AI Is Making Every Brand Sound the Same
AI Is Making Every Brand Sound the Same — Here’s What the Research Actually Shows
Every business using AI to write faster is quietly trading away the one thing a reader actually remembers: distinctiveness.
That is the uncomfortable finding sitting underneath a growing pile of academic research published between 2024 and 2026. Individually, AI-assisted writing tests better. People rate it as clearer, more polished, and more enjoyable than most unassisted drafts. But measured across a group — a market, an industry, a niche — the opposite happens. Content converges. Vocabulary widens while the ideas underneath it narrow. And almost nobody notices, because each individual piece looks better than what its author could have produced alone.
For anyone running a content site, an agency, or a brand, this is not an academic curiosity. It is a direct explanation for why so much AI-assisted marketing copy, blog content, and social posts already feel interchangeable — and it points to a specific, testable way to avoid the trap.
The study that started the conversation
In 2024, researchers Anil Doshi (University College London) and Oliver Hauser (University of Exeter) ran a controlled writing experiment and published it in Science Advances. Participants wrote short stories, some unaided and some with story ideas suggested by GPT-4, and independent evaluators — blind to which was which — scored the results.
The finding: AI-assisted stories were rated as more creative and better written than human-only stories, with the biggest gains going to the writers who started out least creative. But the AI-assisted stories were also measurably more similar to one another than the human-only stories were. The researchers described it as a social dilemma: each individual writer benefits, while the collective range of what gets produced shrinks.
That single trade-off — better on average, narrower as a group — turns out to be the most reproduced result in this entire body of research.
The gap that compounds as the pile grows
A separate research team, led by Kibum Moon alongside Adam Green and Kostadin Kushlev, tested the same question at a much larger scale. They analysed 2,200 college admissions essays across three studies, using a metric they called the diversity growth rate — essentially, how many genuinely new ideas each additional piece of writing contributes to a growing collection.
The finding: Every additional human-written essay added two to eight times more new ideas to the pool than every additional GPT-4-assisted essay. The gap did not shrink when the researchers adjusted prompts and model settings to explicitly encourage variety — it widened as more essays were added to the sample.
If you run a content operation that publishes at volume, that compounding detail matters more than the headline number. It suggests the homogenising effect gets worse, not better, the more AI-assisted content you and your competitors collectively publish into the same space — which is exactly the dynamic behind why a defensible content strategy in the AI era can no longer rely on covering a topic well; it has to say something a competitor’s AI-assisted draft would not.
It shows up before anyone even starts writing
A third study, run by Barrett Anderson, Jash Shah and Max Kreminski and presented at the ACM Creativity and Cognition conference, tested raw idea generation rather than finished prose. Thirty-six participants produced 1,271 ideas between them, half using ChatGPT and half using a non-AI creativity tool.
The ChatGPT group generated more ideas, and rated their own ideas as more detailed. But when the researchers compared ideas across participants rather than within them, the ChatGPT group’s output was measurably less distinct from one another. Each person felt they were exploring; collectively, they were converging on the same handful of concepts.
One secondary finding is worth sitting with: participants who used ChatGPT reported feeling less responsible for the ideas they generated — not less satisfied with them, less responsible for them. An idea nobody feels ownership over is rarely the one a brand is willing to defend, refine, or build a campaign around.
The pattern holds outside the lab
Controlled experiments are useful but small. So the Moon, Kushlev and Green team went looking for the same signal in the real world. In a 2026 preprint, they examined the personal statements of 372,793 real college applicants, comparing essays written before and after ChatGPT’s late-2022 release.
The finding: After ChatGPT arrived, the vocabulary in these essays became measurably more varied — but the underlying ideas became more similar. The researchers named this gap semantic disjunction: writing that sounds more different from other writing while saying more of the same thing.
This is the finding that should worry anyone doing SEO or content marketing in 2026. Search engines and readers alike are increasingly good at detecting semantic overlap even when the wording changes — which is part of why AI Overviews and generative search results have reshaped what ranks. Surface-level variation no longer buys real differentiation.
A fifth data point: the cost shows up in the brain, too
A team at the MIT Media Lab, led by Nataliya Kosmyna, wired 54 participants to an EEG and had them write essays under three conditions: unaided, with a search engine, or with an AI assistant. The paper, released as a preprint, is not yet peer-reviewed and a subsequent comment paper urged more conservative interpretation of its results — but its direction is consistent with the four studies above.
