In November 2022, OpenAI released ChatGPT to the public for the first time. Less than four years later, around half of U.S. adults say they use chatbots powered by artificial intelligence, including 24% who say they use them daily. These tools’ ability to generate human-sounding text has raised a basic question about the modern web: How much content online is now written by AI rather than by other people?
To explore this question, we used the Common Crawl web archive to collect almost half a million English-language webpages from the past five years – starting a couple of years before the release of ChatGPT. We then ran the text of those pages through an AI detection tool called Open Pangram to see how many of them were likely written or substantially edited by AI.
Where online is AI-authored text most common?
AI-authored text is not evenly spread across the web. When ChatGPT was first released, the kinds of linguistic patterns that can signal AI authorship appeared at similar rates across the main top-level web domains (.com, .org, .edu and .gov).
But in samples from 2026, around one-in-ten pages with a .com domain show signs of AI authorship – about double the share on .org domains (4.6%), and 10 times the rate on .edu or .gov domains (both around 1%).
Signs of AI authorship are less common in .edu and .gov domains
% of webpages showing significant signs of AI editing or authorship
Note: Plotted points represent six-month averages.
Source: Pew Research Center analysis of 490,000 English-language webpage texts sampled from the Common Crawl web repository and analyzed using Open Pangram.
PEW RESEARCH CENTER
Signs of AI authorship are less common in .edu and .gov domains
% of webpages showing significant signs of AI editing or authorship
| Crawl date | .com | .org | .edu | .gov | n_com | n_org | n_edu | n_gov |
|---|---|---|---|---|---|---|---|---|
| 2021-01-01 | 1.09 | 0.83 | 0.57 | 0.41 | 29366 | 4850 | 1993 | 445 |
| 2021-07-01 | 1.04 | 1.02 | 0.34 | 0.23 | 23202 | 3828 | 1713 | 407 |
| 2022-01-01 | 1.08 | 0.68 | 0.51 | 0.39 | 11407 | 1924 | 804 | 243 |
| 2022-07-01 | 1.07 | 0.63 | 0.64 | 0.25 | 23734 | 4020 | 1756 | 431 |
| 2023-01-01 | 1.63 | 0.9 | 0.56 | 0.34 | 18631 | 3093 | 1238 | 305 |
| 2023-07-01 | 2.76 | 1.89 | 0.28 | 1 | 12047 | 1925 | 703 | 200 |
| 2024-01-01 | 3.76 | 2 | 1.7 | 1.42 | 23739 | 3590 | 1180 | 328 |
| 2024-07-01 | 4.7 | 2.1 | 0.57 | 1.72 | 35976 | 5049 | 1589 | 545 |
| 2025-01-01 | 5.48 | 2.89 | 0.95 | 1.62 | 35766 | 5012 | 1559 | 536 |
| 2025-07-01 | 6.64 | 2.85 | 1.02 | 1.4 | 34971 | 4727 | 1329 | 509 |
| 2026-01-01 | 9.35 | 4.59 | 1.03 | 0.76 | 41116 | 6363 | 1571 | 669 |
Note: Plotted points represent six-month averages.
Source: Pew Research Center analysis of 490,000 English-language webpage texts sampled from the Common Crawl web repository and analyzed using Open Pangram.
PEW RESEARCH CENTER
What are some common features of AI authored text?
AI detection models aren’t perfect – they sometimes misclassify individual documents that were written by humans as including signs of AI authorship, and vice versa. But if we look at very large collections of texts together, we can start to see that certain types of punctuation, words and phrases show up at much higher rates in AI-generated content than in text written by humans.
WRITTEN BY AI
The internet has evolved into something far beyond a simple network of connected pages — it’s a living, shifting ecosystem shaped by algorithms, creators, and audiences alike, bolstered by waves of AI-generated content that blur the lines between human and machine expression. As users keep delving in to endless streams of posts, videos, and synthetic voices, a pivotal question emerges about authenticity, ownership, and trust — because it’s not just information, it’s influence.
- Em dashes
- Oxford commas
- AI vocabulary
- Negative parallelism
AI models are trained on large datasets of human writing, and they learn to mimic the patterns and styles of that writing. Sometimes certain types of writing are overrepresented in that training process. As a result, models trained on that data can end up using certain words, phrases or linguistic quirks more than humans typically do.
