The uneven rise of AI-generated web content

Adam Visokay; Soumyakanti Pan; Molly Griston; Polina Iasakova; Mark Nepf; Denis Peskoff; Riya Sinha; Jikai Sun; Isabelle Thiara; Letao Wang; Yiwei Wu; Lauren Yan; Tyler H. McCormick

AI summary · arXiv: 2610.36631v1 · cs.AI

The uneven rise of AI-generated web content

Adam Visokay, Soumyakanti Pan, Molly Griston, Polina Iasakova, Mark Nepf, Denis Peskoff, Riya Sinha, Jikai Sun,

and 5 more authors

Isabelle Thiara, Letao Wang, Yiwei Wu, Lauren Yan, Tyler H. McCormick

Original paper Submitted to arXiv on 27 Oct. 2026 12 pages, 4 figures License: CC BY 4.0 Also in: fr

How to read this summary? Illustrated guide

The essentials

AI-generated or AI-edited text reached roughly one-quarter of eligible Common Crawl pages by August 2026, with accelerating growth concentrated in peripheral domains and English-language content.

Question

How has the prevalence of AI-generated or AI-edited web text changed from 2019 to 2026, and how does it vary by domain centrality, language, and domain category? Do these differences persist after correcting detector errors and page composition?

Method

The authors analyzed a two-stage stratified probability sample of 5,118,347 eligible pages from 505,974 domains across 71 Common Crawl snapshots spanning January 2019 to August 2026; eligibility required at least 500 characters of extracted text. They applied Open Pangram to all pages and Binoculars-Falcon to English-containing pages, calibrating detection errors using pre-2022 control crawls and labeled benchmarks. A Bayesian hierarchical measurement-error model jointly estimated prevalence and detector performance, propagated calibration uncertainty, and poststratified estimates to eligible page populations.

Findings

Result By August 2026, approximately one-quarter of eligible crawlable pages contained AI-generated or AI-edited text, with a 95% credible interval of 22.9–24.4%.
See the passage in the paper
In the most recent crawl, in August 2026, we estimate that about one in four crawlable pages had AI-generated text (95% credible interval 22.9-24.4).
Result Estimated annual prevalence growth accelerated from 4.6 percentage points in February 2023–January 2025 to 8.1 percentage points in January 2025–August 2026.
See the passage in the paper
The average annual change was 0.2 percentage points from January 2019 to February 2023, 4.6 percentage points from February 2023 to January 2025, and 8.1 percentage points from January 2025 to August 2026.
Result The two least-central domain tiers accounted for 78.8% of the net prevalence increase between December 2021 and August 2026, primarily through increasing prevalence within those tiers.
See the passage in the paper
The two least-central tiers, Bottom and Tail, hold 70.5% of eligible pages in August 2026 and account for 78.8% of the net increase (conservative one-sided 95% posterior lower bound: 77.9%).
Result Estimated prevalence in August 2026 was 32.8% for English-only pages versus 17.0% for non-English-only pages, and the difference persisted under a common page composition.
See the passage in the paper
At a common composition of centrality tier, TLD and length, prevalence in August 2026 is 32.5% for English-only pages, 16.8% for non-English-only pages and 21.4% for bilingual pages (Appendix C.2).
Result After standardizing page characteristics, .com retained the highest estimated prevalence and government and educational domains the lowest.
See the passage in the paper
Nevertheless, after standardization by page characteristics, .com remains the highest top-level domain category and the institutional domains the lowest.

Limitations

The estimates cover crawler-accessible pages with sufficiently long extractable text, excluding standalone short-form writing and much social-platform content.
See the passage in the paper
Our estimates therefore describe long-form web text, and say nothing about the prevalence of fully AI-generated or AI-edited text on social platforms.
Detector errors are positively dependent on AI-involved benchmark texts, violating the model's conditional-independence assumption, and the effect on prevalence estimates remains unquantified.
See the passage in the paper
We expect the effect on the prevalence estimate to be limited, because Pangram scores every page and carries most of the information, while Binoculars is applied to english-stratum pages only, but we have not quantified it.
Non-English calibration pools heterogeneous languages with incomplete benchmark coverage and therefore cannot support individual-language prevalence estimates.
See the passage in the paper
The evaluations reveal heterogeneous detector performance across languages, so the non-English-only estimate does not describe any individual non-English language.

