![\"\"](\"https://marketing-site.graphite-wf.workers.dev/_next/image?url=%2Fimages%2Fresearcher-jose.webp&w=256&q=75\")
Jose Luis Paredes, PhDSenior Data Specialist![\"\"](\"https://marketing-site.graphite-wf.workers.dev/_next/image?url=%2Fimages%2Fresearcher-gregory.webp&w=256&q=75\")
Gregory Druck, PhDChief AI Officer![\"\"](\"https://marketing-site.graphite-wf.workers.dev/_next/image?url=%2Fimages%2Fresearcher-bevin.webp&w=256&q=75\")
Bevin BensonGrowth Advisor![\"\"](\"https://marketing-site.graphite-wf.workers.dev/_next/image?url=%2Fimages%2Fresearcher-ethan.webp&w=256&q=75\")
Ethan SmithCEOJose Luis Paredes, PhDSenior Data SpecialistGregory Druck, PhDChief AI OfficerBevin BensonGrowth AdvisorEthan SmithCEO
ContentsCopy linkDiscuss with AI
\nHalf of Articles Being Published Online Are Now Generated by AI
\nKey Takeaways
\n![\"\"](\"https://i.ytimg.com/vi/S1tn3MwavCU/maxresdefault.jpg\")
Contents
\nMotivation
\nResults
\nMethodology
\nLimitations
\nAppendix
\nSince ChatGPT launched in November 2022, many companies have explored publishing content generated by LLMs such as ChatGPT, Claude, and Gemini to grow their traffic across channels such as Google Search, social, and advertising. This is a cost-effective alternative to spending hundreds of dollars for humans to write content.
\nThe quality of AI content is rapidly improving. In many cases, AI-generated content is as good or better than content written by humans (MIT Study). It is often hard for people to distinguish whether content is created by AI (Originality.ai Study).
\nWe seek to evaluate the prevalence of AI-generated articles.
\nWe observe significant growth in primarily AI-generated articles, coinciding with the launch of ChatGPT in November 2022. After only 12 months, primarily AI-generated articles accounted for 35.9% of articles published.
\nIn Q1 2025, the quantity of primarily AI-generated articles being published on the web nearly equaled the quantity of human-written articles, 49.6% vs. 50.4%. In Q4 2025, primarily AI-generated articles surpassed human-written at 50.9%, before returning to 49.9% in Q1 2026.
\nStacked time-series chart showing primarily AI-generated articles rising to ~50% of published articles by Q1 2026
\nSource data: Google Sheets
\nWhile primarily AI-generated articles grew dramatically after ChatGPT launched, we do not see that trend continuing. Instead, the proportion of primarily AI-generated articles has remained relatively stable, near 50%, over the last five quarters. We hypothesize that this is because practitioners found that primarily AI-generated articles do not perform well in search, as shown in a separate study.
\nCommon Crawl maintains one of the largest publicly available web archives. It contains billions of pages and is used by researchers and developers. It is a key data source for training large language models.
\nWe need a representative sample of English-language articles on the web. While Common Crawl does not crawl every page, its archive is the best free and publicly available proxy for the web. We want to measure the proportion of all articles being published that are primarily AI-generated, so we do not filter by traffic or use a curated subset. We randomly select 55.4k URLs from Common Crawl, and confirm that each is in English, has an article schema markup, is at least 100 words, has a publish date between January 2020 and March 2026, and is an article or listicle as classified by the Graphite page type classifier.
\nWe classify each article using three AI detectors: Pangram, Copyleaks, and GPTZero. The AI detectors produce different outputs. We provide the output of each detector, and how we transform that output into a binary, primarily AI / primarily human classification below.
\nPangram and Copyleaks provide the proportion of the article's content that is AI-generated.
\nPangram : Output: proportion of the article that is Human, AI-assisted, AI. Classify as primarily AI if: proportion AI + proportion AI-assisted > proportion Human.
\nCopyleaks : Output: proportion of the article that is Human, AI. Classify as primarily AI if: proportion AI > proportion Human.
\nIn contrast, GPTZero provides an article-level prediction. (Its Advanced Sentence Scanning output includes sentences that most impact the classification, but it does not directly provide the proportion of AI-generated content. We prefer to use its article-level output rather than devising our own method for computing the proportions.)
\nGPTZero : Output: prediction (Human, Mixed, AI) and confidence score. Classify as primarily AI if: prediction is AI or Mixed.
\nNote that the labels indicating a mixture of AI and human writing are rarely predicted on our dataset: GPTZero tags 6.4% of articles as Mixed, and Pangram tags 1.9% of articles as having AI-assisted text.
\nAccurate detection of AI-generated content is required to make claims about the prevalence of AI-generated articles on the web. There is considerable disagreement about the accuracy of AI detection algorithms, and many argue that detecting AI is impossible, or at best, highly inaccurate. Therefore, before classifying the articles in our data set, we evaluate the accuracy of the AI detectors.
\nTo evaluate the false positive rate (the percentage of human-written articles classified as primarily AI-generated), we need a dataset of human-written articles. Since the large-scale adoption of AI tools began with ChatGPT, we argue that, with high probability, articles published before its release were written by humans. Therefore, we run each detector on the 15.7k articles in our Common Crawl dataset that were published between January 2020 and November 2022. In the table below, we see that all the AI detectors have low false-positive rates.
\nFalse Positive Rates by AI Detector
\n| Detector | False Positive Rate |
|---|---|
| PPangram | 1.844% |
| CCopyleaks | 1.836% |
| 1.355% |
Source: 15.7k articles from Common Crawl (3/2026) published between 1/2020 and 11/2022. AI detections from Pangram (4/2026), Copyleaks (5/2026), GPTZero (4/2026).
