Ask how much of the internet is now AI-generated and you'll get wildly different answers depending on who you ask — anywhere from a fifth to nearly all of it. That range itself is the first useful data point: nobody has a precise number, but every rigorous study points the same direction, and the pace of change is the part worth taking seriously.
What the most careful study actually found
The most methodologically serious answer so far comes from a joint study by researchers at Stanford University, Imperial College London, and the Internet Archive, published in 2025. Rather than guessing, the team pulled 33 months of website snapshots from the Wayback Machine and ran them through Pangram v3, a purpose-built AI text detector, to classify each page. Their finding: by mid-2025, roughly 35% of newly published websites were AI-generated or AI-assisted — a figure that was essentially zero before ChatGPT launched in November 2022 (Decrypt's coverage of the Stanford/Imperial College/Internet Archive study). One of the paper's co-authors, Imperial College London researcher Jonáš Doležal, described the speed of the shift as staggering — after decades of humans shaping the web, a significant portion of it changed character within roughly three years.
The study's more interesting contribution isn't the headline percentage — it's what the researchers tested and ruled out. They checked six hypotheses about what AI content actually does to the web, and only two held up under real scrutiny. First: AI-generated pages showed pairwise semantic similarity scores 33% higher than human-written pages, meaning AI content trends toward expressing the same ideas in increasingly similar ways — a narrowing of the range of things actually being said, not just how they're said. Second: AI-generated content showed sentiment scores more than 107% more positive than human writing, consistent with the well-documented tendency of language models to default toward agreeable, upbeat framing regardless of the actual subject matter.
The linguistic fingerprint is measurable — and worth checking your own writing against
A separate large-scale analysis, examining a broad 2026 sample of new web pages, found the "linguistic fingerprint" of AI-written content becoming statistically detectable at scale: a doubling in em-dash usage compared to 2023, a 63% increase in Oxford comma usage, more than double the frequency of AI-preferred vocabulary (words like "delve," "tapestry," "interplay," and "pivotal"), and a near-tripling of the "it's not just X, it's Y" sentence construction (coverage of the 2026 Pew Research study on AI writing patterns). No single one of these signals proves a page was AI-written, but in aggregate, they describe a detectable statistical drift in how a growing share of the web now sounds.
This is worth treating as a practical, actionable checklist, not just a curiosity: if your own recent posts are unintentionally leaning on the same tics — heavy em-dash use, reflexive "not just X, it's Y" constructions, or a rotating cast of the same five overused words — that's worth noticing and editing out deliberately, independent of whether AI tools were involved in drafting them at all. Sounding statistically similar to the flood of AI-assisted content is a real risk even for entirely human-written posts that happen to lean on the same current stylistic habits.
The same study also found where this is happening fastest — and where it isn't
The 2026 analysis broke its findings down by domain type, and the pattern is instructive for anyone thinking about where trust still concentrates online. On .com domains, about one in ten pages showed AI-generated characteristics — more than double the rate on .org domains (4.6%), and roughly ten times the rate on .edu and .gov domains (both around 1%). The commercial internet is adopting AI-assisted writing far faster than academic and government institutions, which have both slower publishing cycles and stronger institutional norms around named, credentialed authorship.
For an independent blogger, this is a useful reference point rather than a call to imitate .edu formality: it confirms that named authorship and credentialed, verifiable expertise — the same qualities that keep .edu and .gov content trusted — are precisely the qualities becoming scarcer, and therefore more differentiating, across the commercial web where most blogs compete.
The counterintuitive good news: what actually gets surfaced still skews heavily human
Here's where the doom-laden "90% of the internet will be AI by the end of 2026" framing — a widely repeated but far less rigorously sourced claim than the Stanford/Imperial figure — falls apart against real measurement. A follow-up survey by the SEO research firm Graphite, reported by Axios, found that only 14% of articles actually ranking in Google search results were AI-generated; 86% were still written by humans. Chatbot answer engines showed a similar pattern: 82% of articles cited by ChatGPT and Perplexity in their answers were human-written (University of Technology Sydney's coverage of the Graphite/Axios findings).
Put the two findings together and a coherent picture emerges: AI-assisted content is flooding the web in raw volume — roughly a third of new websites by one rigorous measure — but search engines and AI answer engines themselves appear to be actively filtering much of it out of what actually gets surfaced to a real user. The flood is real. Being drowned by it, for content that's actually well-structured, credentialed, and original, is not the inevitable consequence the loudest headlines suggest.
Why this might get more important, not less, over the next several years
There's a longer-term dynamic worth understanding, because it changes how you might think about the value of your own original writing going forward. In 2024, a paper published in Nature by researchers at Oxford and Cambridge demonstrated a phenomenon called model collapse: when AI systems are trained recursively on data increasingly generated by earlier AI systems, their outputs gradually drift from the true underlying distribution of ideas, losing rare and unusual patterns and becoming progressively more homogeneous — and in worst cases, increasingly meaningless — with each generation. Researchers at Epoch AI have projected that the supply of high-quality, genuinely human-generated text suitable for training future AI models could be effectively exhausted sometime between 2026 and 2032, given how much of new web content is now AI-assisted itself.
If that holds, genuinely original, human-authored writing doesn't just compete better with the current flood of AI-assisted content for a reader's attention — it may become an increasingly scarce and valuable input for the AI industry itself, which needs a continuing supply of real, non-recursive human writing to keep training future models from degrading. That's a strange, second-order reason to keep writing something real, on top of the more immediate reader-trust argument.
What to actually do with this
- Audit your last five posts for the AI linguistic fingerprint — heavy em-dash use, reflexive "it's not just X, it's Y" constructions, and an overreliance on words like "delve" or "pivotal" — regardless of how they were drafted, since sounding statistically like the flood carries its own cost.
- Lean into named, credentialed authorship, the trait separating the still-trusted .edu/.gov corner of the web from the fastest-AI-adopting .com corner.
- Don't assume the flood means you're invisible. The measured data on what actually ranks and gets cited skews far more human than the loudest "dead internet" framing suggests — the opportunity to stand out with genuinely original work is real, not purely theoretical.
- Treat original, human-authored writing as a long-term asset, not just a short-term differentiator — the scarcity argument here may only get stronger as more of the web becomes AI-assisted.
Frequently asked questions
Is the "90% of the internet will be AI by 2026" claim actually true?
It's a widely repeated forecast, but it traces back to far less rigorous sourcing than the peer-adjacent Stanford/Imperial College/Internet Archive study, which measured 35% of newly published websites as AI-generated or AI-assisted using a validated detector across 33 months of archived data. Treat the more extreme figure with real skepticism.
How do I know if my own writing accidentally reads as AI-generated?
Check for the specific, measured linguistic markers: unusually frequent em-dashes, a high rate of Oxford commas relative to your past writing, repeated use of a small set of AI-favored words, and the "it's not just X, it's Y" sentence pattern. None of these alone is conclusive, but a cluster of them is worth a deliberate editing pass.
Does the flood of AI content mean my blog is competing against an impossible volume of new pages?
In raw volume, somewhat — but the evidence on what actually gets surfaced in search results and AI citations shows a strong, measured skew toward human-written content. The flood affects total web volume more than it affects your realistic competition for visibility.
What is "model collapse" and does it actually matter for bloggers?
It's a documented phenomenon where AI models trained on increasingly AI-generated data degrade in quality over successive generations. It matters for bloggers indirectly: if genuinely human, original writing becomes a scarcer resource for AI training, that may add a second, longer-term reason — beyond direct reader trust — for original content to hold or gain value over time.
No comments yet — be the first to share your thoughts.