The headline you keep seeing — "half the internet is now written by AI" — is close to right and slightly wrong at the same time, and the gap between the two matters a great deal if you publish for a living. The number comes from Graphite, an SEO firm that sampled articles from Common Crawl and ran them through AI detectors: it found that the share of newly published English-language articles that are primarily AI-generated climbed from a near-zero baseline before ChatGPT to roughly parity with human-written articles by late 2024, and has plateaued near 50% since. That is a real, defensible finding about a specific slice of the web — newly published text articles — not a claim that half of everything online is machine-made. But the finding most people skip is the one that should change how you work: those AI articles largely do not show up in Google search results or in ChatGPT's answers. The web filled up with AI text; the text mostly didn't win distribution. Volume and visibility came apart. This guide walks through what the studies actually measured and how confident you can be in the numbers, why the curve rose so fast and then flattened, the difference between the raw generation everyone is drowning in and the differentiated content that still gets read, how detection and platform enforcement tightened across 2026, and — the practical part — how to run a high-volume content operation that uses AI without becoming another entry in the pile the study measured.
The claim traveling around the industry is "half the internet is now written by AI." The research behind it is more specific and more useful than the slogan. It comes from Graphite, a growth and SEO firm that sampled articles from Common Crawl — a large, publicly available crawl of the web — and classified each one with AI detection tools. What it found is that among newly published English-language articles, the share that are primarily AI-generated has risen to roughly parity with human-written ones. Not half of all pages that have ever existed; not half of every word online; about half of new text articles in a sample of the open web. Holding that distinction is the difference between using the finding well and repeating a distortion of it.
The shape of the curve is the interesting part. Graphite's first analysis, published in October 2025, found that articles predating ChatGPT registered only a low rate of AI detection — consistent with the tools' baseline false-positive rate, meaning effectively little to no AI authorship. After ChatGPT launched on November 30, 2022, the share climbed fast: to roughly 39% within about a year, then past 50% around November 2024, at which point machine-written articles edged ahead of human-written ones for the first time. Then, notably, it stopped climbing. The line plateaued near parity rather than marching toward total saturation — a detail the doom-laden version of the headline tends to omit.
Reasonably confident about the trend; appropriately humble about the exact figure. Two things bound the certainty. First, the sample: Common Crawl is broad but not the whole web, and it skews toward what crawlers reach, so the number describes a slice, not a census. Second, and more important, AI detectors are probabilistic. They produce false positives (flagging human writing as AI) and false negatives (missing AI writing), and no single detector is authoritative. A number built on one detector inherits that detector's error.
Graphite clearly understood this, because its follow-up analysis changed the method to harden the result. Instead of one detector, it averaged three — Pangram, Copyleaks, and GPTZero — across roughly 55,000 Common Crawl articles, and extended the data into early 2026. Averaging independent detectors reduces the chance that any one tool's bias drives the conclusion. The more conservative approach moved the estimate down by about three percentage points: AI-written content peaked near 50.9% and settled around 49.9% in the first quarter of 2026. That the figure barely moved under a stricter method is itself reassuring — the trend is robust even if the second decimal isn't. The honest way to cite this is "roughly half of newly published sampled articles, with real uncertainty around the precise share," not "exactly 50% of the internet."
The rise needs little explanation: ChatGPT and the tools that followed made competent text nearly free to produce, and a large content industry that was already optimizing for volume simply switched inputs. When the marginal cost of another article falls to near zero, the number of articles rises to meet whatever demand exists — and for years the operating assumption in content marketing was that more indexed pages meant more traffic. AI removed the last constraint on volume, so volume exploded.
The plateau is the more instructive half. If AI text were reliably winning readers and rankings, the share would likely have kept climbing as more publishers piled in. It flattened instead — which fits the finding that these articles largely aren't being distributed. When a tactic stops paying off, adoption stops accelerating. The plateau near 50% is consistent with a market discovering that flooding the zone with machine text has diminishing returns, even as the tools to do it get cheaper and better. The ceiling isn't technical; it's that the strategy underneath the volume stopped working.
Here is the sentence in the research that should reorganize how you think about publishing: Graphite noted that these AI-generated articles largely do not appear in Google search results or in ChatGPT's answers. Sit with what that means. The web absorbed an enormous surge of AI-written text, and that text is mostly not the content winning search visibility or getting cited by AI assistants. Production went up; distribution didn't come with it. The two decoupled.
For a decade, "publish more" was a reasonable proxy for "reach more," because content was expensive enough that volume signaled effort and effort correlated with quality. AI severed that link. When anyone can generate a thousand articles in an afternoon, the article stops being a scarce signal of anything, and the systems that allocate attention — search rankings, AI answer engines, social feeds — adjust to discount it. The commodity got cheaper; the reward for producing the commodity got smaller. This is why out-producing the flood is a trap: you are spending more to compete harder for a prize that is shrinking precisely because everyone is competing for it.
The strategic response is not "stop using AI" — that leaves the productivity gain on the table for no reason. It is to be deliberate about which side of a line your output sits on. On one side is raw generation: undifferentiated text produced from a thin prompt, indistinguishable from the millions of other pieces produced the same way. That is the pile the study measured, and it is the pile that isn't ranking. On the other side is differentiated content: work that carries a recognizable voice, a specific point of view, an owned identity, and a format that can't be reproduced by anyone with the same tool and a similar prompt.
