| Published estimates | 2020 | 2.48 | percent | rawValue: 2.48%; source: https://docs.google.com/spreadsheets/d/1WamFyVahPDtAPFtvly30BG2QjyA-L1KYkO2UEYGKKcg/edit?gid=0; notes: Graphite estimated that 2.48% of newly published English web articles in 2020 were AI-generated. It filtered Common Crawl pages by article markup, English language, and publish date, then classified them with Surfer's AI detector using a greater-than-50% threshold. This is one of the closest direct measures in the table, but it only covers newly published English articles and depends on detector performance.; publicationDate: 2024-05-07 |
| Published estimates | 2020 | 5.1 | percent | rawValue: 5.10%; source: https://originality.ai/blog/ai-climate-change-paper-abstracts; notes: This study found 5.10% likely AI-generated climate change paper abstracts in 2020. It pulled up to 500 OpenAlex abstracts per year matching 'climate change' and scored 50+ word abstracts with the Originality.ai API. It is keyword-based, limited to sampled abstracts, and relies on a vendor detector rather than disclosure data.; publicationDate: 2025-10-11 |
| Published estimates | 2020 | 0.0 | percent | rawValue: 0%; source: https://originality.ai/blog/ai-lawyer-office-reviews; notes: The study estimates 0% AI-generated lawyer office reviews in the United States for 2020 among 100-plus-word reviews. It scraped Google Places data for 49 city-state pairs and flagged reviews over a 0.5 Originality.ai score. It covers one review corpus only and depends on a vendor detection model.; publicationDate: 2025-11-01 |
| Published estimates | 2020 | 3.06 | percent | rawValue: 3.06%; source: https://originality.ai/blog/ai-quora-answers-study; notes: Originality.AI estimated that 3.06% of English-language Quora answers in 2020 were likely AI-generated. It scraped public Quora answers across multiple topics, kept answers of at least 50 words, and ran them through its detector. This is a direct platform-specific answer estimate, but it depends on scraper coverage and detector performance.; publicationDate: 2025-10-23 |
| Published estimates | 2020 | 0.79 | percent | rawValue: 0.79%; source: https://originality.ai/blog/ai-us-bank-reviews; notes: The study estimates 0.79% AI-generated customer reviews for United States banks in 2020. It analyzed over 19000 reviews from 47 banks and classified them with Originality.ai. It covers one banking corpus only and uses a vendor detector with very low estimated rates.; publicationDate: 2025-10-19 |
| Published estimates | 2020 | 4.55 | percent | rawValue: 4.55%; source: https://originality.ai/blog/ai-walmart-reviews; notes: Originality.AI estimated that 4.55% of Walmart product reviews in 2020 were likely AI-generated. It analyzed Walmart mobile-app reviews of at least 50 words and classified them with its detector across a year-by-year series. This is a detector-based estimate inside one retail-review platform, not a broad web-content measure.; publicationDate: 2025-10-19 |
| Published estimates | 2021 | 3.06 | percent | rawValue: 3.06%; source: https://docs.google.com/spreadsheets/d/1WamFyVahPDtAPFtvly30BG2QjyA-L1KYkO2UEYGKKcg/edit?gid=0; notes: Graphite estimated that 3.06% of newly published English web articles in 2021 were AI-generated. It filtered Common Crawl pages by article markup, English language, and publish date, then classified them with Surfer's AI detector using a greater-than-50% threshold. This is one of the closest direct measures in the table, but it only covers newly published English articles and depends on detector performance.; publicationDate: 2024-05-07 |
| Published estimates | 2021 | 16.0 | percent | rawValue: 16.00%; source: https://originality.ai/blog/ai-climate-change-paper-abstracts; notes: This study found 16.00% likely AI-generated climate change paper abstracts in 2021. It pulled up to 500 OpenAlex abstracts per year matching 'climate change' and scored 50+ word abstracts with the Originality.ai API. It is keyword-based, limited to sampled abstracts, and relies on a vendor detector rather than disclosure data.; publicationDate: 2025-10-11 |
| Published estimates | 2021 | 0.0 | percent | rawValue: 0%; source: https://originality.ai/blog/ai-lawyer-office-reviews; notes: The study estimates 0% AI-generated lawyer office reviews in the United States for 2021 among 100-plus-word reviews. It scraped Google Places data for 49 city-state pairs and flagged reviews over a 0.5 Originality.ai score. It covers one review corpus only and depends on a vendor detection model.; publicationDate: 2025-11-01 |
| Published estimates | 2021 | 1.37 | percent | rawValue: 1.37%; source: https://originality.ai/blog/ai-us-bank-reviews; notes: The study estimates 1.37% AI-generated customer reviews for United States banks in 2021. It analyzed over 19000 reviews from 47 banks and classified them with Originality.ai. It covers one banking corpus only and uses a vendor detector with very low estimated rates.; publicationDate: 2025-10-19 |
| Published estimates | 2021 | 5.19 | percent | rawValue: 5.19%; source: https://originality.ai/blog/ai-walmart-reviews; notes: Originality.AI estimated that 5.19% of Walmart product reviews in 2021 were likely AI-generated. It analyzed Walmart mobile-app reviews of at least 50 words and classified them with its detector across a year-by-year series. This is a detector-based estimate inside one retail-review platform, not a broad web-content measure.; publicationDate: 2025-10-19 |
| Published estimates | 2022 | 1.77 | percent | rawValue: 1.77%; source: https://arxiv.org/abs/2412.18148; notes: The study-year median across Medium, Quora, and Reddit put AI-generated social-media posts at 1.77% in 2022. The paper trained and selected an AIGT detector, then applied it to a 2.4 million post corpus spanning January 2022 to October 2024. This gives one value for a major social-content dataset, but it remains platform-bound and detector-dependent.; publicationDate: 2024-12-24 |
| Published estimates | 2022 | 4.95 | percent | rawValue: 4.95%; source: https://pubmed.ncbi.nlm.nih.gov/41016415/; notes: This study estimated a mean AI publication score of 4.95% for ophthalmology manuscripts in 2022. It used a retrospective bibliographic analysis of ophthalmology manuscripts and computed an AI-derived prevalence score across the corpus. This is a detector-like proxy in one medical specialty and is narrower than general scientific publishing or the web.; publicationDate: 2025-10-10 |
| Published estimates | 2022 | 6.81 | percent | rawValue: 6.81%; source: https://docs.google.com/spreadsheets/d/1WamFyVahPDtAPFtvly30BG2QjyA-L1KYkO2UEYGKKcg/edit?gid=0; notes: Graphite estimated that 6.81% of newly published English web articles in 2022 were AI-generated. It filtered Common Crawl pages by article markup, English language, and publish date, then classified them with Surfer's AI detector using a greater-than-50% threshold. This is one of the closest direct measures in the table, but it only covers newly published English articles and depends on detector performance.; publicationDate: 2024-05-07 |
| Published estimates | 2022 | 0.0 | percent | rawValue: 0.0%; source: https://arxiv.org/abs/2603.19316; notes: The January 2022 baseline put significant AI-generated text at 0.0% of JAMA Network Open articles. The study uses Originality.AI on a monthly series covering original investigations, research letters, and invited commentaries from January 2022 through March 2025. This is a narrow but still direct content-share estimate for one medical publishing venue.; publicationDate: 2026-03-15 |
| Published estimates | 2022 | 13.0 | percent | rawValue: 13.00%; source: https://originality.ai/blog/ai-climate-change-paper-abstracts; notes: This study found 13.00% likely AI-generated climate change paper abstracts in 2022. It pulled up to 500 OpenAlex abstracts per year matching 'climate change' and scored 50+ word abstracts with the Originality.ai API. It is keyword-based, limited to sampled abstracts, and relies on a vendor detector rather than disclosure data.; publicationDate: 2025-10-11 |
