AI writing sounds generic because a language model produces the most probable next words, and the most probable words are the ones everyone uses. Each draft is individually fine and collectively the same, and the studies that have measured it find exactly that: people using AI write pieces that score a little higher on their own and a lot more alike as a group. The fix is giving the model something specific to say, and then checking that it said it.
What the research found
In 2024, Anil Doshi and Oliver Hauser published a study in Science Advances in which around 300 writers produced short stories, some unaided, some with one or more story ideas from GPT-4. Judges rated the AI-assisted stories more creative on average (about 8% higher for novelty).
Measured against each other, though, the stories written with AI help were markedly more similar. Access to a single AI idea raised story-to-story similarity by about 11%. The authors called it a social dilemma: each writer does slightly better, and the pool of writing gets narrower. (University of Exeter, PsyPost)
Vishakh Padmakumar and He He got the same shape of result with essays, published at ICLR 2024. Writers helped by an instruction-tuned model produced essays that were measurably more alike than unaided writers’ essays, and the sameness came from the text the model contributed, which made up about 35% of the final drafts. (ICLR 2024, Journal du Net)
A 2025 study by Emily Wenger and Yoed Kenett ran 22 commercial models and 102 people through standard creativity tests. The models matched the average person, and they agreed with each other far more than people did, with models from the same company agreeing most. “If you want your content to be truly unique, probably shy away from using an AI chatbot to generate it,” Wenger said. (arXiv, PsyPost)
And it holds at work. A Harvard Business School field experiment gave GPT-4 to 758 BCG consultants. Their ideas were rated higher in quality and were “less variable and therefore at risk of being more redundant”, in the authors’ words. (HBS working paper, Wharton follow-up)
Why it happens
A model learns from an enormous amount of writing and, when it generates, leans toward what’s common in that writing: the common phrasing, the common structure, the common opinions, so what it writes is an average of what it read, with excellent grammar. That’s a feature when you want a competent first draft of something routine, and a problem when the whole point of the piece is to sound like one particular company and nobody else.
Two things make it worse in a business. Prompts are generic too: “write a LinkedIn post about our new service” contains nothing the model couldn’t have guessed. And the person signing it off rewards safety, so the specific sentence gets questioned and the smooth one gets through. AI didn’t invent that habit, but it’s made it much cheaper to indulge.
What people get wrong
The usual response is to ask for more personality: “make it punchier”, “add some humour”. The model obliges with the most probable version of punchy, which is its own cliché, so you end up with the same generic draft at a higher volume.
The second mistake is treating the voice as the problem. A distinctive voice saying nothing in particular reads as a brand trying hard. The writing that stands out is the writing with a fact, a figure, a client’s name or a decision in it that no competitor could have written, and no model can supply that from a prompt.
The third is assuming the model followed the brief. It was asked to. Whether it did is a separate question, and nobody’s asking it.
What to do instead
Give the model specifics it can’t invent: your approved facts, the words you use and don’t, three pieces of writing that sound like you. Ask for one idea rather than five. Then read the draft back against what you asked for, because the gap between the two is where the generic creeps in.
How BeigeCheck does it
BeigeCheck gives Claude, ChatGPT and Microsoft 365 Copilot your approved facts and voice before they write, so the model starts from what’s true about you rather than what’s probable about everyone. Each person’s own voice, built only from writing they’ve done themselves, keeps colleagues from sounding like the same person.
Then the Beige-ometer reads the draft and scores how authentic it reads and how well it fits the voice, quoting the sentences that pull it toward the middle, with a fix for each. It checks every claim against your facts. Nothing changes until you say so, and a score that moves tells you whether the fix worked.