Ahrefs looked at nearly 900,000 web pages published in a single month this year. Roughly 74% of them contained AI-generated content. A separate analysis of 65,000 articles found that AI-written pieces crossed the halfway mark of all new web content in late 2024, up from about 5% before ChatGPT existed.
More than half of what gets published online now starts with the same handful of models, trained on roughly the same pool of existing text, prompted by people asking roughly the same questions.
It's a description of what's happening, and if you're trying to make something worth paying attention to, it's a problem.
If differentiation is part of your job — you're a product marketer, demand gen lead, or creative — then how AI is used should concern you. Researchers ran a study where people generated ideas, half using ChatGPT and half using a non-AI creativity tool. The result: the group using ChatGPT produced ideas that were more similar to each other, not within one person's output, but across different people. Everyone converged toward the same answers.
That's the mechanism behind the feeling you've probably had scrolling LinkedIn or reading a competitor's blog lately, the sense that everything sounds like it was written by the same person, even when it wasn't. In a sense, it was. Just not a person.
A more recent review in Trends in Cognitive Sciences extends this beyond writing. The argument is that homogenization isn't a side effect of how people use these tools. It's built into how the models are trained. They're optimized to favor patterns that show up often and generalize well, and to smooth over the unusual, the minority, the outlier. Which means the convergence doesn't stop at sentence structure. It applies to the strategies, frameworks, and positioning angles that come out of those sentences too. If the tool helping you think pulls everyone toward the same statistically likely answers, your competitors' positioning workshops are pulling toward the same place as yours.
None of this means AI doesn't belong in creative work. It's genuinely useful for the parts of the process that used to eat the most time: early research, rough drafts, a wide first pass of options before narrowing down. Used that way, it gets you to a starting point faster than ever.
The problem isn't that AI helps you go faster. It's that the speed comes from a process that, by design, can't do the one thing creative work eventually requires.
AI doesn't invent. It synthesizes. Those sound similar but they're not the same thing.
Margaret Boden, a philosopher who spent decades studying what creativity actually is, broke it into three types: combinational (new combinations of ideas that already exist), exploratory (finding new things within a space whose rules are already set), and transformational (changing the rules of the space itself). AI is genuinely good at the first two. Give it a brief and it'll combine and explore within the boundaries of what it's seen, often in ways that feel surprising. What it consistently struggles with is the third kind, the move that changes what's even possible within the category.
That third kind is what separates imagination from hallucination. Imagination is a new connection that holds up when it meets the real world. It's surprising, but it's also relevant, useful, true in some way other people recognize once they see it. Hallucination is a new connection with nothing to check itself against. The model has no contact with reality, no product it built, no customer it talked to, no failure it lived through. It only has more text. So it can recombine everything it's been shown, and explore the space defined by what it's been shown, but it has no way to step outside that space, because outside the space is exactly where it has nothing.
This is the technical reason AI output tends to converge. It's not a flaw that better prompting fixes. It's simply what the system is.
Speed has become the default measure of content quality, which is a strange thing to say out loud. Faster isn't better. Faster is just faster.
MIT Media Lab ran a study worth knowing about if you make anything for a living. Over four months, they tracked 54 people writing essays, some using ChatGPT, some using search engines, some using nothing but their own thinking. The group that relied on ChatGPT showed the lowest brain engagement, the worst memory of what they'd written, and the least sense of ownership over the result. Two English teachers reviewing the essays blind, without knowing which group wrote what, independently described the AI-assisted essays as "soulless."
A separate study out of Stanford found something similar from the other direction. Researchers had AI generate research ideas and compared them to ideas from human experts. Reviewers rated the AI ideas as more novel. Then 43 researchers spent an average of over 100 hours actually building out the ideas they'd been assigned. The novelty didn't hold up. Human ideas were judged more feasible, and the AI ideas, the ones that looked more original on paper, didn't translate into better outcomes once someone tried to execute them.
Novel on the page and valuable in practice are two different things. Speed gets you the first one. It doesn't get you the second.
