of voice documents I've come across describe the brand as something like friendly, confident and human, which is what plenty of their competitors' documents say too, so the model reaches straight back for its average idea of friendly. What it needs is material it can actually copy: the phrases the brand really uses, the words it would never use, examples of it at its best and some way of checking a draft before it goes anywhere.

Where Brand Brain comes in

That's the thinking behind Brand Brain, which is what we make at InkByte. A Client Brand Brain is a done-for-you voice pack for one brand, made up of a PDF voice guide and ready-to-paste prompts for ChatGPT and Claude. It's built from the brand's own public copy and the three to five on-brand samples you send with a short brief, with no calls. AI tools help us build and test each pack, and a person checks every fact and prompt before it reaches you. The rule I care about most is that we don't invent anything. Every fact and approved phrase traces back to the client's public material, your samples or your brief, and anything we can't source is marked [TO CONFIRM] for you to check rather than guessed at. One brand costs £99.

Our free public sample is built for Hobcart Coffee Roasters, a fictional Yorkshire roaster we invented. This is the default AI draft for a new coffee:

And this is the version written with the Brand Brain in place:

The second one tells you what the coffee tastes like in words you'd actually use and gives you a brewing tip with a number in it, and both details came from the brief rather than the model's imagination.

Most brands have access to the same tools now, and a model given nothing of a brand's own will tend to drift towards the same middle. The brands I'd want to follow are the ones that give the machine something only they have. If you'd like to see what that looks like, the full Hobcart sample is free here on the site. Have a read before you write your next brief.