AI-generated vs human content statistics at a glance
| Category | Statistic | Source |
|---|---|---|
| Volume | 49.9% of new English articles were primarily AI-generated in Q1 2026, stable near 50% for five quarters | Graphite, May 2026 |
| Volume | 74.2% of 900,000 new web pages created in April 2025 contained some AI content; 2.5% were pure AI | Ahrefs, May 2025 |
| Volume | Up to 24% of corporate press release text was LLM-assisted by late 2024 | Liang et al., Stanford, 2025 |
| Volume | NewsGuard has identified 3,749 AI content farm news sites in 16 languages as of June 2026 | NewsGuard, 2026 |
| Rankings | 14% of articles ranking in Google are AI-generated; 86% are human-written (June 2025 data) | Graphite, October 2025 |
| Rankings | 80.5% of position-1 pages are human-written vs 10% AI-generated, across 42,000 blog pages | Semrush, April 2026 |
| Rankings | Correlation between a page's AI share and its Google position is 0.011, effectively zero | Ahrefs, July 2025 |
| Citations | 18% of articles cited by ChatGPT and Perplexity are AI-generated | Graphite, October 2025 |
| Google cut low-quality, unoriginal content in results by 45% after its March 2024 update | Google, April 2024 | |
| Detection | Seven detectors flagged 61.22% of human-written TOEFL essays as AI-generated on average | Liang et al., Stanford, 2023 |
| Readers | 53% of US adults are not confident they can tell AI content from human content | Pew Research Center, September 2025 |
| Readers | 12% are comfortable with news made entirely by AI; 62% with news made entirely by humans | Reuters Institute, October 2025 |
| Readers | An "AI-generated" label cut perceived accuracy by 2.66 points, a third of the effect of a "false" label | Altay and Gilardi, PNAS Nexus, 2024 |
| Producers | 95% of B2B marketers say their organisation uses AI applications | Content Marketing Institute, October 2025 |
How much new web content is AI-generated?
About half of new English-language web articles are primarily AI-generated in 2026, and that share has sat near 50% since early 2025. The answer moves a lot with the definition, though: count any AI involvement and the figure passes 70%; count only untouched AI output and it falls under 3%.
| Statistic | Source |
|---|---|
| 49.9% of new English-language articles published in Q1 2026 were primarily AI-generated, against 50.9% in Q4 2025 and 49.6% in Q1 2025. Graphite randomly sampled 55,400 English-language articles from Common Crawl, published January 2020 to March 2026, and averaged three detectors (Pangram, Copyleaks, GPTZero). | Graphite, May 2026 |
| Twelve months after ChatGPT launched, 35.9% of new articles were already primarily AI-generated. | Graphite, May 2026 |
| Graphite's first study, which used one detector, put the moment AI articles overtook human ones at November 2024. The 2026 rerun with three detectors came in 3.3 percentage points lower. | Graphite, October 2025 and May 2026 |
| 74.2% of 900,000 new English web pages created in April 2025 contained AI-generated content: 71.7% mixed human and AI, 2.5% pure AI, meaning Ahrefs' detector flagged the whole page. | Ahrefs, May 2025 |
| Only 25.8% of the same 900,000 new pages were purely human-written, measured with Ahrefs' own detector, one page per domain. | Ahrefs, May 2025 |
| By late 2024, LLMs wrote or modified up to 24% of corporate press release text, about 18% of US financial consumer complaint text, nearly 14% of UN press release content and just under 10% of job posting text at small firms. Adoption stabilised during 2024. | Liang et al., Stanford, February 2025 |
| At least 13.5% of 2024 biomedical abstracts in PubMed were processed with large language models, rising to 40% in some fields and countries, from a corpus of over 15 million abstracts. | Kobak et al., Science Advances, July 2025 |
| Over 5% of newly created English Wikipedia articles in August 2024 were flagged as AI-generated, with detector thresholds set for a 1% false positive rate. | Brooks et al., Princeton, October 2024 |
| NewsGuard has identified 3,749 AI content farm news and information websites in 16 languages, sites that publish mostly AI text with little human oversight and no disclosure. | NewsGuard AI Tracking Center, June 2026 |
The plateau is the finding I'd underline. The AI share of new articles jumped for 18 months after ChatGPT and then stopped near half, in Graphite's data and in Stanford's study of press releases and complaints. That looks less like a flood and more like a market settling: publishers who wanted cheap text switched early, and the rest decided they didn't. The next question is whether any of that text reaches a reader.
