Why did AI article share stop growing after 2025?

The excerpt does not explain why the primarily AI-generated share stopped rising, and it offers two candidate explanations it cannot test. The first is demand-side. Graphite hypothesizes that practitioners "found that primarily AI-generated articles do not perform well in search," and it suspects that such articles draw less traffic than human-written ones without measuring traffic. The second is measurement-side. The flat line could come from detectors that miss more text as models improve, since the excerpt warns that "AI models continue to improve, and may become harder to detect." Either explanation could produce the flat line Graphite reports, and nothing in the excerpt tells them apart.

The search explanation rests on the motive the excerpt gives for publishing. Companies have "explored publishing content generated by LLMs" to "grow their traffic across channels such as Google Search, social, and advertising," as "a cost-effective alternative to spending hundreds of dollars for humans to write content." If those articles fail to rank, the cost advantage buys less traffic and the incentive to publish them weakens, which would cap the share. That chain is a reading of the excerpt's stated motives, not a result Graphite reports. The detector explanation rests on a narrower base: detector accuracy was checked only as false-negative rates, on three named models, and not for mixed drafts.

The measured plateau is the one Has AI-generated content stopped growing on the web? reports, and this note asks what that plateau means. The mixed-draft gap is where the two explanations meet. Do writers actually edit AI-generated text before publishing? reports that writers rarely edit AI text much, which would keep the mixed-draft gap small in practice. A separate analogy points the other way on error direction: the Why do fake news detectors flag AI-generated truthful content? study finds classifier error tied to LLM style is systematic rather than random. By analogy, a detector keyed to the same style would bend a measured share in one direction, not simply add noise. That analogy comes from a different detector type and is not something the excerpt tests.

The excerpt establishes none of the search or traffic claims, since the separate study it cites is not included. It reports no false-positive rates and no accuracy for edited drafts, and it tests no models beyond the three named. The 50% figure is a measurement of published articles under current detectors. Reading it as a saturation point, as if the web had stopped absorbing AI text, or as evidence that search penalizes AI articles, goes beyond the evidence in either direction. The next checks follow from the excerpt's own gaps: the traffic share of the classified articles, and detector accuracy on newer models and on edited drafts.