Why it matters

The authors present AI detection as a measurement-error problem rather than treating detector flags as definitive authorship labels, enabling population-level comparisons with calibration uncertainty. They argue that uneven growth in AI-influenced web text matters both for public communication and for how earlier language models shape future models trained on web content.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: AI-generated web content Common Crawl AI detection Web centrality Measurement uncertainty

Read further

Comprehensive Summary
Layman's Summary
Blog article
Original abstract

“The Uneven Rise of AI-Generated Web Content” examines the prevalence of generative AI in public-facing web text, its growth between 2019 and 2026, and differences across languages, domain categories, and positions in the web’s hyperlink network. The authors define AI-generated text as both fully generated and AI-edited writing. By examining content in successive web snapshots rather than only newly published material, they aim to describe the accumulated stock of AI-influenced pages. The study analyzes 5,118,347 eligible pages from 505,974 domains across 71 Common Crawl snapshots, spanning January 2019 through August 2026. Eligible pages were successfully captured and contained at least 500 characters of extracted main text. The researchers sampled domains within six PageRank-based centrality tiers, then randomly sampled pages within those domains. They applied Open Pangram to every page and Binoculars-Falcon to pages containing English. Rather than treating detector flags as definitive authorship labels, the researchers used a Bayesian hierarchical measurement-error model to estimate prevalence alongside detector sensitivity and specificity. Pre-2022 crawls served as negative controls for estimating false positives, while labeled benchmarks informed sensitivity estimates. Poststratification adjusted the results to represent the eligible Common Crawl population rather than the sample’s composition. The main estimate indicates that 23.6% of eligible pages contained AI-generated or AI-edited text in August 2026, with a 95% credible interval of 22.9–24.4%. Estimated prevalence was low before widespread generative-AI adoption, then increased substantially. Average annual growth rose from 4.6 percentage points between February 2023 and January 2025 to 8.1 percentage points between January 2025 and August 2026. These endpoint-based averages do not establish a precisely dated change point. Growth was concentrated toward the web’s periphery. In August 2026, prevalence reached 27.5% in the least-central tier, compared with 11.6% in the most-central tier. The two least-central tiers contained 70.5% of eligible pages but accounted for 78.8% of the overall prevalence increase since December 2021. Most of their contribution reflected rising prevalence within those tiers rather than changes in their share of pages. The centrality gradient persisted after standardizing for language, domain category, and text length. Differences across languages and domain categories were also substantial. Estimated prevalence was 32.8% for English-only pages, 17.0% for non-English-only pages, and 20.2% for bilingual pages. The gap between English-only and non-English-only pages widened from 0.6 percentage points in February 2023 to 15.8 percentage points in August 2026. Across domain categories, prevalence was 29.5% for .com pages, compared with 20.4% for .net, 14.6% for .org, 6.1% for .edu, and 5.8% for .gov. Adjustments for composition narrowed some differences but did not eliminate the broad patterns. The authors conclude that generative AI is becoming a substantial but uneven contributor to public web communication, potentially influencing institutional writing conventions and future language-model training data. However, the findings apply to longer-form text observed in web crawls, not to the entire internet, social platforms, or actual audience exposure. Several limitations warrant caution. Detector calibration remains uncertain, particularly for non-English languages pooled together despite limited benchmark coverage. The two detectors also exhibit correlated errors on AI-involved text, contrary to a modeling assumption, and the model does not explicitly account for domain-level clustering. Crucially, the headline estimate depends on assuming that detector performance shifts once after the control period; a more flexible drift model yielded poorly identified estimates. The reported intervals therefore do not capture every source of uncertainty, and detector outputs should not be treated as conclusive evidence of individual authorship.

AI Writing Is Spreading Across the Web—but Not Evenly

How much of the public web now contains writing shaped by generative AI? According to “The uneven rise of AI-generated web content,” nearly one in four eligible pages in an August 2026 web crawl contained AI-generated or AI-edited text. But that headline masks substantial differences: AI-influenced writing was more prevalent on English-language pages, commercial domains, and sites near the web’s hyperlink periphery.

The study, by Adam Visokay, Soumyakanti Pan, Molly Griston, Polina Iasakova, Mark Nepf, Denis Peskoff et al., offers a large-scale view of this transformation. Its findings also come with an important warning: estimating AI involvement across millions of pages is not the same as reliably identifying who—or what—wrote any individual page.

Measuring the Web’s Accumulated Text

The researchers analyzed 5,118,347 eligible pages from 505,974 domains across 71 Common Crawl snapshots, spanning January 2019 through August 2026. Pages qualified if they were successfully captured and contained at least 500 characters of extracted main text.