\nTo evaluate the false negative rate (the percentage of primarily AI-generated articles classified as human-written), we use GPT-5, Gemini 3.1 Pro, and Claude Opus 4.6 to generate 2,000 articles using each, covering the same topics as a set of reference articles published before November 2022. For each reference article, we first generate a 100-word summary of the article using GPT-5, then we use the summary to AI-generate an article using the system prompt:
\nand user prompt:
\nAll detectors have low false negative rates, especially for GPT-5, the most popular LLM as of May 2026.
\nFalse Negative Rates by Detector and Generator Model
\n| Detector | GPT-5 | Gemini 3.1 Pro | Claude Opus 4.6 | Average |
|---|---|---|---|---|
| PPangram | 0.35% | 3.50% | 0.35% | 1.40% |
| CCopyleaks | 0.40% | 4.90% | 0.60% | 1.97% |
| 0.00% | 0.10% | 0.10% | 0.07% |
Source: 2k articles generated by each of GPT-5, Gemini 3.1 Pro, and Claude Opus 4.6. AI detections from Pangram (4/2026), Copyleaks (5/2026), GPTZero (4/2026).
\nThe raw data for this evaluation is available here.
\nFinally, we classify all 55.4k articles in our dataset using each detector to evaluate the percentage of articles that are primarily AI-generated. First, we compute the percentage of articles published in each quarter that are primarily AI-generated using each AI detector. Then, we simply take the average of those AI detector-level estimates.
\nThe raw data with classifications is available here. Note that we do not include the URLs to avoid identifying specific companies that may be publishing AI-generated articles.
\nWe previously published a study on the same topic in October 2025. The differences from our prior study are:
\nThe overall story is the same: a steep rise in primarily AI-generated articles after ChatGPT's release and a plateau near 50% more recently. However, the percentage of primarily AI-generated articles we find by using multiple detectors is slightly lower than before (3.3 percentage points, on average), due to the more robust averaging method.
\nMany people incorporate AI into their content creation process. One strategy is to ask AI to create a first draft, then have a human in the loop to edit or rewrite it. We did not evaluate the accuracy of AI detectors using this strategy.
\nAI models continue to improve, and may become harder to detect. We only evaluate the false negative rate on articles generated by GPT-5, Gemini 3.1 Pro, and Claude Opus 4.6. The AI detection algorithm may have lower accuracy when applied to articles generated by other models.
\nWe are grateful to Pangram, Copyleaks, and GPTZero for allowing us to use their AI detectors for this study.
\nWe are also grateful to Common Crawl for providing free web crawl data to researchers since 2008.
\nResults by AI detector:
\nPer-detector time-series comparing Pangram, Copyleaks, and GPTZero classifications across quarters
\nSource data: Google Sheets
\nJose Luis Paredes, PhDSenior Data Specialist
Senior Data Scientist at Graphite. Ph.D. and Master's from the University of Delaware, followed by more than two decades as a professor at the University of Los Andes. Author of over 50 peer-reviewed research papers and holder of 5 U.S. patents. Previously Head of Data Science at GoToDigital.
\n\nGregory Druck, PhDChief AI Officer
Chief AI Officer at Graphite, leading the team building AI tools for growth and researching how AI is reshaping marketing. Previously Chief Data Scientist at Yummly and an NLP and search researcher at Yahoo! Research, with internships at Google and Microsoft. Ph.D. in machine learning from UMass Amherst, advised by Andrew McCallum.
\n\n\nGrowth advisor focused on AEO, SEO, and AI based marketing, previously in operations at Tesla.
\n\n\nFounder and CEO of Graphite, the research-driven growth agency behind work for Webflow, Adobe, and Upwork. Teaches SEO and AEO at Reforge and is an adjunct professor at IE Business School. Research published in the Financial Times, Axios, and The Atlantic. Previously a growth advisor to Masterclass, Robinhood, and Honey.
\n","headings":[{"level":1,"text":"AI Now Writes as Many Online Articles as Humans","id":"ai-now-writes-as-many-online-articles-as-humans"},{"level":2,"text":"Motivation","id":"motivation"},{"level":2,"text":"Results","id":"results"},{"level":3,"text":"Primarily AI-Generated Article Growth Has Plateaued","id":"primarily-ai-generated-article-growth-has-plateaued"},{"level":2,"text":"Methodology","id":"methodology"},{"level":3,"text":"Common Crawl","id":"common-crawl"},{"level":3,"text":"Selection of Articles","id":"selection-of-articles"},{"level":3,"text":"AI Detection","id":"ai-detection"},{"level":3,"text":"Evaluation of False Positive Rates","id":"evaluation-of-false-positive-rates"},{"level":3,"text":"Evaluation of False Negative Rates","id":"evaluation-of-false-negative-rates"},{"level":3,"text":"Quantifying Primarily AI-Generated Articles on the Web","id":"quantifying-primarily-ai-generated-articles-on-the-web"},{"level":3,"text":"Comparison with Our Prior Study","id":"comparison-with-our-prior-study"},{"level":2,"text":"Limitations","id":"limitations"},{"level":3,"text":"AI-Assisted Articles","id":"ai-assisted-articles"},{"level":3,"text":"AI Models","id":"ai-models"},{"level":3,"text":"Acknowledgements","id":"acknowledgements"},{"level":2,"text":"Appendix","id":"appendix"},{"level":2,"text":"Researchers","id":"researchers"}]}}