Does AI-generated content rank as well as human content in Google?
AI-generated articles rank far less often than their publishing volume predicts: Graphite found they are 14% of the articles in Google results and 7% of position-1 results, while roughly half of new articles are AI-generated. The picture changes with definitions again, because most ranking pages contain some AI text.
| Statistic | Source |
|---|---|
| 86% of articles ranking in Google Search are human-written and 14% are AI-generated, across 31,493 keywords in 10 categories, collected in June 2025 with Surfer's detector. | Graphite, October 2025 |
| Only 7% of position-1 articles in the same study were AI-generated, half the 14% baseline. Human-written pages ranked higher with statistical significance (p < 1e-6). | Graphite, October 2025 |
| A page in position 1 has an 80.5% probability of being human-written and a 10% probability of being AI-generated, across 42,000 blog pages ranking for 20,000 keywords. | Semrush, April 2026 |
| 86.5% of 600,000 pages ranking in Google's top 20 for 100,000 keywords contain some AI-generated content. | Ahrefs, July 2025 |
| 4.6% of those top-20 pages are pure AI, 13.5% pure human and 81.9% a mix. | Ahrefs, July 2025 |
| The correlation between a page's share of AI content and its ranking position is 0.011, which Ahrefs calls effectively zero. | Ahrefs, July 2025 |
| 17.31% of Google's top 20 results for 500 informational keywords were AI-generated in September 2025, up from 2.27% in February 2019. This is vendor data from a detector company measuring with its own tool. | Originality.ai, September 2025 |
The studies don't contradict each other once you read the definitions: Ahrefs counts any page with a trace of AI and finds no penalty for using it. Graphite and Semrush count pages that are mostly AI and find those lose at the top. My read: AI as a drafting tool is invisible to rankings, and AI as the author is visible, because the output tends to say what 50 other pages already say. Getting cited by an AI assistant follows the same pattern.
How often do ChatGPT and Perplexity cite AI-written pages?
ChatGPT and Perplexity cite AI-generated articles slightly more often than Google ranks them: 18% of the articles they cite are AI-generated, against 14% of the articles in Google results, in Graphite's October 2025 study.
| Statistic | Source |
|---|---|
| 82% of articles cited by ChatGPT are human-written and 18% AI-generated. Perplexity shows the same 82/18 split. | Graphite, October 2025 |
| Graphite sampled 100 keywords per category for the answer engine test and classified each cited article in 500-word chunks. | Graphite, October 2025 |
Eighteen percent is still well under the roughly 50% that AI text holds in new publishing. Answer engines pull from the same pool of pages that rank, so they inherit Google's preference for human-written sources. If you want the numbers on what does get cited, the generative engine optimization statistics page covers citation share, and the original research content statistics page carries the academic GEO experiments in full. Google's own rules explain part of why AI text struggles.
What does Google say about AI-generated content?
