A data essay published by Pew Research Center's Data Labs on August 20, 2026, reporting an original measurement of how much English-language web text shows signs of having been written or substantially edited by AI (Source: pewresearch.org). The analysis was conducted and written by Samuel Bestvater, senior data scientist; Carson TerBush produced the graphics and animations, assisted by Chris Baronavski, lead engineer for editorial content; Janakee Chavda produced the report for the web; and Aaron Smith, director of Data Labs research, provided editorial guidance. Analysis code was reviewed by Skyler Seets and the essay checked by Anna Lieb.
The essay frames the question against Pew's own survey finding that around half of U.S. adults say they use AI-powered chatbots, including 24% who say they use them daily — a figure the Center's broader public-opinion work also carries (see Americans and AI 2026: Chatbots, Smart Devices and Views on Impact).
Method
The corpus is drawn from Common Crawl web archive snapshots collected between January 2021 and July 2026 — a window beginning roughly two years before ChatGPT's November 2022 release. The published chart source line describes the base as "Pew Research Center analysis of 490,000 English-language webpage texts sampled from the Common Crawl web repository and analyzed using Open Pangram"; the essay text describes it as "almost half a million" pages. Detection used Open Pangram, an open-weight AI-detection model created by Pangram, which classifies text by statistical differences in how human authors and language models use language. A separate methodology page accompanies the essay (pewresearch.org) and is not captured in the raw file.
Plotted points in both published tables represent six-month averages, so the most recent tabulated row is dated 2026-01-01 even though the headline snapshot is from July 2026.
Headline findings
In a random sample of 10,000 webpages collected in July 2026, 10% showed significant signs of AI authorship. Pew notes that a random sample of the web mixes new and old material, and that many pages in such a sample could not have been written by AI at all. Filtering to only pages published after ChatGPT's release raises the figure: in the July 2026 snapshot, signs of AI authorship appear in over one-third of post-release pages. Pew states this is consistent with other studies finding that large shares of recently published pages were likely written or substantially edited by AI.
The essay dates the upward trend to late 2022, continuing as other chatbots including Claude and Gemini followed.
Distribution by top-level domain
Pew reports that at the time of ChatGPT's release the linguistic patterns signalling AI authorship appeared at similar rates across the main top-level domains, and that the rates have since diverged. The published series:
| Crawl date (six-month average) | .com | .org | .edu | .gov |
|---|---|---|---|---|
| 2021-01-01 | 1.09% | 0.83% | 0.57% | 0.41% |
| 2022-01-01 | 1.08% | 0.68% | 0.51% | 0.39% |
| 2023-01-01 | 1.63% | 0.90% | 0.56% | 0.34% |
| 2024-01-01 | 3.76% | 2.00% | 1.70% | 1.42% |
| 2025-01-01 | 5.48% | 2.89% | 0.95% | 1.62% |
| 2026-01-01 | 9.35% | 4.59% | 1.03% | 0.76% |
Sample sizes for the 2026-01-01 row were 41,116 .com pages, 6,363 .org, 1,571 .edu and 669 .gov. The essay summarizes the 2026 samples as around one in ten .com pages, about double the .org share, and roughly ten times the .edu and .gov rates, which it describes as both around 1%.
Linguistic markers
Pew reports that certain punctuation, words and phrases occur at higher rates in AI-generated text than in human-written text, and tracks four such markers per 10,000 words across pages published after November 30, 2022:
| Date (six-month average) | Em dash | Oxford comma | AI vocabulary | Negative parallelism |
|---|---|---|---|---|
| 2023-01-01 | 5.79 | 34.04 | 11.94 | 0.87 |
| 2024-01-01 | 5.32 | 46.25 | 25.66 | 1.90 |
| 2025-01-01 | 6.30 | 48.35 | 24.79 | 2.52 |
| 2026-01-01 | 11.19 | 55.51 | 26.02 | 2.36 |
Comparing 2026 against a 2023 baseline, the essay reports that em dashes appear about twice as frequently, Oxford commas increased 63%, AI-favored vocabulary more than doubled, and negative parallelism — structuring a comparison as "it's not just X, it's Y" — nearly tripled while remaining rare overall. The AI-vocabulary list comprises "additionally," "align with," "boasts," "bolstered," "crucial," "delve," "emphasizing," "enduring," "enhance," "essential," "fostering," "garner," "highlight," "interplay," "intricate," "key," "landscape," "meticulous," "perfectly," "pivotal," "showcase," "significant," "tapestry," "testament," "underscore," "valuable," and "vibrant."
Pew attributes the pattern to overrepresentation of certain writing styles in training data, noting that journalistic and academic writing uses the em dash and that models use it more than humans typically do, and that models are also more likely to list items in threes and to use Oxford commas.
Stated limitations
The essay states that AI detection models "aren't perfect" and sometimes misclassify human-written documents as AI-influenced and the reverse, and that individual markers such as an em dash or an Oxford comma do not on their own indicate AI authorship. Its claim is about aggregate patterns across very large collections of texts rather than about any individual document, and it notes that the detection model draws on subtler statistical patterns in word choice and sentence structure beyond the four illustrated markers. Because the measure is a classifier output rather than a record of authorship, the known error modes of detection classifiers — decay as models improve, and false positives on non-native English and on AI-assisted human writing — apply to the reported figures.
Provenance
Full text captured from pewresearch.org on August 23, 2026, including both published data tables. The byline spells the graphics engineer's name "Chris Baronavski" while the acknowledgments spell it "Chris Baranovski"; both spellings are as published. The linked methodology companion was not captured and no DOI is issued for a Pew data essay.
Relationships
- supports: AI Content Provenance — supplies a measured prevalence figure for the problem provenance tooling addresses.
- supports: Model Collapse — quantifies the synthetic-text share of the web corpus on which the recursive-training concern depends.
- supports: AI Content Saturation (\"AI Slop\") — an independent measurement alongside the Washington Post markers and the Pangram social-platform data.
- related: The Impact of AI-Generated Text on the Internet — Dolezal, Alam, Graham, Bohacek (Imperial / Internet Archive / Stanford, April 2026) — an earlier Common Crawl / Internet Archive measurement of the same phenomenon.
- related: Americans and AI 2026: Chatbots, Smart Devices and Views on Impact — the Center's survey work on public AI use, cited in this essay's framing.
- related: Media, Journalism & Entertainment — AI Deployment — Pangram-based detection is also the tooling at issue in platform disputes over AI-writing enforcement.