For instance, lots of journalistic and academic writing uses the em dash (—), which is a punctuation mark used to indicate a break in thought or set off a parenthetical statement. AI models tend to use these dashes a lot more than humans typically do. They’re also more likely than human authors to list items in threes and to use Oxford commas in lists.
And as more AI-generated content appears online, these and other AI “tells” have become more common across webpages. Comparing the internet of today to a snapshot from 2023:
- Em dashes appear about twice as frequently.
- Oxford commas see a 63% increase.
- Certain words that AI models like to use (such as “delve,” “interplay” or “testament”) have more than doubled in usage.
- Using “negative parallelism” to structure a comparison (“it’s not just X, it’s Y”) has nearly tripled – although this is still fairly rare overall.
Key words, phrases and punctuation favored by AI have become more common on the web
Uses of __ per 10,000 words
Note: AI-typical vocabulary includes “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,” “vibrant.” Samples filtered to pages published after Nov. 30, 2022. Plotted points represent six-month averages.
Source: Pew Research Center analysis of 490,000 English-language webpage texts sampled from the Common Crawl web repository.
PEW RESEARCH CENTER
Key words, phrases and punctuation favored by AI have become more common on the web
Uses of __ per 10,000 words
| Date | Em dash (per 10k words) | Oxford comma (per 10k words) | AI vocabulary (per 10k words) | Negative parallelism (per 10k words) |
|---|---|---|---|---|
| 2023-01-01 | 5.79 | 34.04 | 11.94 | 0.87 |
| 2023-07-01 | 4.2 | 42.57 | 21.42 | 1.88 |
| 2024-01-01 | 5.32 | 46.25 | 25.66 | 1.9 |
| 2024-07-01 | 4.58 | 44.06 | 23.56 | 2.08 |
| 2025-01-01 | 6.3 | 48.35 | 24.79 | 2.52 |
| 2025-07-01 | 8.58 | 51.08 | 28.15 | 2.72 |
| 2026-01-01 | 11.19 | 55.51 | 26.02 | 2.36 |
Note: AI-typical vocabulary includes “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,” “vibrant.” Samples filtered to pages published after Nov. 30, 2022. Plotted points represent six-month averages.
Source: Pew Research Center analysis of 490,000 English-language webpage texts sampled from the Common Crawl web repository.
PEW RESEARCH CENTER
Of course, em dashes or Oxford commas on their own don’t necessarily mean a particular piece of writing was produced using AI – humans use these in their writing too! But if we look at a lot of documents together, we can start to see patterns in how often human authors and AI models reach for these features on average. And sophisticated detection models like the one we used for this analysis have more to go on than just looking for a few telltale signals. They also learn to recognize more subtle statistical patterns in word choice and sentence structure which help them determine whether a particular piece of writing was likely AI-authored.
About this research
This Pew Research Center essay explores how signs of AI-authored text have increased across the internet over the years since mainstream AI-powered chatbots first hit the market.
Why did we do this?
Pew Research Center does research to help the public, media and decision-makers understand important topics. Tracking the rise of AI and understanding people’s experiences with it are key priorities for us.
Learn more about Pew Research Center and our data science research.
How did we do this?
The data for this essay comes from Common Crawl web archive snapshots collected between January 2021 and July 2026. To analyze this data, we used an open-weight AI detection model created by Pangram. Models like this one look for statistical differences in how human authors and AI models use language, helping them identify when a text shows signs of AI authorship.
Read more about how we did this in this essay’s methodology.
Acknowledgments
This analysis was conducted by Samuel Bestvater, senior data scientist, who also wrote this essay. Carson TerBush, associate information graphics designer, created the graphics and animations and contributed to web development, assisted by Chris Baranovski, lead engineer – editorial content. Janakee Chavda, associate digital producer, produced the report for web. Analysis code was reviewed by Skyler Seets, computational social science assistant, and the essay was checked by Anna Lieb, computational social science analyst. It was copy edited by Anna Jackson, editorial specialist. Aaron Smith, director of Data Labs research, provided editorial guidance.