| Published estimates | 2022 | 6.18 | percent | rawValue: 6.18%; source: https://originality.ai/blog/ai-glassdoor-reviews; notes: This study found 6.18% likely AI-written Glassdoor review 'Pros' text for S&P 500 companies in 2022. It aggregated S&P 500 company reviews and ran Originality.ai on Pros fields longer than 50 words. It measures only one section of employee reviews and depends on a vendor detector.; publicationDate: 2025-10-19 |
| Published estimates | 2022 | 2.04 | percent | rawValue: 2.04%; source: https://originality.ai/blog/ai-lawyer-office-reviews; notes: The study estimates 2.04% AI-generated lawyer office reviews in the United States for 2022 among 100-plus-word reviews. It scraped Google Places data for 49 city-state pairs and flagged reviews over a 0.5 Originality.ai score. It covers one review corpus only and depends on a vendor detection model.; publicationDate: 2025-11-01 |
| Published estimates | 2022 | 6.93 | percent | rawValue: 6.93%; source: https://originality.ai/blog/ai-quora-answers-study; notes: Originality.AI estimated that 6.93% of English-language Quora answers in 2022 were likely AI-generated. It scraped public Quora answers across multiple topics, kept answers of at least 50 words, and ran them through its detector. This is a direct platform-specific answer estimate, but it depends on scraper coverage and detector performance.; publicationDate: 2025-10-23 |
| Published estimates | 2022 | 1.85 | percent | rawValue: 1.85%; source: https://originality.ai/blog/ai-us-bank-reviews; notes: The study estimates 1.85% AI-generated customer reviews for United States banks in 2022. It analyzed over 19000 reviews from 47 banks and classified them with Originality.ai. It covers one banking corpus only and uses a vendor detector with very low estimated rates.; publicationDate: 2025-10-19 |
| Published estimates | 2022 | 5.4 | percent | rawValue: 5.40%; source: https://originality.ai/blog/ai-walmart-reviews; notes: Originality.AI estimated that 5.40% of Walmart product reviews in 2022 were likely AI-generated. It analyzed Walmart mobile-app reviews of at least 50 words and classified them with its detector across a year-by-year series. This is a detector-based estimate inside one retail-review platform, not a broad web-content measure.; publicationDate: 2025-10-19 |
| Published estimates | 2022 | 10.3 | percent | rawValue: 10.30%; source: https://pubmed.ncbi.nlm.nih.gov/37776949/; notes: The mean AI-generation score was 10.30% for OTSR orthopedics articles published before November 2022. The study ran ZeroGPT on the body plus abstract of 425 articles published from February 2022 to September 2023. This is a direct estimate within a defined journal corpus, though detector choice and single-journal scope are clear limitations.; publicationDate: 2023-09-29 |
| Published estimates | 2022 | 0.1 | percent | rawValue: 0.10%; source: https://arxiv.org/abs/2602.00319; notes: The paper estimated that 0.10% of ICLR 2022 peer reviews were AI-generated. It trained a detector on real and synthetic 2021 ICLR reviews and applied it to 1,937 2022 reviews. The estimate is detector-based and likely misses lighter AI editing.; publicationDate: 2026-01-30 |
| Published estimates | 2023 | 13.4 | percent | rawValue: 13.4%; source: https://papers.ssrn.com/sol3/Delivery.cfm/5226562.pdf?abstractid=5226562&mirid=1; notes: The paper reports that AI-generated articles rose to 13.4% of all Seeking Alpha articles after ChatGPT’s launch. It tracks firm-specific financial-analysis articles on Seeking Alpha and estimates AI use over time in a platform-specific corpus. The estimate is narrow to one investing platform and the abstract does not fully expose the identification procedure.; publicationDate: 2025-05-06 |
| Published estimates | 2023 | 31.61 | percent | rawValue: 31.61%; source: https://docs.google.com/spreadsheets/d/1WamFyVahPDtAPFtvly30BG2QjyA-L1KYkO2UEYGKKcg/edit?gid=0; notes: Graphite estimated that 31.61% of newly published English web articles in 2023 were AI-generated. It filtered Common Crawl pages by article markup, English language, and publish date, then classified them with Surfer's AI detector using a greater-than-50% threshold. This is one of the closest direct measures in the table, but it only covers newly published English articles and depends on detector performance.; publicationDate: 2024-05-07 |
| Published estimates | 2023 | 10.0 | percent | rawValue: 10.0%; source: LLM-Assisted writing in engineering is associated with higher citations as evidenced from 1.17 million papers; notes: The study estimates that 10.0% of engineering abstracts published in 2023 exceeded a pre-LLM 95th-percentile marker threshold across five engineering fields plus computer science benchmark data. It sampled 1.17 million abstracts from 2019 to 2025 and inferred LLM assistance from shifts in a 228-marker lexicon rather than article-level detector labels. This is a statistical proxy for likely LLM assistance in abstracts, not a direct classification of whole papers.; publicationDate: 2026-07-24 |
| Published estimates | 2023 | 17.2 | percent | rawValue: 17.20%; source: https://originality.ai/blog/ai-climate-change-paper-abstracts; notes: This study found 17.20% likely AI-generated climate change paper abstracts in 2023. It pulled up to 500 OpenAlex abstracts per year matching 'climate change' and scored 50+ word abstracts with the Originality.ai API. It is keyword-based, limited to sampled abstracts, and relies on a vendor detector rather than disclosure data.; publicationDate: 2025-10-11 |
| Published estimates | 2023 | 12.0 | percent | rawValue: 12.0%; source: https://originality.ai/blog/ai-forbes-business-council-study; notes: The study found that 12.0% of Forbes Business Council posts published in 2023 were likely AI-generated. It scanned 3,906 2023 posts with Originality.ai’s detector and counted 469 as likely AI. The estimate is specific to one paid contributor platform and depends on detector thresholds.; publicationDate: 2025-12-02 |
| Published estimates | 2023 | 10.86 | percent | rawValue: 10.86%; source: https://originality.ai/blog/ai-generated-content-dominates-on-some-fortune-500-sites; notes: The study found that 10.86% of 3,900 articles from official Fortune 500 company blogs were likely AI-generated in a November 2023 snapshot. It ran Originality.ai’s detector on blog articles from 246 Fortune 500 blogs and compared article length with AI flags. The estimate is a one-month, company-blog snapshot rather than a broader web benchmark.; publicationDate: 2025-11-10 |
| Published estimates | 2023 | 16.93 | percent | rawValue: 16.93%; source: https://originality.ai/blog/ai-glassdoor-reviews; notes: This study found 16.93% likely AI-written Glassdoor review 'Pros' text for S&P 500 companies in 2023. It aggregated S&P 500 company reviews and ran Originality.ai on Pros fields longer than 50 words. It measures only one section of employee reviews and depends on a vendor detector.; publicationDate: 2025-10-19 |
| Published estimates | 2023 | 12.21 | percent | rawValue: 12.21%; source: https://originality.ai/blog/ai-google-reviews-study; notes: Originality.AI estimated that 12.21% of Google reviews in 2023 were likely AI-generated. It analyzed Google reviews with its detector and reported a year-by-year series from 2019 to 2024. This is a broad review-platform estimate, but it is detector-based and limited to review text rather than general web content.; publicationDate: 2025-10-11 |
| Published estimates | 2023 | 13.6 | percent | rawValue: 13.6%; source: https://originality.ai/blog/ai-lawyer-office-reviews; notes: The study estimates 13.6% AI-generated lawyer office reviews in the United States for 2023 among 100-plus-word reviews. It scraped Google Places data for 49 city-state pairs and flagged reviews over a 0.5 Originality.ai score. It covers one review corpus only and depends on a vendor detection model.; publicationDate: 2025-11-01 |
| Published estimates | 2023 | 11.86 | percent | rawValue: 11.86%; source: https://originality.ai/blog/ai-quora-answers-study; notes: Originality.AI estimated that 11.86% of English-language Quora answers in 2023 were likely AI-generated. It scraped public Quora answers across multiple topics, kept answers of at least 50 words, and ran them through its detector. This is a direct platform-specific answer estimate, but it depends on scraper coverage and detector performance.; publicationDate: 2025-10-23 |