There's a reason "sleep on it" has survived as advice for centuries, and it's not just folk wisdom. Researchers have studied what's called the incubation effect, what happens to a problem when you stop consciously working on it, for decades, and the findings are consistent: stepping away measurably improves creative output. Something keeps working on the problem while your attention is elsewhere.
A 2025 study added a useful detail. Researchers had people write short stories, take a break, then continue. Some breaks were demanding, requiring focus on another task. Some were boring enough to let the mind wander. The boring breaks produced bigger improvements in the second half of the story. Mind-wandering, not rest, was doing the work.
This part gets lost in a culture optimized for speed. The slow part of the process, the part that looks like nothing is happening, is often where the actual thinking happens.
So what does this look like in practice, especially for the content that actually matters: product pages, technical posts, anything where being right is as important as being readable.
Start with the people who know the thing best. For most companies, that's not marketing. It's engineering, product, and leadership, the people who built whatever you're writing about and understand it at a level no outside tool can. Have them draft the outline: the claims, the structure, the order the ideas need to go in. This isn't a nice-to-have. AI hallucination rates climb sharply on niche and technical topics, with some research putting the increase as high as 50% in domains where training data is thin, compared to a much lower baseline on well-covered ground. The categories where your content needs to be most accurate are the categories where AI is least equipped to be accurate on its own.
Once there's agreement on what needs to be said, marketing takes the first pass at how to say it, and the team reviews it together. Timing matters here in a way that's been studied directly. A 2025 study found that bringing AI into the ideation process from the start, before people had generated their own ideas, reduced the number of original ideas people came up with and lowered how much ownership they felt over the result. The researchers' recommendation was straightforward: let people think first, and bring AI in after.
AI's role comes last, and it's a narrower one than "write this for us." A pass for brand alignment, tone, grammar, consistency. This is precisely the kind of work AI is built for: pattern-matching against an existing standard, catching what doesn't fit. It's combinational creativity, the thing Boden's framework says AI does well, applied to a task where there's no fact to get wrong, just a voice to match.
Humans first, AI second. The order is what determines whether AI is helping you say something or helping you sound like everyone else.
None of this is an argument against AI. It's an argument for the slow parts of the process — the drafting, the disagreement, the overnight pause before something goes out. That's where the work actually happens.
Everyone has access to the same tools now, pulling from roughly the same pool of everything that's ever been written. The thing that doesn't show up in that pool is what you know that nobody else does, and the time it takes to say it in a way that sounds like you said it.
At this point, it might be the only thing left to compete on.
Sources
- Ahrefs, AI content prevalence study (April 2025): https://ahrefs.com/blog/ai-content-prevalence
- Axios, on Graphite's analysis of AI-generated articles: https://www.axios.com/2025/10/14/ai-generated-writing-humans
- Anderson, Shah, Kreminski, "Homogenization Effects of Large Language Models on Human Creative Ideation" (2024): https://arxiv.org/abs/2402.01536
- "The homogenizing effect of large language models on human expression and thought," Trends in Cognitive Sciences (2026): https://www.cell.com/trends/cognitive-sciences/abstract/S1364-6613(26)00003-3
- The Marginalian, on Margaret Boden's framework for creativity: https://www.themarginalian.org/2025/08/22/margaret-boden-creativity/
- MIT Media Lab, "Your Brain on ChatGPT: Accumulation of Cognitive Debt" (2025): https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- "The Ideation-Execution Gap" (Stanford, 2025): https://arxiv.org/pdf/2506.20803
- Ritter & Dijksterhuis, "Creativity—the unconscious foundations of the incubation period" (2014): https://pmc.ncbi.nlm.nih.gov/articles/PMC3990058/
- "Mind wandering during creative incubation predicts increases in creative performance," Scientific Reports (2025): https://www.nature.com/articles/s41598-025-09736-y
- SQ Magazine, LLM hallucination statistics on niche domains: https://sqmagazine.co.uk/llm-hallucination-statistics/
- Qin et al., "Timing Matters: How Using LLMs at Different Timings Influences Writers' Perceptions and Ideation Outcomes" (CHI 2025): https://arxiv.org/abs/2502.06197