Rather than examining only newly published material, the study looked at successive snapshots of the web. This approach aims to measure the accumulated stock of AI-influenced pages, including material that remains accessible over time. Its definition of AI-generated content includes both fully generated writing and human writing edited with AI.

To capture differences in the web’s structure, the researchers sampled domains within six tiers based on PageRank centrality, then randomly sampled pages within those domains. Centrality reflects a site’s position in the hyperlink network—not necessarily its audience size or readership.

Why Detector Flags Were Not Treated as Facts

The team applied Open Pangram to every page and Binoculars-Falcon to pages containing English. Because detectors can miss AI-involved text or falsely flag human writing, their outputs were treated as imperfect measurements rather than definitive authorship labels.

A Bayesian hierarchical measurement-error model estimated prevalence alongside detector sensitivity and specificity. Pre-2022 crawls served as negative controls to help estimate false positives, while labeled benchmarks informed sensitivity estimates. Poststratification then adjusted results to represent the eligible Common Crawl population rather than the sample’s particular composition.

A Growing Share, With Faster Recent Increases

The main estimate put AI-generated or AI-edited content at 23.6% of eligible pages in August 2026, with a 95% credible interval of 22.9–24.4%. Estimated prevalence was low before widespread generative-AI adoption and increased substantially afterward.

The average annual increase rose from 4.6 percentage points between February 2023 and January 2025 to 8.1 percentage points between January 2025 and August 2026. These endpoint-based comparisons suggest faster recent growth, but they do not establish a precisely dated turning point.

The Strongest Growth Was at the Periphery

In August 2026, estimated prevalence reached 27.5% in the least-central tier, compared with 11.6% in the most-central tier. The two least-central tiers contained 70.5% of eligible pages but accounted for 78.8% of the overall prevalence increase since December 2021.

Most of that contribution reflected rising prevalence within those tiers, rather than an increase in their share of pages. The centrality gradient also persisted after standardizing for language, domain category, and text length. In other words, those compositional differences did not fully explain the pattern.

Language and Domain Differences Were Substantial

English-only pages had considerably higher estimated prevalence than other language groups:

English-only pages: 32.8%.
Non-English-only pages: 17.0%.
Bilingual pages: 20.2%.

The English-only versus non-English-only gap widened from 0.6 percentage points in February 2023 to 15.8 percentage points in August 2026.

Domain categories showed another pronounced divide. Estimated prevalence was 29.5% for .com pages, compared with 20.4% for .net, 14.6% for .org, 6.1% for .edu, and 5.8% for .gov. Adjusting for composition narrowed some differences without eliminating the broad patterns. These figures describe variation, however; they do not establish why particular groups adopted AI more extensively.

What the Numbers Cannot Tell Us

The findings apply to longer-form text observed in web crawls—not the entire internet, social platforms, or actual audience exposure. A share of pages is not a share of what people read.

Detector calibration remains uncertain, especially for non-English languages pooled together despite limited benchmark coverage. The detectors also showed correlated errors on AI-involved text, contrary to a modeling assumption, and the model did not explicitly account for domain-level clustering.

Crucially, the headline estimate depends on assuming detector performance shifts once after the control period. A more flexible drift model produced poorly identified estimates. The reported credible interval therefore does not capture every source of uncertainty.

A Changing Web, Not a Uniform One

The study suggests that generative AI has become a substantial but uneven contributor to public web communication, with possible consequences for institutional writing conventions and future language-model training data. Its strongest lesson is not simply that AI writing is growing, but that its growth depends on where—and in which language—we look. Detector outputs remain population-level evidence to interpret cautiously, not conclusive proof of individual authorship.

Created on 02 Nov. 2026

Was this synthesis useful?

Score: 0

Ask questions about this paper to our AI assistant

How to chat with this paper? Illustrated guide

Compare with the PDF

You can also chat with multiple papers at once here.

AI assistant instructions?

Tools for this paper

Explain it differently 3 free rewrites left today

Save to a collection Group papers in collections to compare them, chat with them and export them. Create a free account Log in

Export BibTeX · RIS (Zotero, Mendeley, EndNote) · Obsidian (Markdown) · Notion (Markdown)

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero

Citation alerts Citation alerts guide

Who works on it? The people and institutions who publish on its topic

Weekly letter Presented in "Artificial intelligence this week" of 2 November 2026. Artificial intelligence this week: when better scores hide the wrong answer

Journal club Premium+ Slides to present it Propose for the journal club

Versions What changed between versions

Talk about it in public Premium+ Press kit: release, questions, hype check Press kit guide

Papers around this one

See them on a map

Ranked by how close their title and abstract are to this paper’s. Select some to read them together.