Brain connectivity was strongest among unaided writers, weaker among search-engine users, and weakest among the AI-assisted group. Minutes after finishing, most participants in the AI group could not quote a line from the essay carrying their own name. And within that group, the essays themselves converged around the same names and phrases — homogenisation, again, measured by a completely different method.
What this actually means for a content business
None of this is an argument against using AI. It is an argument for being precise about which part of the work you hand over and which part you don’t.
Hand over the research gathering, the outline, the first structural pass, the tedious hunt for a source you half-remember. Those are exactly the tasks the research above shows AI genuinely improves.
Keep, without exception, the parts a machine cannot generate because it was never in the room for them:
- A specific decision you made and its actual outcome — a campaign that underperformed, a pricing change that backfired, a client call that changed your approach.
- A number from your own operation, not an industry-average statistic.
- A stated opinion you’re willing to be wrong about in public, rather than a balanced summary of “both sides.”
- Editing passes that cut your favourite lines — the discipline that turns competent AI-assisted drafts into something with a point of view.
A useful test before publishing anything: could a competitor’s AI tool have produced this exact piece from the same prompt? If yes, it is competing in the flattened middle the research describes. If no — because it contains something only you could have contributed — it is doing the one job content still reliably does, which is making a brand memorable rather than merely present.
Frequently asked questions
Does AI-assisted writing actually get worse ratings than human writing?
No — the opposite, in individual tests. Doshi and Hauser’s 2024 study found AI-assisted short stories were rated as more creative and better written than unaided stories. The problem researchers identified isn’t quality; it’s that AI-assisted pieces converge toward similar ideas when compared across many writers.
How much less original is AI-assisted content, in concrete terms?
In the largest study to quantify it, Moon, Green and Kushlev found each additional human-written essay contributed two to eight times more genuinely new ideas to a collection than each additional GPT-4-assisted essay, across 2,200 college essays.
Can better prompting fix the homogenisation problem?
The researchers behind the 2,200-essay study specifically tested this by varying prompts and model parameters to push for more diverse output. The gap in idea diversity persisted regardless of prompting strategy, suggesting the effect is structural rather than something a clever prompt can prompt its way out of.
What is “semantic disjunction”?
It’s the term Moon, Kushlev and Green use for a pattern they found in nearly 373,000 real college application essays: after ChatGPT’s release, the vocabulary in essays became more varied even as the underlying ideas became more similar. Writing looked more different on the surface while saying the same thing underneath.
Does using AI to write actually change brain activity?
An MIT Media Lab study led by Nataliya Kosmyna found EEG-measured brain connectivity was lowest among participants who wrote essays with AI assistance, compared to those using a search engine or writing unaided. This is a preprint that has not completed peer review, and other researchers have published a comment urging conservative interpretation, so treat it as a signal rather than a settled finding.
Should marketers stop using AI for content production?
The research doesn’t support that conclusion. It supports being deliberate about the split: use AI for research, structure, and drafting speed, and reserve firsthand decisions, real numbers, a stated point of view, and final editing for a human who can be wrong in public and defend it.
Why does this matter more for SEO than it used to?
As AI-generated content narrows the range of ideas being published on a given topic, search engines and AI Overviews increasingly reward pieces that add genuinely new information or perspective rather than another well-structured summary of the same consensus. Semantic overlap is easier to detect than surface-level wording changes suggest.
Is this only a problem for writers, or does it affect brainstorming and idea generation too?
It shows up before writing even starts. Anderson, Shah and Kreminski’s study of 1,271 ideas from 36 participants found that ChatGPT users generated more detailed ideas individually, but those ideas were less distinct from one another when compared across the group — and participants reported feeling less ownership over ideas the tool helped produce.
What’s a simple test to check if a piece of content is too generic?
Ask whether a competitor’s AI tool, given the same prompt or brief, could plausibly have produced the same piece. If so, it’s likely competing in the flattened middle these studies describe. Content that includes a specific firsthand decision, an original number, or a defended opinion tends to fail that test in the right way.
Sources: Doshi, A.R. & Hauser, O.P., “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances, 2024. Moon, K., Green, A.E. & Kushlev, K., “Homogenizing effect of large language models on creative diversity,” Journal of Creativity. Moon, Kushlev, Green et al., “The link between diverse words and original ideas is weakening in the AI-era college admissions” (preprint, 2026). Anderson, B.R., Shah, J.H. & Kreminski, M., “Homogenization effects of large language models on human creative ideation,” ACM Creativity & Cognition, 2024. Kosmyna, N. et al., “Your brain on ChatGPT: accumulation of cognitive debt,” MIT Media Lab, arXiv preprint, 2025.