As AI search tools increasingly refer to AI-generated content, warnings are also growing that answers could become more uniform and more vulnerable to manipulation.

Axios reported on Wednesday that AI search visibility consultancy Graphite said an “AI search collapse” could occur when AI search tools pull into results pages derived from earlier AI answers. The claim borrows the concept of “model collapse” presented in a 2024 Nature paper, referring to the risk that quality gradually deteriorates when AI outputs are repeatedly used for retraining.

Graphite said simulations using OpenAI, Gemini and Anthropic APIs found that models relying on AI-generated references increasingly tended to produce identical recommendations.

According to Wharton School research, individuals who used ChatGPT as a research partner produced stronger ideas, but groups using it together tended to converge on similar concepts. Wharton School professor Gideon Nave said, “If everyone relies on LLMs, diversity, randomness and room for exploration disappear.”

Companies are already trying to manipulate AI search in their favour by using so-called generative engine optimisation (GEO) and posting chatbot-friendly content so AI systems cite and recommend them.

But Axios said the research does not prove that AI search is already collapsing.

Axios also said the findings should be considered in light of the fact that it is a simulation study by a company with interests in the AI search visibility field, not an academic study that has undergone peer review.