Two things reliably push content to the differentiated side. The first is a governed voice. Generic AI prose reads generic because it was given nothing specific to be — no stated point of view, no editorial rules, no vocabulary that belongs to you. Feed the model a real specification of your voice and it stops sounding like everyone else's model. The second is format. The Graphite studies counted written articles, which is telling: the surge is concentrated in the cheapest, most commoditized format there is. Short-form video, avatar content, brand-exact carousels, and native social posts are far harder to mass-produce from a one-line prompt, and they compete in feeds that aren't yet saturated the way the article graveyard is. Moving up the format ladder is, by itself, a differentiation strategy.
There is also a downside risk to churning generic text that didn't exist a year ago. Across 2026, the platforms tightened their handling of low-effort AI content. Substack added a reader-facing AI detector so audiences can scan a post for machine-written text; Snapchat moved to deprioritize fully AI-generated content in Spotlight; YouTube, TikTok, and search all pressed on quality and disclosure signals. The effect is that undifferentiated AI content isn't just failing to earn reach — it can actively cost you, in reader trust and in distribution, if it reads as slop. For the wider pattern across surfaces, see the cross-platform AI content quality crackdown, and for the broader saturation dynamic, AI-generated content is flooding every platform.
The compliance and quality bars are converging on the same instruction: produce content a human stands behind. That is good news, because it means the same move — a governed voice plus a human review step — solves the differentiation problem and the enforcement problem at once. You don't have to choose between using AI and staying credible; you have to choose between raw output and reviewed, on-brand output.
The practical challenge is that "differentiated, governed, multi-format, human-reviewed" sounds like the opposite of "high-volume," and the whole appeal of AI is volume. Resolving that tension is a workflow problem, and it comes down to concentrating the human decisions where they define your identity — at the front and the end of the pipeline — while automating the mechanical middle. That is the shape Kompozy is built around, and it maps directly onto what the Graphite data says still works.
The front of the pipeline is a governed voice. Kompozy's Persona Brief is a written specification of your point of view, style, and — crucially — a banned-word filter that strips the tells of generic model prose before they ever ship. That brief governs every generation, so a batch of output reads as one recognizable identity rather than as a dozen anonymous drafts. This is the single most effective lever against "sounds AI-written," because the reason generic content sounds generic is that it was never told to be anything specific. The same brief drives an AI Influencer persona whose face and voice stay consistent across formats, giving the whole body of work an owned identity a competitor can't clone from a prompt.
The middle is format breadth, which is where you climb off the commoditized text tier the study measured. From a single source, Kompozy generates captioned Persona Shorts and avatar video, brand-exact Carousel Posts rendered through HyperFrames, quote graphics, photo posts, blog articles, and email newsletters — one idea, many differentiated formats, not another lone article dropped into the pile. And the end of the pipeline is the part the enforcement trend now demands: a per-post human review gate. On Autopilot, Kompozy schedules and publishes across the eight social platforms plus blog and email, but a person approves and edits every piece before it goes out. That review step is the human sign-off that keeps the work on the differentiated, credible side of the line — and it's also where reach gets solved, because you're distributing to owned and social channels you control instead of betting on search visibility the study says AI articles rarely get.
The lesson from the data isn't that AI ruined content or that you should abandon it. It's that the era when volume alone bought reach is over — the web proved it by filling with AI text that mostly doesn't rank. What still works is what always worked, now under more pressure: a distinct voice, formats worth the reader's time, and a human accountable for the result. Use AI to make that faster and broader, not to add one more indistinguishable page to a pile that's already half-machine and half-invisible.
The most-cited figure comes from Graphite, which sampled articles from Common Crawl and ran them through AI detectors. It found the share of newly published English-language articles that are primarily AI-generated is at roughly parity with human-written ones — its follow-up study, using three averaged detectors through early 2026, put the figure near 49.9%. That is best understood as 'about half of new articles in this sample are primarily AI-written,' not 'half of everything on the internet is AI.' An article counted as AI-generated when more than half of its text read as machine-written.
Around November 2024, according to Graphite's first analysis, primarily AI-generated articles crossed 50% for the first time and edged past human-written output — then the trend plateaued rather than continuing to climb. The rise tracks ChatGPT, which launched November 30, 2022: articles published before that date showed only a low, false-positive-level rate of AI detection, and within about a year the AI share had reached roughly 39%.
Largely not, per Graphite. The firm observed that despite the volume of AI content online, these articles mostly do not appear in Google search results or ChatGPT answers. So the surge in AI text did not translate into a surge in distribution — volume and visibility came apart. For a publisher, that is the single most important finding in the research: producing more of the commodity everyone is producing does not buy reach.
It is a credible estimate of a specific thing, not a precise census of the whole internet. The samples come from Common Crawl (a broad but not exhaustive crawl), and AI detectors are probabilistic — they produce false positives and false negatives. Graphite's follow-up study averaged three detectors specifically to reduce that error, which lowered the estimate by about three percentage points versus its earlier single-detector number. Read it as 'roughly half of newly published sampled articles,' with real uncertainty around the exact figure.
Compete on differentiation, not volume. Give the AI a governed voice — a written brief with your point of view, style rules, and banned words — so output reads as you rather than as generic model prose; keep a human review step before anything publishes; and produce formats beyond plain text (short-form video, carousels, avatar content) that resist one-prompt reproduction. The study measured commodity text articles; a recognizable, multi-format, human-reviewed body of work is precisely what it isn't measuring, and what still earns attention.
According to 2026 research from the SEO firm Graphite, primarily AI-generated articles are at roughly parity with human-written ones among newly published web articles — up from a near-zero baseline before ChatGPT launched in November 2022, crossing 50% around late 2024 and then plateauing. The crucial caveat: these AI articles largely do not appear in Google search or ChatGPT answers, so volume rose but distribution didn't follow.
Get started → · ← All guides · Compare Kompozy vs other tools