Google Search does not penalise content for being AI-generated; its spam policies target content made at scale to manipulate rankings, "no matter whether content is produced through automation, human efforts," or both, in Google's March 2024 wording.
| Statistic | Source |
|---|---|
| On 8 February 2023, Google wrote that its ranking systems reward original, high-quality content "however it is produced", and that "Appropriate use of AI or automation is not against our guidelines." Using AI mainly to manipulate rankings breaks its spam policies. The post is dated, from 2023, and still live on Search Central. | Google Search Central Blog, February 2023 |
| On 5 March 2024, Google introduced the scaled content abuse policy, defined as many pages generated "for the primary purpose of manipulating Search rankings and not helping users." It applies to automation, humans or a mix. | Google Search Central Blog, March 2024 |
| Google expected the March 2024 core update and spam work to reduce low-quality, unoriginal content in results by 40%. When the rollout finished on 19 April 2024, it reported a 45% reduction. | Google Keyword blog, March 2024, updated 26 April 2024 |
Google's current documentation on generative AI content adds one practical line: using AI to "generate many pages without adding value for users" may violate the scaled content abuse policy (Google Search Central documentation).
Read the three documents in date order and the position hasn't moved since 2023. Google never promised to hunt AI text. It promised to hunt scale without value, and the 45% figure is what that looked like in practice. The hard part is that nobody outside Google can see how it tells the two apart, and the public tools that try get it wrong often enough to matter.
How accurate are AI content detectors?
AI content detectors range from near-perfect to unusable depending on the tool and the writer: independent academic tests found false positive rates from below 1% for the best commercial detector to 61% for non-native English essays across seven popular tools. Detector vendors grade themselves, so the academic numbers below carry more weight than any vendor's own accuracy claim.
| Statistic | Source |
|---|---|
| Seven widely used GPT detectors misclassified 61.22% of 91 human-written TOEFL essays as AI-generated on average, while classifying native speakers' essays accurately. | Liang et al., Stanford, Patterns, July 2023 |
| 97.8% of those TOEFL essays were flagged as AI by at least one detector, and 19.78% by all seven. | Liang et al., Stanford, July 2023 |
| A test of 14 detection tools, including Turnitin and PlagiarismCheck, concluded they are "neither accurate nor reliable" and lean towards labelling AI text as human. Paraphrasing made them worse. The test is dated, from 2023, and predates current detector versions. | Weber-Wulff et al., International Journal for Educational Integrity, 2023 |
| On 1,992 human passages across six genres, the open-source RoBERTa detector labelled more than 90% of human passages as AI even at its strictest threshold and missed up to 51% of AI passages. | Jabarian and Imas, University of Chicago, NBER Working Paper, September 2025 |
| In the same test, Pangram reached essentially zero false positives and false negatives on medium and long passages; GPTZero and Originality.ai formed a second tier. | Jabarian and Imas, NBER, September 2025 |
| Surfer's detector misclassified 4.2% of human articles as AI and 0.6% of AI articles as human in Graphite's validation. This is a detector tested by the research firm using it. | Graphite, October 2025 |
| In Graphite's own validation of the three detectors it uses, false positive rates on 15,700 articles published before ChatGPT were 1.84% for Pangram, 1.84% for Copyleaks and 1.36% for GPTZero. False negative rates on 2,000 articles from each of three AI models averaged 0.07% to 1.97%. | Graphite, May 2026 |
A 4.2% false positive rate sounds small until you apply it: that's 42 wrongly flagged writers in every 1,000. The Stanford result is dated, from 2023, and detectors have improved, but the lesson hasn't aged: detectors fail hardest on plain, simple prose, which is how non-native writers and many good editors write. I'd use them to estimate shares across 50,000 pages, which is what Graphite does, and never to judge one person's article. Readers, it turns out, are worse at this than the tools.
Can readers tell AI-written text from human writing?
Most readers cannot reliably tell AI-written text from human writing: 53% of US adults say they are not confident they could, according to Pew Research Center's June 2025 survey, and in blind tests non-experts identify AI poems at below-chance rates.
| Statistic | Source |
|---|---|
| 76% of US adults say it is extremely or very important to be able to tell whether pictures, videos and text were made by AI or by people. The survey covered 5,023 adults in June 2025. | Pew Research Center, September 2025 |
| 53% of US adults are not too or not at all confident they can detect whether content was made by AI or a person. | Pew Research Center, September 2025 |
| Non-expert readers identified AI-generated poems with 46.6% accuracy, below chance, and judged AI poems as human-written more often than real human poems. They rated the AI poems higher for rhythm and beauty. | Porter and Machery, Scientific Reports, November 2024 |
Put the two Pew numbers side by side and you get the whole problem: three in four Americans want to know, and half admit they can't tell. That gap is why labels matter so much, and why what a label does to trust deserves its own section.
