Generative AI makes each person’s work better. The problem starts when everyone uses it the same way: what gets produced, and what gets decided, starts to look alike. It has never been easier to create. It has never been harder to stand out.
New content is now mostly a mix of human and AI
Ahrefs analyzed 900,000 new English-language web pages detected in April 2025. 74.2% contained AI-generated content. But what dominates is not text made by a machine alone, which barely reaches 2.5%: it’s the mix. Seven out of ten new pages combine human and AI work.
That has a less visible consequence: a growing share of what is on the internet, which is where models learn from, has already passed through a model.
When models learn from themselves, everything converges to the average
In 2024, a team of researchers published in Nature what happens when a model is trained, generation after generation, on data produced by earlier models. They called it model collapse. It isn’t a sudden failure but a gradual degradation: what is rare disappears first, the extremes flatten out and everything moves toward the average. The model ends up learning an impoverished version of reality.
The finding comes with an important nuance. That collapse happens when synthetic data replace real data. Another 2024 study showed that if real data are kept and synthetic data are added to them, collapse is avoided. The lesson isn’t that AI is doomed. It’s that real data are the anchor.
It’s a different risk from hallucinations, which OpenAI attributes to models being trained and evaluated in ways that reward a confident answer over admitting they don’t know. Hallucinations make knowledge wrong. Collapse makes it average.
The same thing happens to companies
The average isn’t just a problem for models. An experiment published in Science Advances in 2024 asked almost 300 people to write a short story. Some received ideas from an AI model and others did not. Six hundred evaluators rated the results.
View data
| Novelty | Usefulness | |
|---|---|---|
| All writers | 8.1% | 9.0% |
| Least creative writers | 10.7% | 11.5% |
Stories with AI ideas came out more novel and better rated, especially those by the least creative writers. But they were also more similar to each other than stories written without AI. The authors call it a social dilemma: each person is better off, and the whole loses variety.
The same happens in companies. When everyone uses the same tools, with similar data and similar questions, the answers converge: more homogeneous strategies, less distinct campaigns, products that look alike and standardized customer service. Using AI was an advantage; today it’s an expectation, and soon it will be irrelevant.
If we all decide with similar models, trained on similar data, how different can our decisions be?
What to do: anchor AI in what only you have
- Use proprietary data. Data on your customers, your operations and your history are what competitors don’t have. In a company, they’re the equivalent of keeping the real data: the anchor against the average.
- Save judgment for what sets you apart. AI can produce; the decisions that define strategy, brand and offer need human judgment.
- Measure whether you’re starting to look alike. Compare messages, offers and campaigns with your competitors’. If they converge, AI is diluting your difference.
In a world where everything looks alike, differentiation stops being technical and becomes human: independent thinking, real experience and judgment.
The real risk isn’t that artificial intelligence will think for us. It’s that we’ll stop thinking.
