AI-generated content now makes up roughly half of the new English-language articles on the web. That is the finding of a 2025 study by the SEO firm Graphite, which sampled 43,000 articles and classified each one with an AI detector. The line below plots Graphite's own monthly figures, and the sections after it explain how the number was produced and why it is an estimate, not a head count.
How much of the web is written by AI?
About half of new articles, by Graphite's measure. In January 2020 the detector flagged 2.16% of sampled articles as AI-generated. By 2025 the share averaged 51.4% across the months January to May. The first single month above 50% was November 2024, at 51.1%, and the highest was January 2025, at 55.1%.
The yearly averages from the same data show how sharp the break was:
| Year | AI-generated share of sampled articles |
|---|---|
| 2020 | about 2–3% (monthly range 1.95–3.16%) |
| 2022 | 6.8% |
| 2023 | 31.6% |
| 2024 | 45.9% |
| 2025 (January–May) | 51.4% |
Little moves until late 2022. The curve then bends upward through 2023, the year after ChatGPT launched, and slows near the 50% line. After mid-2024 the monthly values wobble between about 41% and 55% rather than climbing in a straight line.
How did Graphite measure it?
Graphite took a random sample of 43,000 English-language pages from Common Crawl, the public web archive. Each page had article markup, at least 100 words and a publish date between January 2020 and May 2025. Each article was split into 500-word chunks and scored with Surfer's AI detector. An article counted as AI-generated when more than half of its text was predicted to be AI-written. The full write-up is on Graphite's site.
The study published its own error rates. On 15,894 articles from before ChatGPT existed, which it treats as human-written, the detector wrongly flagged 4.2%. On 6,009 articles generated by GPT-4o it missed 0.6%. Both are small, but a false-positive rate of 4.2% means the 2020 baseline of about 2% may be the detector's noise floor rather than evidence of AI writing before the tools existed.
What the number does not tell you
Treat the half-and-half headline as a well-sourced estimate with four limits:
- One detector. Detector accuracy is contested, and this version of the study used a single one.
- Volume, not readership. It counts articles published, not articles read. Graphite itself suspects AI articles are not viewed in proportion to their number, but did not test it.
- Human-edited AI drafts are not measured. A draft written by a model and rewritten by a person is classed by how it reads, not how it started.
- A newer version exists. Graphite has since published a revision that averages three detectors and uses data into 2026. By Graphite's account its AI share is on average 3.3 percentage points lower than the version charted here.
If you quote the figure, say "about half" and name the source. Do not quote a precise percentage for a month as if it were a census.
Chart it yourself
This is a single measure over 65 months, which is the job of a line chart. The data came as a public Google Sheet with one row per month; saving it as a CSV and uploading it is all the data preparation the chart needed. The line chart guide covers the clicks. Two choices are worth copying:
- Plot one series, not two. Human and AI shares sum to 100%, so the second line is the first one mirrored. One line with a clear title says more.
- Put the caveat in the footer. The source line on the chart above names the detector, the sample size and says it is an earlier version of the study, so the caveat travels with the picture when someone shares it.
For a smoother yearly view, an area chart of the same data works, and a stacked column chart of human against AI share per year shows the 100% split directly. If you want the shift to play out over time, the same table can be turned into a line chart race video. To have an assistant build the chart for you, connect the PlotSet MCP server; the ten-prompt guide shows what to ask. Charts that rely on detector output deserve the same scepticism as any estimate, which is the subject of AI makes your charts now and can't tell when they lie.
FAQ
Is more than half of the internet written by AI?
Not by this evidence. Graphite's sample found about half of new English-language articles were classified as AI-generated in 2025, with monthly values between roughly 41% and 55% after mid-2024. That covers articles in Common Crawl, not every page, video or social post on the internet.
When did AI-written articles overtake human-written ones?
In Graphite's monthly series the first month above 50% was November 2024 (51.1%). Graphite also reported that AI articles made up 39% of the total about a year after ChatGPT's launch.
How accurate is AI content detection?
Imperfect. Graphite measured a 4.2% false-positive rate on pre-ChatGPT articles and a 0.6% false-negative rate on GPT-4o text, but tested only one detector and one generator. Edited or mixed human and AI text is the hardest case.
Does AI-written content rank in search?
Graphite's author attributes the flat share since mid-2024 to AI articles not performing well in search results, but that is an interpretation, not something the study tested.
Where can I get the data?
Graphite published the monthly figures in a public spreadsheet linked from its study. Download it as a CSV, then upload it to a PlotSet line chart.