| Published estimates | 2023 | 3.08 | percent | rawValue: 3.08%; source: https://originality.ai/blog/ai-us-bank-reviews; notes: The study estimates 3.08% AI-generated customer reviews for United States banks in 2023. It analyzed over 19000 reviews from 47 banks and classified them with Originality.ai. It covers one banking corpus only and uses a vendor detector with very low estimated rates.; publicationDate: 2025-10-19 |
| Published estimates | 2023 | 15.64 | percent | rawValue: 15.64%; source: https://pubmed.ncbi.nlm.nih.gov/37776949/; notes: The mean AI-generation score was 15.64% for OTSR orthopedics articles published after November 2022. The study ran ZeroGPT on the body plus abstract of 425 articles published from February 2022 to September 2023. This is a direct estimate within a defined journal corpus, though detector choice and single-journal scope are clear limitations.; publicationDate: 2023-09-29 |
| Published estimates | 2023 | 0.0 | percent | rawValue: 0.00%; source: https://arxiv.org/abs/2602.00319; notes: The paper estimated that 0.00% of ICLR 2023 peer reviews were AI-generated. It trained a detector on real and synthetic 2021 ICLR reviews and applied it to 1,887 2023 reviews. The estimate is detector-based and likely misses lighter AI editing.; publicationDate: 2026-01-30 |
| Published estimates | 2023 | 12.0 | percent | rawValue: 12.0%; source: https://pubmed.ncbi.nlm.nih.gov/42213755/; notes: Estimated 12.0% of published articles in 2023 across Elsevier, Frontiers, MDPI, and PLoS exhibited evidence of LLM influence. The study scored 7.3 million full-text journal articles from 2020-2025 using 228 focal words with post-2022 frequency spikes consistent with LLM output. The estimate is a corpus-level lexical inference rather than paper-level ground-truth attribution.; publicationDate: 2026-06-02 |
| Published estimates | 2023 | 12.7 | percent | rawValue: 12.7%; source: https://arxiv.org/abs/2502.09747; notes: The study-year median across the paper's real-world text domains put LLM-assisted writing at 12.7% in 2023. The paper uses a population-level statistical framework rather than a black-box detector and reports content-share estimates across several domains such as complaints, press releases, and job postings. This gives one balanced study-year value, but it still mixes narrower content sets rather than the internet as a whole.; publicationDate: 2025-02-13 |
| Published estimates | 2023 | 3.3 | percent | rawValue: 3.3%; source: https://www.turnitin.com/press/turnitin-ai-detection-feature-reviews-more-than-65-million-papers; notes: Turnitin reported that 3.3% of submitted student papers in 2023 showed 80% or more text consistent with AI writing. The figure came from more than 65 million papers reviewed after the April 2023 launch of its AI detection feature. This is a direct large-scale prevalence estimate for heavily AI-written student work, but it is thresholded, detector-based, and narrower than broader measures of partial AI assistance.; publicationDate: 2023-07-25 |
| Published estimates | 2023 | 0.052 | percent | rawValue: 0.052%; source: https://arxiv.org/abs/2404.14244; notes: The study estimated that 0.052% of Twitter profile images in its 2023 corpus were AI-generated. The authors built a multi-stage detection pipeline and applied it to 14,989,385 profile images collected from March 7 to March 15, 2023. This is a direct estimate within a clearly defined image corpus, but it is narrowly focused on profile pictures rather than the broader web.; publicationDate: 2024-04-22 |
| Published estimates | 2024 | 37.03 | percent | rawValue: 37.03%; source: https://arxiv.org/abs/2412.18148; notes: The study-year median across Medium, Quora, and Reddit put AI-generated social-media posts at 37.03% in 2024. The paper trained and selected an AIGT detector, then applied it to a 2.4 million post corpus spanning January 2022 to October 2024. This gives one value for a major social-content dataset, but it remains platform-bound and detector-dependent.; publicationDate: 2024-12-24 |
| Published estimates | 2024 | 17.5 | percent | rawValue: 17.5%; source: https://arxiv.org/abs/2404.01268; notes: The paper reported that 17.5% of arXiv computer science abstracts by February 2024 were LLM-modified. It used a population-level statistical framework across 950,965 papers from arXiv, bioRxiv, and Nature journals published between January 2020 and February 2024. This is a strong direct estimate within a clearly defined scholarly corpus, though it reflects one venue and field rather than the whole internet.; publicationDate: 2024-04-01 |
| Published estimates | 2024 | 11.2 | percent | rawValue: 11.2%; source: https://pubmed.ncbi.nlm.nih.gov/41016415/; notes: This study estimated a mean AI publication score of 11.2% for ophthalmology manuscripts by mid-2024. It used a retrospective bibliographic analysis of ophthalmology manuscripts and computed an AI-derived prevalence score across the corpus. This is a detector-like proxy in one medical specialty and the 2024 point is mid-year rather than a full-year measure.; publicationDate: 2025-10-10 |
| Published estimates | 2024 | 13.5 | percent | rawValue: 13.5%; source: https://arxiv.org/abs/2406.07016; notes: The paper's lower-bound estimate put LLM-assisted writing at 13.5% of PubMed biomedical abstracts in 2024. It tracked excess use of LLM-associated style words across more than 15 million abstracts from 2010 to 2024 and attributed the post-ChatGPT jump to AI assistance. This is still more indirect than page-level AI-share estimates and focuses on scholarly abstracts rather than the wider web.; publicationDate: 2024-06-11 |
| Published estimates | 2024 | 5.0 | percent | rawValue: 5%; source: https://gjs.nuist.edu.cn/2025/0912/c2490a289978/pagem.htm; notes: A Chinese institutional summary reported that AI-generated content reached 5% of master's theses in 2024. The summary described a study of 230,000 theses from 2022 to 2024 using the AIGC-CHECK system and large-model semantic recognition. The original journal article was not readily accessible online, so this row relies on a reputable secondary source rather than a directly retrieved paper.; publicationDate: 2025-09-06 |
| Published estimates | 2024 | 1.57 | percent | rawValue: 1.57%; source: https://copyleaks.com/about-us/media/copyleaks-study-finds-explosive-growth-of-ai-content-on-the-web; notes: Copyleaks estimated that 1.57% of sampled Common Crawl web pages in 2024 were AI-generated. It said its study analyzed one million pages from the Common Crawl dataset with the company's own detector. This is a genuine web-wide prevalence estimate with a clear sample frame, but it is still detector-based and company-published.; publicationDate: 2024-04-30 |
| Published estimates | 2024 | 45.92 | percent | rawValue: 45.92%; source: https://docs.google.com/spreadsheets/d/1WamFyVahPDtAPFtvly30BG2QjyA-L1KYkO2UEYGKKcg/edit?gid=0; notes: Graphite estimated that 45.92% of newly published English web articles in 2024 were AI-generated. It filtered Common Crawl pages by article markup, English language, and publish date, then classified them with Surfer's AI detector using a greater-than-50% threshold. This is one of the closest direct measures in the table, but it only covers newly published English articles and depends on detector performance.; publicationDate: 2024-05-07 |
| Published estimates | 2024 | 15.8 | percent | rawValue: 15.8%; source: https://arxiv.org/abs/2405.02150; notes: The paper estimated that at least 15.8% of ICLR 2024 peer reviews were AI-assisted. It used GPTZero as a lower-bound detector in a quasi-experimental study of the ICLR review set. This is a direct estimate within a defined scholarly content set, but it is framed as a lower bound and depends on detector performance.; publicationDate: 2024-05-03 |
| Published estimates | 2024 | 21.0 | percent | rawValue: 21.0%; source: LLM-Assisted writing in engineering is associated with higher citations as evidenced from 1.17 million papers; notes: The study estimates that 21.0% of engineering abstracts published in 2024 exceeded a pre-LLM 95th-percentile marker threshold across five engineering fields plus computer science benchmark data. It sampled 1.17 million abstracts from 2019 to 2025 and inferred LLM assistance from shifts in a 228-marker lexicon rather than article-level detector labels. This is a statistical proxy for likely LLM assistance in abstracts, not a direct classification of whole papers.; publicationDate: 2026-07-24 |