How do readers respond when they know content is AI-generated?
Readers trust content less once they are told or believe it is AI-generated, even when the content is true or human-written: in a Reuters Institute survey of six countries, 12% were comfortable with news made entirely by AI against 62% for news made entirely by humans.
| Statistic | Source |
|---|---|
| 12% of people across six countries are very or somewhat comfortable with news made entirely by AI, rising to 21% with human oversight. | Reuters Institute, Generative AI and News Report, October 2025 |
| 43% are comfortable with news made mainly by humans with some AI help, and 62% with news made entirely by humans. YouGov fielded the survey from 5 June to 15 July 2025, about 2,000 people per country in Argentina, Denmark, France, Japan, the UK and the US. | Reuters Institute, October 2025 |
| Labelling a headline "AI-generated" lowered its perceived accuracy and sharing intent by 2.66 percentage points, whether the headline was true, false or written by a human. Labelling it "false" had an effect of 9.33 points. Two experiments, 4,976 US and UK participants. | Altay and Gilardi, PNAS Nexus, October 2024 |
| Altay and Gilardi traced the penalty to an assumption: people read the label as full automation with no human supervision. Explaining that AI only helped removed the effect. | Altay and Gilardi, PNAS Nexus, 2024 |
| In a nationally representative sample of 3,861, an AI label on a policy news article reduced its perceived accuracy but had no significant effect on support for the policy. Explaining how AI was used shrank the penalty. | Wang, Sturgis and de Kadt, arXiv, June 2025, revised February 2026 |
| People who thought an article was AI-generated rated it 48% less trustworthy, 57% less authentic and 60% lower on emotional connection, whether or not AI wrote it. Ads next to suspected AI content saw a 14% drop in purchase consideration. Vendor data: a study of 3,000 US adults commissioned by Raptive. | Raptive, August 2025 |
Three separate experiments agree on something awkward for anyone who labels content: the penalty attaches to the label and the suspicion, not to the text. In the Scientific Reports poetry study, readers preferred the AI poems as long as they didn't know which was which. The fix is to say what the human did, because every study that explained the human role saw the penalty shrink. Marketers mostly haven't absorbed that yet, judging by how they describe their own use.
How many marketers use AI to write content?
Nearly every content team now uses AI in some form: 87% of 879 content marketers in Ahrefs' survey use it to create or help create content, and 95% of B2B marketers say their organisation uses AI applications, according to the Content Marketing Institute's 2026 research.
| Statistic | Source |
|---|---|
| 87% of 879 surveyed content marketers use AI to create or help create content; 13% don't use it. | Ahrefs, May 2025 |
| 95% of 1,015 B2B marketers say their organisation uses AI-powered applications, and 89% of those use AI tools for written copy. The survey ran from June to August 2025. | Content Marketing Institute and MarketingProfs, October 2025 |
| 87% of SEO teams say their content is fully human-created or heavily human-led, and only 19% say AI improves content quality, in a survey of 224 SEO professionals. | Semrush, April 2026 |
| 72% of SEO professionals who use AI content say AI-assisted content ranks at least as well as human-written content, up from 64% in Semrush's 2024 study. | Semrush, April 2026 |
One Semrush report holds both sides: 72% of SEOs who use AI content believe it ranks as well, and the same Semrush report found position 1 is 80.5% human-written. Both can be true, because "AI content" in the survey mostly means human-led work with AI help. The practitioners are describing the mixed pages Ahrefs measured, and those do rank. Fully automated output is the part the data punishes, and that's also the part old content ages worst in, as the content decay statistics page shows.