Closest papers

Very close Same question
Delving into: the quantification of Ai-generated content on the internet (synthetic data)
D. H. Spennemann · 2025 · arXiv.org · preprint, not peer reviewed · cited 23 times · Not summarized yet

The study estimates how much web text is AI-generated, using ChatGPT-associated keywords rather than detector-calibrated sampling.

Abstract
While it is increasingly evident that the internet is becoming saturated with content created by generated Ai large language models, accurately measuring the scale of this phenomenon has proven challenging. By analyzing the frequency of specific keywords commonly used by ChatGPT, this paper demonstrates that such linguistic markers can effectively be used to esti-mate the presence of generative AI content online. The findings suggest that at least 30% of text on active web pages originates from AI-generated sources, with the actual proportion likely ap-proaching 40%. Given the implications of autophagous loops, this is a sobering realization.
Very close Same question
The Impact of AI-Generated Text on the Internet
J. Dolezal, Sawood Alam, Mark Graham et al. · 2026 · arXiv.org · preprint, not peer reviewed · cited 9 times · Not summarized yet

The study estimates the prevalence of AI-generated or AI-edited text on websites from 2022 to 2025 using an Internet Archive sample and a text detector.

Abstract
The proliferation of AI-generated and AI-assisted text on the internet is feared to contribute to a degradation in semantic and stylistic diversity, factual accuracy, and other negative developments (sometimes subsumed under the Dead Internet Theory). What has hindered answering these questions is that it has not been understood just how much of the internet is actually AI-generated or AI-edited. To this end, we construct a representative sample of websites published on the internet between 2022 and 2025 using the Internet Archive, and apply a state-of-the-art AI text detector on them. We find that by mid-2025, roughly 35% of newly published websites were classified as AI-generated or AI-ass…
Very close Same question
The Rise of AI-Generated Content in Wikipedia
Creston Brooks, Samuel Eggert, Denis Peskoff · 2024 · Proceedings of the First Workshop on Advancing Natural Language Processing for Wikipedia · cited 44 times · Not summarized yet

The study measures AI-generated text in recently created Wikipedia pages with two detectors, rather than across web domains.

Abstract
The rise of AI-generated content in popular information sources raises significant concerns about accountability, accuracy, and bias amplification. Beyond directly impacting consumers, the widespread presence of this content poses questions for the long-term viability of training language models on vast internet sweeps. We use GPTZero, a proprietary AI detector, and Binoculars, an open-source alternative, to establish lower bounds on the presence of AI-generated content in recently created Wikipedia pages. Both detectors reveal a marked increase in AI-generated content in recent pages compared to those from before the release of GPT-3.5. With thresholds calibrated to achieve a 1% false posit…
Very close Broader context
A Study on Influence of Generative AI on the Future of Internet
Binoy Manmadhan, M. Rodrigues · 2025 · The Voice of Creative Research · Not summarized yet

The study discusses generative AI’s effects on the internet, including content creation and concerns about misinformation and online information reliability.

Abstract
The internet has evolved from its origins as a military and academic communications project initiated by the Advanced Research Projects Agency (ARPA) into a decentralized global network for knowledge sharing, communication, and commerce. Today with the rapid advancements in artificial intelligence (AI), Generative AI the branch of AI capable of producing text, images, videos, and other media—has emerged as a trans- formative force across multiple industries by assisting professionals in various tasks. While this technology enhances content creation and problem-solving capabilities, it also raises concerns regarding misinformation, intellectual property, and the reliability of online informat…
Very close Another approach to the problem
Detecting AI Generated Text
I. Chandana, Oruganti Mariya Reshma, Nerella Geetha Sree et al. · 2024 · 2024 2nd World Conference on Communication & Computing (WCONF) · cited 4 times · Not summarized yet

The project trains machine-learning models to distinguish AI-generated from human-written text, while this paper uses detectors to estimate web prevalence.