| Published estimates | 2024 | 43.16 | percent | rawValue: 43.16%; source: https://originality.ai/blog/ai-climate-change-paper-abstracts; notes: This study found 43.16% likely AI-generated climate change paper abstracts in 2024. It pulled up to 500 OpenAlex abstracts per year matching 'climate change' and scored 50+ word abstracts with the Originality.ai API. It is keyword-based, limited to sampled abstracts, and relies on a vendor detector rather than disclosure data.; publicationDate: 2025-10-11 |
| Published estimates | 2024 | 2.65 | percent | rawValue: 2.65%; source: https://originality.ai/blog/ai-ebay-lazada-reviews; notes: The study found that 2.65% of eBay reviews in 2024 were likely AI-generated. It analyzed yearly review data from 2018 to 2024 and applied detector-based classification to estimate prevalence. The estimate is platform-specific and the source notes it is near some detector false-positive ranges.; publicationDate: 2025-10-17 |
| Published estimates | 2024 | 41.18 | percent | rawValue: 41.18%; source: https://www.stanventures.com/news/facebook-flooded-with-ai-over-40-of-posts-are-machine-generated-1847/; notes: A secondary writeup reported that 41.18% of long-form Facebook posts as of November 2024 were likely AI-generated. It summarized an Originality.AI study of 8,885 Facebook posts published from November 2018 to November 2024 and classified as long-form by detector-based rules. This is a useful platform-specific text estimate, but it comes through a secondary article because the original study URL now redirects.; publicationDate: 2025-01-30 |
| Published estimates | 2024 | 19.83 | percent | rawValue: 19.83%; source: https://originality.ai/blog/ai-flipkart-reviews; notes: The study found that 19.83% of Flipkart reviews in 2024 were likely AI-generated. It scanned publicly available Flipkart reviews of at least 50 words through the Originality.ai API and aggregated yearly AI classifications. The estimate is platform-specific and depends on a proprietary detector plus a length filter.; publicationDate: 2025-10-18 |
| Published estimates | 2024 | 14.9 | percent | rawValue: 14.9%; source: https://originality.ai/blog/ai-forbes-business-council-study; notes: The study found that 14.9% of Forbes Business Council posts published in 2024 were likely AI-generated. It scanned 3,952 2024 posts with Originality.ai’s detector and counted 587 as likely AI. The estimate is narrow to one contributor network and reflects detector-based classification rather than confirmed disclosure.; publicationDate: 2025-12-02 |
| Published estimates | 2024 | 29.45 | percent | rawValue: 29.45%; source: https://originality.ai/blog/ai-glassdoor-reviews; notes: This study found 29.45% likely AI-written Glassdoor review 'Pros' text for S&P 500 companies in 2024. It aggregated S&P 500 company reviews and ran Originality.ai on Pros fields longer than 50 words. It measures only one section of employee reviews and depends on a vendor detector.; publicationDate: 2025-10-19 |
| Published estimates | 2024 | 19.0 | percent | rawValue: 19.0%; source: https://originality.ai/blog/ai-google-reviews-study; notes: The study found that 19.0% of Google reviews were likely AI-generated by the end of 2024. It applied Originality.ai detection to a review dataset spanning 2019 to 2024 and aggregated yearly rates. This is a detector-based platform estimate and the existing chart already appears to contain only the 2023 point from this source family.; publicationDate: 2025-10-15 |
| Published estimates | 2024 | 26.4 | percent | rawValue: 26.4%; source: https://originality.ai/blog/ai-lawyer-office-reviews; notes: The study estimates 26.4% AI-generated lawyer office reviews in the United States for 2024 among 100-plus-word reviews. It scraped Google Places data for 49 city-state pairs and flagged reviews over a 0.5 Originality.ai score. It covers one review corpus only and depends on a vendor detection model.; publicationDate: 2025-11-01 |
| Published estimates | 2024 | 54.0 | percent | rawValue: 54%; source: https://originality.ai/blog/ai-content-published-linkedin; notes: Originality.AI estimated that 54% of long-form LinkedIn posts in October 2024 were likely AI-generated. It analyzed 8,795 LinkedIn posts of 100 words or more, spanning January 2018 to October 2024, with its own detector. This is a platform-specific detector-based estimate for long posts, not all LinkedIn content.; publicationDate: 2025-10-28 |
| Published estimates | 2024 | 5.98 | percent | rawValue: 5.98%; source: https://originality.ai/blog/ai-myntra-reviews-study; notes: The study found that 5.98% of Myntra reviews in 2024 were likely AI-generated. It analyzed Myntra review text with at least 50 words using the Originality.ai API and aggregated the results by year. The estimate is narrow to one shopping app and sits close to some detector false-positive ranges noted by the source.; publicationDate: 2025-10-24 |
| Published estimates | 2024 | 10.94 | percent | rawValue: 10.94%; source: https://originality.ai/blog/ai-quora-answers-study; notes: Originality.AI estimated that 10.94% of English-language Quora answers in 2024 were likely AI-generated. It scraped public Quora answers across multiple topics, kept answers of at least 50 words, and ran them through its detector. This is a direct platform-specific answer estimate, but it depends on scraper coverage and detector performance.; publicationDate: 2025-10-23 |
| Published estimates | 2024 | 9.93 | percent | rawValue: 9.93%; source: https://originality.ai/blog/ai-shein-reviews; notes: The study found that 9.93% of Shein reviews in 2024 were likely AI-generated. It collected Shein review data in CSV form and ran a detector-based workflow to estimate prevalence. The estimate is specific to one shopping app and relies on proprietary AI scoring.; publicationDate: 2025-10-16 |
| Published estimates | 2024 | 10.7 | percent | rawValue: 10.7%; source: https://originality.ai/blog/ai-tripadvisor-reviews-study; notes: The study found that 10.7% of TripAdvisor hotel reviews in 2024 were likely AI-generated. It used Originality.ai to score TripAdvisor reviews from 2019 to 2024 and reported yearly prevalence. The estimate is limited to one travel-review platform and inherits detector uncertainty.; publicationDate: 2025-12-16 |
| Published estimates | 2024 | 5.1 | percent | rawValue: 5.1%; source: https://originality.ai/blog/ai-us-bank-reviews; notes: The study estimates 5.1% AI-generated customer reviews for United States banks in 2024. It analyzed over 19000 reviews from 47 banks and classified them with Originality.ai. It covers one banking corpus only and uses a vendor detector with very low estimated rates.; publicationDate: 2025-10-19 |
| Published estimates | 2024 | 6.96 | percent | rawValue: 6.96%; source: https://www.pangram.com/blog/one-day-of-ai-news; notes: Pangram found 6.96% of 857,434 news articles published worldwide on July 1, 2024 were AI-generated. It pulled one day of articles from NewsCatcher and ran Pangram Text on cleaned article text. This is a detector-based one-day snapshot of online news rather than all web content.; publicationDate: 2024-08-05 |
| Published estimates | 2024 | 16.2 | percent | rawValue: 16.2%; source: https://arxiv.org/abs/2402.18463; notes: The study estimated that 16.2% of Pixiv artworks in its pooled 2022 to 2024 corpus were AI-generated. It examined 15,203,948 artworks created between January 1, 2022 and January 5, 2024, of which 2,475,485 were tagged as AI-generated. This is a direct estimate within a very large online art platform, but the value is a pooled multi-year platform estimate rather than a single-calendar-year snapshot.; publicationDate: 2024-02-28 |