How we calculated the original numbers
Four figures on this page do not appear in any of the sources. Here is the arithmetic, so you can check it or swap the inputs.
- 3.5x: the ranking gap. Graphite's May 2026 study puts the AI share of new English articles at 49.6% in Q1 2025, the quarter closest to the June 2025 search sample. Graphite's October 2025 search study found AI-generated articles are 14% of Google results and 7% of position-1 results. 49.6 / 14 = 3.5, and 49.6 / 7 = 7.1. Both studies come from Graphite, but they use different samples and detectors: the search study used Surfer alone, the publishing study averaged three. Google's results also include articles of every age, many written before ChatGPT, while the 49.6% covers only new articles. Part of the gap is age, not quality, so treat the ratio as the size of the gap, not a precise multiplier.
- 49.6%: the error-corrected publishing share for Q1 2026. The error rates come from Graphite's own validation of the three detectors it uses, not from an independent test. Averaged, the detectors show a 1.68% false positive rate ((1.844 + 1.836 + 1.355) / 3, on 15,700 pre-ChatGPT articles) and a 1.15% false negative rate ((1.40 + 1.97 + 0.07) / 3, on AI articles from three models). Correcting the measured Q1 2026 share of 49.9% for both errors gives (0.499 - 0.0168) / (1 - 0.0115 - 0.0168) = 49.6%. That it matches Graphite's measured Q1 2025 figure is a coincidence. Detector error moves the headline by 0.3 points; the definition (a majority of an article's text flagged as AI) moves it far more.
- 30x: the definition swing. In Ahrefs' April 2025 sample of 900,000 new pages, 74.2% contained some AI content (any share of the page flagged) and 2.5% were pure AI (the whole page flagged): 74.2 / 2.5 = 29.7. Both figures come from the same study, sample and month. For Ahrefs' ranking pages the same split is 86.5% against 4.6%, a 19x swing. The question "how much content is AI?" has answers 30 times apart inside one dataset, depending on the threshold.
- 9 points: what human oversight buys back. The Reuters Institute's Generative AI and News Report 2025 asked people in Argentina, Denmark, France, Japan, the UK and the US how comfortable they are with news made in different ways. 12% are very or somewhat comfortable with news made entirely by AI and 62% with news made entirely by humans, a 50-point comfort gap. Adding human oversight to AI-made news lifts comfort to 21%, which closes only 9 points of that 50-point gap. News made mainly by a human with some AI help reaches 43%, closing 31 points. Same survey, same six-country averages, so the comparison holds.
What the 2026 numbers say
Volume stopped growing a year and a half ago. Graphite's share has sat near 50% for five quarters, and Stanford's press release and complaint data flattened in 2024 as well. Anyone projecting an internet that is 90% AI by some near date is extrapolating a curve that already bent.
Publishing and being found are different games. AI text is half of what gets published and 14% of what Google ranks, 7% at position 1, 18% of what ChatGPT cites. Google says it doesn't care who wrote the page, and I believe that, because it doesn't need to care. Scaled, interchangeable text loses on its own merits, and the 45% cleanup in 2024 removed a lot of it.
The trust penalty is about the label, so write the label well. Readers can't spot AI text, they want to know, and a bare "AI-generated" tag costs credibility even on true, human-written headlines. The one lever every study agrees on is explaining what the person did. The byline and a line on process now do more work than the model choice.
FAQ
What percentage of new online content is AI-generated?
49.9% of new English-language web articles were primarily AI-generated in Q1 2026, according to Graphite's study of 55,400 Common Crawl URLs. The share has held near 50% since early 2025. Counting any AI involvement instead, Ahrefs found 74.2% of new pages in April 2025 contained some AI text.
Does AI-generated content rank lower in Google?
14% of articles in Google results are AI-generated, against roughly half of new articles, according to two Graphite studies (October 2025 and May 2026) with different samples. That's a 3.5x gap by our calculation, and at position 1 the AI share falls to 7%. Ahrefs found no correlation (0.011) between a page's AI share and its position when mixed pages are included.