Abstract
The exponential progression in artificial intelligence (AI) technologies has spearheaded the evolution towards proliferation of AI-generated text across various online platforms. This phenomenon poses significant challenges in maintaining the integrity of information dissemination and combating misinformation. In response, this project proposes a comprehensive approach to detecting AI-generated text through ML methods. By leveraging a diverse dataset of both AI-generated and human-generated text, the project aims to develop robust models capable of accurately discerning AI-generated content. The methodology involves feature engineering, model training, and validation using state-of-the-art N…
Close Broader context
AI-Generated Content on Social Media: Correlations with Persuasion and Disinformation
Peteingthanin Hajumpee, Jesadakom Cheeranoravanich, Charnyote Pluempitiwiriyawe · 2026 · International Joint Conference on Computer Science and Software Engineering · cited 2 times · Not summarized yet

The study examines AI-generated political posts on Bluesky and their correlations with persuasion and disinformation, rather than prevalence across web pages.

Abstract
The rapid proliferation of large language models (LLMs) and generative AI in creating fluent political messaging has raised significant concerns regarding persuasion and the spread of disinformation. This study analyzes AI-generated content and its correlation with these phenomena using a 24-month dataset of 120,000 U.S. politics-related posts from the Bluesky platform. Utilizing OSM-Det for AI detection and PCoT for quantifying persuasion and disinformation, we examine the shifting landscape of automated political discourse. Our findings reveal: (i) a substantial increase in both AI-generated content and disinformation; (ii) a strong positive correlation between AI-generated output and disi…
Close Shared data, tool or method
Counter Turing Test CT^2: AI-Generated Text Detection is Not as Easy as You May Think -- Introducing AI Detectability Index
Megha Chakraborty, S. M Towhidul Islam Tonmoy, S M Mehedi Zaman et al. · 2023 · Computation and Language · cited 35 times · AI summary available

The study examines how detectable AI-generated text is and introduces an AI Detectability Index, addressing detector performance relevant to prevalence estimation.

Abstract
With the rise of prolific ChatGPT, the risk and consequences of AI-generated text has increased alarmingly. To address the inevitable question of ownership attribution for AI-generated artifacts, the US Copyright Office released a statement stating that 'If a work's traditional elements of authorship were produced by a machine, the work lacks human authorship and the Office will not register it'. Furthermore, both the US and the EU governments have recently drafted their initial proposals regarding the regulatory framework for AI. Given this cynosural spotlight on generative AI, AI-generated text detection (AGTD) has emerged as a topic that has already received immediate attention in researc…
Close Review
AI-Generated Content in Cross-Domain Applications: Research Trends, Challenges and Propositions
Jianxin Li, Liang Qu, Tao-Tao Cai et al. · 2025 · Knowledge-Based Systems · cited 22 times · Not summarized yet

The paper reviews AIGC applications, research trends, and challenges across domains including marketing, education, and public health.

Abstract
Artificial Intelligence Generated Content (AIGC) has rapidly emerged with the capability to generate different forms of content, including text, images, videos, and other modalities, which can achieve a quality similar to content created by humans. As a result, AIGC is now widely applied across various domains such as digital marketing, education, and public health, and has shown promising results by enhancing content creation efficiency and improving information delivery. However, there are few studies that explore the latest progress and emerging challenges of AIGC across different domains. To bridge this gap, this paper brings together 16 scholars from multiple disciplines to provide a cr…

Go further

How are these papers found?

We compare the title and abstract of this paper with those of every paper summarized on this site, and with the papers Semantic Scholar recommends, those it cites and those citing it (for a much-cited paper, its most recent citations and, from OpenAlex, the most cited ones). Among the prior work and the follow-ups, the papers that use each other in depth, a method it uses and the much-cited follow-ups come a little higher. The levels (very close, close, related) come from the similarity of the texts, not from a judgment of quality; * marks a paper without an abstract, compared by its title. The sentence under a paper is written by an AI from the two abstracts: check it in the papers. A paper marked not relevant by three readers leaves the list.

Updated on 2 November 2026.

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.

The uneven rise of AI-generated web content

The essentials

Question

Method

Findings

Limitations

Why it matters

Read further

Definitions

AI Writing Is Spreading Across the Web—but Not Evenly

Measuring the Web’s Accumulated Text

Why Detector Flags Were Not Treated as Facts

A Growing Share, With Faster Recent Increases

The Strongest Growth Was at the Periphery

Language and Domain Differences Were Substantial

What the Numbers Cannot Tell Us

A Changing Web, Not a Uniform One

Ask questions about this paper to our AI assistant

Tools for this paper

Advanced tools Premium

Papers around this one

Closest papers

Go further

Cite this paper (BibTeX)