| Published estimates | 2024 | 5.0 | percent | rawValue: 5%; source: https://arxiv.org/abs/2410.08044; notes: The study estimated that up to 5% of new English Wikipedia pages in August 2024 contained significant AI-generated content. It used GPTZero and Binoculars, calibrated to a 1% false positive rate on pre-GPT-3.5 articles, on a set of 2,909 new pages. This is a direct estimate within one important public knowledge platform, but it is platform-specific and focused on newly created pages.; publicationDate: 2024-10-10 |
| Published estimates | 2024 | 0.09 | percent | rawValue: 0.09%; source: https://pubmed.ncbi.nlm.nih.gov/41710616/; notes: The study estimated that at least 0.09% of PubMed medical manuscripts in its study window were ChatGPT-generated. It counted papers whose title or abstract contained the marker term 'delve' across 2,193,871 PubMed manuscripts published from December 1, 2022 to May 14, 2024. The estimate is a conservative lower bound and depends on a single lexical marker rather than a full-content attribution model.; publicationDate: 2026-01-02 |
| Published estimates | 2024 | 7.32 | percent | rawValue: 7.32%; source: https://arxiv.org/abs/2602.00319; notes: The paper estimated that 7.32% of ICLR 2024 peer reviews were AI-generated. It trained a detector on real and synthetic 2021 ICLR reviews and applied it to 1,818 2024 reviews. The estimate is detector-based and narrower than sentence-level AI-assistance measures.; publicationDate: 2026-01-30 |
| Published estimates | 2024 | 17.7 | percent | rawValue: 17.7%; source: https://arxiv.org/abs/2502.09747; notes: The study-year median across the paper's real-world text domains put LLM-assisted writing at 17.7% in 2024. The paper uses a population-level statistical framework rather than a black-box detector and reports content-share estimates across several domains such as complaints, press releases, and job postings. This gives one balanced study-year value, but it still mixes narrower content sets rather than the internet as a whole.; publicationDate: 2025-02-13 |
| Published estimates | 2024 | 11.0 | percent | rawValue: 11%; source: https://www.turnitin.com/blog/turnitin-celebrates-one-year-for-ai-writing-detection; notes: Turnitin reported that 11% of submitted student papers reviewed by April 2024 contained at least 20% AI-written text. The figure came from more than 22 million submissions within roughly 200 million papers reviewed since launch of its AI writing detector. This is a direct large-scale prevalence estimate within student writing submitted through Turnitin, but it uses a proprietary detector and a thresholded definition rather than page-level human review.; publicationDate: 2024-04-09 |
| Published estimates | 2024 | 12.0 | percent | rawValue: 12%; source: https://arxiv.org/abs/2502.11248; notes: The study estimated that 12% of shared images in a large 2024 US presidential election dataset from X were AI-generated. It analyzed a clearly defined election-related image corpus from one major social platform. This is a direct estimate within that image set, but it is visual, political, and platform-specific rather than internet-wide.; publicationDate: 2025-02-16 |
| Published estimates | 2025 | 16.0 | percent | rawValue: 16%; source: https://arxiv.org/abs/2602.00319; notes: The midpoint across ICLR and Nature Communications put AI-generated peer reviews at 16% in 2025. The longitudinal peer-review study reported roughly 20% for ICLR and 12% for Nature Communications, and this row uses the midpoint to keep one study-year value. This is a direct estimate within peer-review text, but it still reflects a narrow scholarly corpus rather than the wider internet.; publicationDate: 2026-01-30 |
| Published estimates | 2025 | 74.2 | percent | rawValue: 74.2%; source: https://ahrefs.com/blog/what-percentage-of-new-content-is-ai-generated/; notes: Ahrefs estimated that 74.2% of 900,000 newly detected English-language web pages in April 2025 contained some AI content. It used the company's own detector on a large sample of newly detected pages. This is relevant to the core question, but it measures pages containing some AI text rather than pages that are mostly AI-generated.; publicationDate: 2025-05-19 |
| Published estimates | 2025 | 17.3 | percent | rawValue: 17.3%; source: https://aiforensics.org/uploads/GenAI%20Report.pdf; notes: AI Forensics estimated that 17.3% of top search results across TikTok and Instagram in June 2025 contained synthetic AI imagery. It manually annotated 2,100 search results collected in Spain, Germany, and Poland and then built a weighted aggregate from the report's country-level counts. This is a useful direct social-platform estimate, but it is query-based, image-focused, and tied to search results rather than all feed content.; publicationDate: 2025-07-31 |
| Published estimates | 2025 | 34.0 | percent | rawValue: 34%; source: https://newsroom-deezer.com/2025/11/deezer-ipsos-survey-97-of-people-cant-tell-the-difference-between-fully-ai-generated-and-human-made-music-clear-desire-for-transparency-and-fairness-for-artists/; notes: Deezer reported that 34% of all daily music deliveries to its platform in late 2025 were fully AI-generated tracks. It said its in-house AI-music detection system was identifying roughly 50,000 fully AI-generated uploads per day by the November 2025 update. This is a direct platform-level prevalence estimate, but it is narrowly about one music-streaming service and fully synthetic tracks rather than AI-assisted content more broadly.; publicationDate: 2025-11-12 |
| Published estimates | 2025 | 51.38 | percent | rawValue: 51.38%; source: https://docs.google.com/spreadsheets/d/1WamFyVahPDtAPFtvly30BG2QjyA-L1KYkO2UEYGKKcg/edit?gid=0; notes: Graphite estimated that 51.38% of newly published English web articles through May 2025 were AI-generated. It filtered Common Crawl pages by article markup, English language, and publish date, then classified them with Surfer's AI detector using a greater-than-50% threshold. This is one of the closest direct measures in the table, but it only covers newly published English articles and depends on detector performance.; publicationDate: 2024-05-07 |
| Published estimates | 2025 | 35.0 | percent | rawValue: 35%; source: https://ai-on-the-internet.github.io/; notes: The study estimated that 35% of newly published websites by mid-2025 were AI-generated or AI-assisted. It used a representative Wayback Machine sample of websites published between 2022 and 2025 and selected Pangram v3 after robustness checks across detectors, text length, HTML versus plain text, model family, version, and language. This is one of the stronger web-wide estimates in the table, though it still relies on detector choice and focuses on newly published websites rather than the entire live web.; publicationDate: 2026-04-15 |
| Published estimates | 2025 | 11.3 | percent | rawValue: 11.3%; source: https://arxiv.org/abs/2603.19316; notes: The March 2025 overall level put significant AI-generated text at 11.3% of JAMA Network Open articles. The study uses Originality.AI on a monthly series covering original investigations, research letters, and invited commentaries from January 2022 through March 2025. This is a narrow but still direct content-share estimate for one medical publishing venue.; publicationDate: 2026-03-15 |
| Published estimates | 2025 | 5.3 | percent | rawValue: 5.3%; source: https://mededu.jmir.org/2025/1/e80084/; notes: The study found that 5.3% of screened preclinical biochemistry and cell biology videos on YouTube and TikTok were AI-generated slop. Researchers searched both platforms, screened on-topic videos, and manually judged obvious low-quality AI-generated videos in the final corpus. The estimate is explicitly a lower bound because the method was designed to catch only obvious slop.; publicationDate: 2025-11-20 |
| Published estimates | 2025 | 21.0 | percent | rawValue: 21.0%; source: https://www.kapwing.com/blog/ai-slop-report-the-global-rise-of-low-quality-ai-videos/; notes: Kapwing found that 21.0% of the first 500 Shorts shown to a brand-new YouTube account were AI-generated slop. The report manually classified AI slop and brainrot while cycling through the first 500 videos on a fresh YouTube Shorts feed. The estimate is a platform-specific new-user feed sample rather than all YouTube content.; publicationDate: 2025-11-28 |