Does Google penalize AI-generated content?
Since 8 February 2023, Google Search Central has said that appropriate use of AI "is not against our guidelines," so AI text is not penalised as such. Its March 2024 scaled content abuse policy targets pages made mainly to manipulate rankings, whether by automation, humans or both.
How accurate are AI content detectors?
61.22% of human-written essays by non-native English speakers were flagged as AI by seven detectors in a 2023 Stanford study, while a 2025 University of Chicago test (NBER) found Pangram made almost no errors on longer passages. Vendor figures are self-reported, so check an independent test before trusting one.
Can people tell AI writing from human writing?
53% of US adults are not confident they can, according to Pew Research Center (June 2025). In a Scientific Reports study, non-experts identified AI poems with 46.6% accuracy, below chance, and preferred the AI poems.
Do readers trust AI-generated content less?
12% of people are comfortable with news made entirely by AI, against 62% for news made entirely by humans, in the Reuters Institute's six-country survey (2025). Adding human oversight lifts comfort only to 21%. Experiments show the penalty falls when publishers explain what the human did.
Sources
- Graphite, AI Now Writes as Many Online Articles as Humans Do (May 2026)
- Graphite, More Articles Are Now Created by AI Than Humans (October 2025, superseded by source 1)
- Graphite, AI Content in Search and LLMs (October 2025)
- Ahrefs, What Percentage of New Content Is AI-Generated? (May 2025)
- Ahrefs, AI-Generated Content Does Not Hurt Your Google Rankings (July 2025)
- Semrush, Does AI Content Rank in Search? Survey and Data Study (April 2026)
- Originality.ai, AI Content in Google Search Results (vendor tracker, September 2025)
- Google Search Central Blog, Google Search's guidance about AI-generated content (February 2023)
- Google Search Central Blog, What web creators should know about our March 2024 core update and new spam policies (March 2024)
- Google, New ways we're tackling spammy, low-quality content on Search (March 2024, updated April 2024)
- Google Search Central, Guidance on using generative AI content on your website (accessed September 2026)
- Liang et al., The Widespread Adoption of Large Language Model-Assisted Writing Across Society (February 2025)
- Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances (July 2025)
- Brooks, Eggert and Peskoff, The Rise of AI-Generated Content in Wikipedia (October 2024)
- NewsGuard, AI Tracking Center (June 2026)
- Liang et al., GPT detectors are biased against non-native English writers, Patterns (July 2023)
- Weber-Wulff et al., Testing of Detection Tools for AI-Generated Text, International Journal for Educational Integrity (2023)
- Jabarian and Imas, Artificial Writing and Automated Detection, NBER Working Paper 34223 (September 2025)
- Pew Research Center, How Americans View AI and Its Impact on People and Society (September 2025)
- Porter and Machery, AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably, Scientific Reports (November 2024)
- Reuters Institute, Generative AI and News Report 2025 (October 2025)
- Altay and Gilardi, People are skeptical of headlines labeled as AI-generated, PNAS Nexus (October 2024)
- Wang, Sturgis and de Kadt, AI labeling reduces the perceived accuracy of online content but has limited broader effects (June 2025, revised February 2026)
- Raptive, The "AI stink" is real, and it's costing brands (August 2025)
- Content Marketing Institute and MarketingProfs, B2B Content and Marketing Trends: Insights for 2026 (October 2025)
Methodology. Numbers were collected in September 2026 and checked against the original publisher's page, paper or PDF, not against statistics roundups. Detector companies' figures (Originality.ai, and detector error rates reported by firms that use a detector) are labelled as vendor data, and independent academic tests from Stanford, the University of Chicago and the International Journal for Educational Integrity are shown beside them. Google's position is quoted from dated Search Central posts. Where studies define "AI content" differently (any AI involvement versus mostly AI), the definition is stated in the row. Stats older than 24 months are flagged in the text. Last updated 24 September 2026.