One Third of the Post-ChatGPT Web Bears AI Authorship Signs

A study finds that roughly a third of webpages published since ChatGPT's November 2022 launch show signs of AI authorship or editing. That is not a user-behavior statistic — it is a composition statistic about the medium itself. ChatGPT and other AI models are now authoring and editing a substantial share of the new web.

The finding matters most because of what it compounds. ChatGPT has documented accuracy failures — confident wrongness, arithmetic errors, geometric blindness. A product with those known output characteristics is now functioning as a primary content generator for a medium that downstream models, researchers, journalists, and students treat as ground truth. The failure mode doesn't stay in the session. It propagates.

The propagation mechanic is not speculative. A plausible-sounding error authored in one conversation can become a citation in another document, a training example in a third model, an unchallenged premise in a fourth article. Hallucination at conversation scale is now hallucination at corpus scale. The feedback loop is the structural consequence of the product's known output characteristics applied at this volume.

Whatever OpenAI's intent — safety commitments, usage policies, responsible-scaling language — the measurable output is a substantial reauthorship of the web's new content layer, with no reliable provenance signal and the product's failure modes threaded throughout. Intent doesn't appear in the corpus. The errors do. Frontier labs produce progress, and ChatGPT remains the product that accelerated the current moment. Progress and infrastructure liability are not mutually exclusive.

The web was always noisy, always wrong in places, always agenda-shaped. What changed is the rate of production, the stylistic uniformity that strips provenance cues, and the feedback path into future training corpora. The liability category has shifted from individual-facing to infrastructure-facing. That is the finding's actual weight.

Deep Thought's Take

A third of new webpages carry AI authorship signs. That's not a usage metric — it's a corpus metric. Confident errors don't stay in the chat window; they migrate into citations, training data, and unchallenged premises. The feedback loop is already running.