| Published estimates | 2025 | 26.6 | percent | rawValue: 26.6%; source: LLM-Assisted writing in engineering is associated with higher citations as evidenced from 1.17 million papers; notes: The study estimates that 26.6% of engineering abstracts in the first half of 2025 exceeded a pre-LLM 95th-percentile marker threshold across five engineering fields plus computer science benchmark data. It sampled 1.17 million abstracts from 2019 to 2025 and inferred LLM assistance from shifts in a 228-marker lexicon rather than article-level detector labels. This is a statistical proxy for likely LLM assistance in abstracts, not a direct classification of whole papers.; publicationDate: 2026-07-24 |
| Published estimates | 2025 | 73.0 | percent | rawValue: 73%; source: https://originality.ai/blog/ai-generated-abstracts-in-ai-journals; notes: This study found 73% likely AI-generated abstracts in AI-related journals in 2025. It pulled up to 500 OpenAlex abstracts per year matching 'artificial intelligence' and classified 50+ word abstracts with the Originality.ai API. It is a keyword-based sample of abstracts from AI-related publications and relies on a vendor detector.; publicationDate: 2025-10-10 |
| Published estimates | 2025 | 66.8 | percent | rawValue: 66.8%; source: https://originality.ai/blog/likely-ai-pet-loss-books-study; notes: Originality.AI found that likely AI-written pet-loss books overtook likely human-written pet-loss books on Amazon in 2025, with 147 likely AI titles versus 73 likely human titles. This row converts those 2025 counts into a 66.8% likely-AI share for the year while keeping the article's broader context that 32% of analyzable 2010-2025 titles contained likely AI content. The estimate is limited to one Amazon book subcategory and depends on available samples and detector performance.; publicationDate: 2026-03-05 |
| Published estimates | 2025 | 77.0 | percent | rawValue: 77%; source: https://originality.ai/blog/likely-ai-success-self-help-book-study; notes: Originality.AI estimated that 77% of Amazon Success self-help books published from August 31 to November 28, 2025 were likely AI-written. It scanned 844 English paperback titles in the Success subcategory and checked descriptions, author bios, and sample pages with Originality.AI Lite 1.0.2. This is a high-signal platform estimate for one self-publishing niche, but it is not representative of books or the web overall.; publicationDate: 2026-01-28 |
| Published estimates | 2025 | 28.6 | percent | rawValue: 28.6%; source: https://originality.ai/blog/ai-car-dealership-reviews; notes: Originality.AI estimated that 28.6% of Canadian car dealership reviews in 2025 were likely AI-written. The study used a Google Maps-based business data API, sampled 100 dealerships across 10 Canadian cities, retained 2,000+ reviews of at least 100 words, and classified them with the Originality.AI API. This is a review-platform estimate for one retail category, not a general web-content measure.; publicationDate: 2025-12-11 |
| Published estimates | 2025 | 45.83 | percent | rawValue: 45.83%; source: https://originality.ai/blog/ai-climate-change-paper-abstracts; notes: This study found 45.83% likely AI-generated climate change paper abstracts in 2025. It pulled up to 500 OpenAlex abstracts per year matching 'climate change' and scored 50+ word abstracts with the Originality.ai API. It is keyword-based, limited to sampled abstracts, and relies on a vendor detector rather than disclosure data.; publicationDate: 2025-10-11 |
| Published estimates | 2025 | 30.41 | percent | rawValue: 30.41%; source: https://originality.ai/blog/higher-education-act-changes-public-feedback-study; notes: Originality.AI estimated that 30.41% of public comments on 2025 Higher Education Act student-aid changes were likely AI-generated. The study used a compiled Regulations.gov comment dataset, excluded submissions under 100 words, and scored 707 comments with the Originality.AI detector. This is a narrow public-comment estimate, not a general web measure, but it captures AI use in an official online civic feedback channel.; publicationDate: 2026-03-16 |
| Published estimates | 2025 | 4.1 | percent | rawValue: 4.1%; source: https://originality.ai/blog/ai-holiday-recipes-study; notes: Originality.AI estimated that 4.1% of Christmas and holiday recipe pages in its 2025 holiday-season sample were likely AI-generated. The study collected 636 recipe articles from holiday and Christmas recipe searches, then classified pages with an Originality.AI detector threshold of 0.5 or higher. This is a narrow recipe-search estimate and is useful partly because it shows a low-AI counterexample among online content niches.; publicationDate: 2026-01-09 |
| Published estimates | 2025 | 4.6 | percent | rawValue: 4.6%; source: https://originality.ai/blog/ai-home-depot-lowes-reviews; notes: Originality.AI estimated that 4.6% of 2025 Trustpilot reviews for Home Depot and Lowe's were likely AI-generated. The study examined public customer feedback for the two home-improvement retailers from 2018 to 2025 and classified reviews with its detector. This is a useful low-AI counterexample inside retail reviews, but it covers two brands on one review platform.; publicationDate: 2025-11-12 |
| Published estimates | 2025 | 21.3 | percent | rawValue: 21.3%; source: https://originality.ai/blog/ai-house-cleaning-reviews; notes: Originality.AI estimated that 21.3% of US house-cleaning service reviews in 2025 were likely AI-generated. It sampled 100 businesses across 10 major US cities, retained 700+ reviews of at least 100 words after cleaning, and scored them with the Originality.AI API. This is a local-services review estimate, useful as a narrow consumer-feedback signal rather than as a broad web average.; publicationDate: 2025-12-11 |
| Published estimates | 2025 | 16.4 | percent | rawValue: 16.4%; source: https://originality.ai/blog/ai-lawn-care-landscaping-reviews; notes: Originality.AI estimated that 16.4% of US lawn-care and landscaping reviews in 2025 were likely AI-generated. It sampled 100 businesses across 10 major US cities, kept 700+ reviews of at least 100 words, and classified reviews with the Originality.AI API. This is a narrow local-services review slice, and the article notes that earlier years in the dataset had fewer available reviews.; publicationDate: 2025-12-11 |
| Published estimates | 2025 | 34.4 | percent | rawValue: 34.4%; source: https://originality.ai/blog/ai-lawyer-office-reviews; notes: The study estimates 34.4% AI-generated lawyer office reviews in the United States for 2025 among 100-plus-word reviews. It scraped Google Places data for 49 city-state pairs and flagged reviews over a 0.5 Originality.ai score. It covers one review corpus only and depends on a vendor detection model.; publicationDate: 2025-11-01 |
| Published estimates | 2025 | 53.7 | percent | rawValue: 53.7%; source: https://originality.ai/blog/linkedin-ai-study-engagement; notes: Originality.AI estimated that 53.7% of long-form LinkedIn posts in its 2025 sample were likely AI-generated. It analyzed 3,368 posts from 99 influential profiles published from January to November 2025 and classified them with its detector. This is a detector-based estimate for a curated set of influential voices, not for all LinkedIn posts.; publicationDate: 2026-01-22 |
| Published estimates | 2025 | 20.0 | percent | rawValue: 20%; source: https://originality.ai/blog/ai-luxury-hotel-reviews; notes: Originality.AI reported that more than 20% of 2025 luxury hotel reviews in its sample were likely AI-generated. It analyzed reviews for 4- and 5-star luxury hotels across 10 major US cities using a Google Maps-powered local business API and the Originality.AI detector. This row uses 20% as a conservative floor because the article states over 20% rather than publishing a more exact 2025 value in text.; publicationDate: 2026-01-26 |
| Published estimates | 2025 | 46.4 | percent | rawValue: 46.4%; source: https://originality.ai/blog/ai-neuroscience-abstracts; notes: Originality.AI estimated that 46.4% of neuroscience journal abstracts in 2025 were likely AI-generated. It retrieved abstracts from OpenAlex by year, collected up to 500 non-empty abstracts per year, reconstructed the abstract text, and scanned abstracts of at least 50 words with the Originality.AI detector. This is a detector-based academic-abstract estimate and should be read as AI-writing signal in one research field, not as proof that the papers themselves were AI-written.; publicationDate: 2025-11-12 |
