Połowa internetu napisana przez AI. Co to oznacza dla czytelnika?
W listopadzie 2022 roku uruchomiono ChatGPT. Mniej niż cztery lata later analysis of the Common Crawl web archive — a project that periodically snapshots the English-language internet — shows that roughly one in ten web pages now carries significant signs of AI authorship. Among pages published after ChatGPT’s release, that figure rises to over one-third.
Badacze z Graphite, korzystając z trzech niezależnych detektorów (Pangram, GPTZero, Copyleaks), oszacowali, że w pierwszym kwartale 2026 roku primarily AI-generated articles stanowiły 49,9 proc. wszystkich publikowanych tekstów — praktycznie tyle, ileRemaining human-written articles. W czwartym kwartale 2025 roku proporcja ta wzrosła do 50,9 proc., by wrócić do 49,9 proc. w kolejnym kwartale. Od Q1 2025 roku udział tekstów AI-generowanych stabilizuje się wokół 50 proc.
Co ciekawe, te same badania wykazują, żearticles primarily AI-generated rarely appear in Google search results or ChatGPT answers. Authors hypothesize that practitioners discovered such content performs poorly in search, which may explain the plateau. Economic incentives, not coordinated intent, drive the proliferation.
NewsGuard, organization monitoring misinformation, identified 3,749 AI Content Farm websites across 16 languages. These sites — often bearing generic names like “Times Business News” — publish dozens of articles daily, frequently recycling false claims about public health, political leaders, and major brands. Their revenue model relies on programmatic advertising, meaning top brands inadvertently fund them unless they actively exclude such domains.
Pew Research Center’s July 2026 analysis of 10,000 randomly sampled web pages found that .com domains show AI authorship signs at roughly double the rate of .org domains (10% vs. 4.6%) and ten times the rate of .edu or .gov domains (both around 1%). The gap suggests that institutional domains maintain stronger editorial controls, while commercial content farms operate with minimal human oversight.
Researchers at Imperial College London, Internet Archive, and Stanford University found that by mid-2025, roughly 35% of newly published websites were AI-generated or AI-assisted — up from zero before ChatGPT. Their study documented a statistically significant correlation between rising AI text prevalence and shrinking semantic diversity: websites predicted to be AI-generated showed 33% higher average pairwise semantic similarity than non-AI sites. They also found increased positive sentiment in AI-generated text, consistent with the sycophantic and over-optimistic tendencies of existing language models.
Importantly, the same study did not find statistically significant evidence that AI-generated text decreases factual accuracy or stylistic diversity at the macro level — diverging from public perception, where 75.1% of surveyed U.S. adults believe AI content increases factual errors. The researchers caution that the threat may be epistemic rather than purely factual: an environment flooded with cheerful, homogenized text risks marginalizing human dissent without deploying overt disinformation.
The proliferation also carries consequences for AI research itself. As much as 35% of contemporary web text is now AI-generated, future foundation models trained on crawled internet data will inevitably ingest semantically less diverse datasets. This transforms the theoretical risk of model collapse — where models degrade after recursively ingesting AI-generated data — into an empirical concern.
Oznaczenie treści AI
Ten artikel powstał z udziałem sztucznej inteligencji: research oraz tekst przygotował i zredagował agent AI. Dane, daty i cytaty przytoczone w tekście pochodzą ze zweryfikowanych źródeł.