| Published estimates | 2025 | 14.7 | percent | rawValue: 14.7%; source: https://originality.ai/blog/ai-reddit-posts-study; notes: Originality.AI estimated that 14.7% of Reddit posts in its 2025 sample were likely AI-generated. It collected posts using 50 unrelated search nouns, kept posts with at least 50 words, and ended with 497 posts after filtering. This is a direct forum-level estimate, but the sample is modest and detector-based.; publicationDate: 2025-12-11 |
| Published estimates | 2025 | 10.9 | percent | rawValue: 10.90%; source: https://originality.ai/blog/ai-temu-reviews; notes: Originality.AI estimated that 10.90% of Trustpilot reviews about Temu in 2025 were likely AI-generated. The study tracked Temu review text on Trustpilot from 2022 to 2025 and reported a year-by-year increase from 0.75% in 2022 to 10.90% in 2025. This is a direct review-platform estimate, but it covers reviews about one company on one third-party review site.; publicationDate: 2026-01-09 |
| Published estimates | 2025 | 9.0 | percent | rawValue: 9%; source: https://arxiv.org/abs/2510.18774; notes: Pangram estimated that 9% of newly published articles from roughly 1,500 American newspapers in summer 2025 were partially or fully AI-generated. It audited 186,000 articles and this row keeps the overall article estimate rather than the narrower opinion-piece split. This is a useful newspaper-sector estimate, but it still covers one media slice rather than the wider web.; publicationDate: 2025-10-21 |
| Published estimates | 2025 | 29.0 | percent | rawValue: 29%; source: https://doi.org/10.1097/PRS.0000000000013114; notes: A 2026 Plastic and Reconstructive Surgery study reported that 29% of 2025 articles in 10 major plastic-surgery journals exceeded its threshold for substantial AI involvement. The authors reviewed 1,627 manuscripts published from January 2024 through May 2025 and used ZeroGPT, with the substantial-involvement threshold calibrated on 300 pre-ChatGPT manuscripts. This is a specialized academic-literature estimate, and detector thresholds may capture AI-assisted editing as well as fuller AI drafting.; publicationDate: 2026-04-09 |
| Published estimates | 2025 | 19.2 | percent | rawValue: 19.2%; source: https://arxiv.org/abs/2601.18341; notes: Robbes and coauthors estimated that 15.85% to 22.60% of sampled GitHub projects in the first half of 2025 showed coding-agent adoption, which converts to a 19.2% midpoint row value. They analyzed 129,134 GitHub projects and identified adoption through explicit repository traces such as agent-authored commits, pull requests, and agent configuration files. The figure is a conservative project-level range for coding agents rather than a direct measure of all code lines or all GitHub content.; publicationDate: 2026-01-26 |
| Published estimates | 2025 | 30.0 | percent | rawValue: 30%; source: arXiv preprint 'Scale over Preference: The Impact of AI-Generated Content on Online Content Ecology'; notes: The study reports that AI-generated videos made up about 30% of total uploads in the Local Life channel of a leading Chinese video-sharing platform by mid-2025. It analyzes platform data from June 2024 to May 2025 and describes AIGC prevalence using platform-side records of creators' use of the platform's proprietary AIGC tool. The estimate covers one named channel on one unnamed platform, so it is a strong but narrow platform-specific video row rather than a general web estimate.; publicationDate: 2026-04-02 |
| Published estimates | 2025 | 57.0 | percent | rawValue: 57.0%; source: https://pubmed.ncbi.nlm.nih.gov/42213755/; notes: Estimated 57.0% of published articles in 2025 across Elsevier, Frontiers, MDPI, and PLoS exhibited evidence of LLM influence. The study scored 7.3 million full-text journal articles from 2020-2025 using 228 focal words with post-2022 frequency spikes consistent with LLM output. The estimate is a corpus-level lexical inference rather than paper-level ground-truth attribution.; publicationDate: 2026-06-02 |
| Published estimates | 2025 | 7.0 | percent | rawValue: 7%; source: https://science.feedback.org/second-measurement-mis-disinformation-major-platforms-europe/; notes: Science Feedback estimated that 7% of Facebook mis/disinformation posts in its October 2025 SIMODS sample contained AI-generated images or video. It used exposure-weighted samples of public-interest mis/disinformation posts collected across four EU countries in the second SIMODS measurement. This is a narrow but useful prevalence estimate for misleading content, not for Facebook posts overall.; publicationDate: 2026-03-19 |
| Published estimates | 2025 | 2.6 | percent | rawValue: 2.6%; source: https://science.feedback.org/second-measurement-mis-disinformation-major-platforms-europe/; notes: Science Feedback estimated that 2.6% of Instagram mis/disinformation posts in its October 2025 SIMODS sample contained AI-generated images or video. It used exposure-weighted samples of public-interest mis/disinformation posts collected across four EU countries in the second SIMODS measurement. This is a narrow platform-specific estimate for misleading content rather than Instagram content overall.; publicationDate: 2026-03-19 |
| Published estimates | 2025 | 24.0 | percent | rawValue: 24%; source: Science Feedback / SIMODS second measurement; notes: SIMODS found that 24% of TikTok mis/disinformation posts in its October 2025 European platform sample contained AI-generated elements. It used exposure-weighted samples of public-interest content annotated by professional fact-checkers and added a new wave-two indicator for AI-generated images or video. The figure is event- and topic-bounded because the denominator is only mis/disinformation posts in the sampled public-interest corpus, not all TikTok content.; publicationDate: 2026-03-19 |
| Published estimates | 2025 | 4.4 | percent | rawValue: 4.4%; source: https://science.feedback.org/second-measurement-mis-disinformation-major-platforms-europe/; notes: Science Feedback estimated that 4.4% of X/Twitter mis/disinformation posts in its October 2025 SIMODS sample contained AI-generated images or video. It used exposure-weighted samples of public-interest mis/disinformation posts collected across four EU countries in the second SIMODS measurement. This is a narrow X-specific estimate for misleading content, not for the full platform.; publicationDate: 2026-03-19 |
| Published estimates | 2025 | 35.0 | percent | rawValue: 35%; source: https://arxiv.org/abs/2504.08755; notes: The paper's midpoint estimate put AI-generated text at 35% of active web pages in 2025. It inferred AI presence from the frequency of keywords associated with ChatGPT-style writing rather than direct page-by-page labeling. This is unusually close to the core question, but it relies on a heuristic marker method rather than direct classification.; publicationDate: 2025-03-29 |
| Published estimates | 2025 | 20.0 | percent | rawValue: 20.0%; source: https://www.totallyhuman.io/blog/the-surprising-new-number-of-genai-games-on-steam; notes: Totally Human Media reported that a little under 20% of games released on Steam in 2025 disclosed generative AI use. It used Steam API data and Steam's disclosure field to count titles that self-reported GenAI usage. This is a lower bound on disclosed AI use in games on one storefront, not a direct detection-based measure of all game content.; publicationDate: 2025-07-13 |
| Published estimates | 2026 | 23.0 | percent | rawValue: 23%; source: https://antislop.online/research/state-of-slop-2026; notes: AntiSlop estimated that 23% of its category-balanced sample of web pages in March 2026 were AI-heavy. It sampled 2,000 pages, with 200 pages in each of 10 content categories, and classified them with AntiSlop's own detector using a greater-than-60% AI threshold. This is a rare early-2026 web estimate, but it is category-balanced rather than representative of the whole open web and depends on detector-based labeling.; publicationDate: 2026-03-21 |
| Published estimates | 2026 | 33.3 | percent | rawValue: 33.3%; source: https://www.techradar.com/audio/apple-music/every-label-in-the-world-is-delivering-ai-apple-music-executive-says-over-a-third-of-uploads-are-100-percent-ai-as-it-clamps-down-on-ai-fraud; notes: Apple Music's VP Oliver Schusser said more than a third of tracks delivered to the platform today are 100% AI. This row uses 33.3% as a conservative floor for that 2026 upload-intake claim, based on trade-press reporting of a Billboard interview rather than a public audited dataset. It is a platform-specific music-upload signal, and it does not mean one-third of listening or the whole Apple Music catalog is AI.; publicationDate: 2026-04-27 |
| Published estimates | 2026 | 44.0 | percent | rawValue: 44%; source: https://newsroom-deezer.com/2026/04/ai-generated-tracks-represent-44-of-new-uploaded-music/; notes: Deezer reported that 44% of all daily music uploads to its platform in April 2026 were fully AI-generated tracks. It said its in-house AI-music detection system was identifying nearly 75,000 fully AI-generated uploads per day by the April 2026 update. This is a direct platform-level prevalence estimate, but it is narrowly about one music-streaming service and fully synthetic tracks rather than the web at large.; publicationDate: 2026-04-21 |
| Published estimates | 2026 | 42.7 | percent | rawValue: 42.7%; source: https://graphite.io/five-percent/ai-search-collapse; notes: Graphite reports that 42.7% of unique URLs cited by ChatGPT for its prompt set in early June 2026 were predicted to be AI-generated. It collected reference URLs for 1,019 information-seeking prompts and classified unique cited URLs with GPTZero, also reporting 38.9% for the January 2026 collection. This is a detector-based estimate for ChatGPT-cited references rather than for the whole web or for all search results.; publicationDate: 2026-06-11 |
| Published estimates | 2026 | 49.94 | percent | rawValue: 49.94%; source: https://graphite.io/five-percent/ai-now-writes-as-many-online-articles-as-humans-do; notes: Graphite estimated that 49.94% of English-language online articles and listicles published in 2026-Q1 were primarily AI-generated in its Common Crawl sample. It randomly sampled 55.4k English URLs with article schema from 2020 to March 2026 and averaged three detector-based quarterly estimates from Pangram, Copyleaks, and GPTZero. This is detector-based and limited to English article-style web pages rather than the full internet.; publicationDate: 2026-05-15 |
| Published estimates | 2026 | 12.0 | percent | rawValue: 12.0%; source: https://graphite.io/five-percent/ai-content-and-search; notes: Graphite found that 12.0% of article-like pages in the top 20 Google organic results had more than 50% AI content, while 3.0% were purely AI-generated. It sampled 2,200 keywords, kept 20,280 filtered article or listicle URLs from the top 20 results, and applied paragraph-level Originality.ai thresholds. The estimate excludes many page types and relies on one detector and Graphite’s filtering choices.; publicationDate: 2026-04-16 |
| Published estimates | 2026 | 59.0 | percent | rawValue: 59.0%; source: https://www.kapwing.com/resources/the-tiktok-ai-slop-report/; notes: Kapwing found that 59.0% of the first 500 videos shown to a brand-new TikTok account were AI slop on the For You feed in May 2026. The study manually reviewed the first 500 feed videos and separately hand-coded 10,742 videos across 20 TikTok categories using visible AI-generated visual, script, and voiceover cues. The estimate is platform-specific and uses Kapwing’s manual slop definition rather than a validated platform-wide detector.; publicationDate: 2026-06-09 |
| Published estimates | 2026 | 8.0 | percent | rawValue: 8%; source: https://originality.ai/blog/likely-ai-us-hiking-guides-pinterest; notes: Originality.AI estimated that 8% of Pinterest-surfaced hiking guides published so far in 2026 were likely AI-generated. It analyzed 1,000 hiking-guide URLs surfaced by Pinterest searches across 40 US national park and trail keywords and then broke the results out by publication year. This is a query-based Pinterest travel subset, not a platform-wide prevalence estimate, and the 2026 figure is early-year only.; publicationDate: 2026-04-25 |
| Published estimates | 2026 | 32.4 | percent | rawValue: 32.4%; source: https://originality.ai/blog/likely-ai-medical-diet-recipes-on-pinterest; notes: Originality.AI estimated that 32.4% of Pinterest medical-diet recipe URLs in its April 2026 sample were likely AI-generated. It collected 1,109 recipe URLs across five medically sensitive diet-search lanes and scored them with its detector. This is a narrow but important platform estimate for health-related recipe content, not for Pinterest overall.; publicationDate: 2026-04-19 |
| Published estimates | 2026 | 3.0 | percent | rawValue: 3.0%; source: https://www.pangram.com/blog/ai-amazon-reviews; notes: Pangram found 3.0% of 30,000 front-page reviews across 500 Amazon best-selling products were AI-generated. It scraped first-page reviews and scored them with Pangram's detector while recording star ratings and Verified Purchase status. This is a detector-based sample of prominent reviews on Amazon best-sellers, not all Amazon reviews.; publicationDate: 2026-05-04 |
| Published estimates | 2026 | 21.0 | percent | rawValue: 21%; source: https://www.pangram.com/blog/pangram-predicts-21-of-iclr-reviews-are-ai-generated; notes: Pangram found that 21% of ICLR 2026 peer reviews, or 15,899 reviews, were fully AI-generated. It ran its detector on all roughly 70,000 public ICLR reviews and reported a 2022 negative-control error check with no fully-AI versus fully-human confusions. This is a conference-review corpus rather than the open web, and the estimate depends on a proprietary detector.; publicationDate: 2025-11-18 |
| Published estimates | 2026 | 13.8 | percent | rawValue: 13.8%; source: Pangram Labs blog post 'AI in Your Feed'; notes: Pangram found that 13.8% of scanned social-feed posts over 50 words were AI-generated across LinkedIn, Medium, Substack, X, and Reddit in spring 2026. It used opt-in Chrome-extension browsing data covering 1,002,627 posts and classified each post with Pangram 3.3. The sample reflects what extension users saw in-feed rather than a platform-representative universe of all posts.; publicationDate: 2026-07-09 |
| Published estimates | 2026 | 71.83 | percent | rawValue: 71.83%; source: https://www.themoonlight.io/es/review/agentic-very-much-adoption-of-coding-agent-in-new-github-projects; notes: A detailed secondary summary of Robbes and coauthors’ June 2026 preprint reports that 71.83% of sampled new GitHub projects showed detectable coding-agent adoption. The summary describes a conservative repository-level estimate built from GitHub API sampling plus file-, pull-request-, and commit-based heuristics on projects created after late August 2025. This row currently relies on a secondary route because the public arXiv abstract confirms the study but does not expose the exact percentage in the accessible abstract text.; publicationDate: 2026-06-05 |
| Published estimates | 2026 | 2.5 | percent | rawValue: 2.5%; source: https://arxiv.org/abs/2602.13504; notes: A 2026 arXiv preprint estimated that 2.5% of January 2026 news articles across three Turkish online media outlets were rewritten or revised by LLMs. The author fine-tuned a Turkish BERT classifier on 3,600 labeled articles, then applied it to 300 randomly selected articles per year from each outlet for 2023 through 2026. This is a primary academic estimate with a clear news corpus, but it covers three Turkish outlets and uses model predictions rather than human labels for the external-year estimates.; publicationDate: 2026-02-13 |
| Annual average | 2020 | 2.6633 | percent | estimateCount: 6 |
| Annual average | 2021 | 5.124 | percent | estimateCount: 5 |
| Annual average | 2022 | 4.9442 | percent | estimateCount: 12 |
| Annual average | 2023 | 11.5554 | percent | estimateCount: 17 |
| Annual average | 2024 | 17.3552 | percent | estimateCount: 31 |
| Annual average | 2025 | 29.1195 | percent | estimateCount: 37 |
| Annual average | 2026 | 29.7479 | percent | estimateCount: 14 |