AI Content Detection Statistics 2026 (Plus a Risk Score)

Graphite, a digital marketing agency, pulled 55,400 English-language article URLs from the Common Crawl archive and ran them through three AI detectors. The result, reported by Axios in May 2026, is that roughly half of new articles published since early 2025 are mostly AI-written, and that share has stopped climbing. Half the web now carries a machine fingerprint, and the tools that read those fingerprints disagree with each other, with the vendors that build them, and sometimes with reality.

That matters to anyone who publishes for a living. Editors, professors, search engines and AI answer engines are all making calls about what a human wrote. Below are the AI content detection statistics that hold up to a source check, what they mean for PR and earned media, and a scoring model we built to rate the detection risk of a single piece before you send it.

How much of the web is AI-generated in 2026?

The short answer is about half of new articles, and most new pages contain at least some AI text. The longer answer depends on how you count mixed content.

Graphite’s time series is the cleanest trend line available. Within a year of ChatGPT’s November 2022 launch, about 35.9% of new online articles were mostly AI-written. Within two years the share hit 48%. Since early 2025 it has hovered around half. Graphite averaged results across Pangram, GPTZero and Copyleaks and only counted a page as AI if most of its text was flagged, so a press release that is 40% machine-drafted would land in the human column.

Ahrefs took a broader view. Its team analyzed 900,000 newly created English pages in April 2025 with an in-house detector and found that 74.2% contained AI-generated content. Only 2.5% were pure AI. Pure human pages made up 25.8%. The other 71.7% were mixed, ranging from light assistance to text that was 71% to 99% machine-written.

Fully synthetic pages are rare. Hybrid pages are the norm.

Visibility tells a different story than volume. Graphite’s October 2025 follow-up checked the first two Google results pages for 31,493 keywords and found 86% of ranking articles were human-written. ChatGPT and Perplexity citations ran 82% human. Ahrefs, studying 500,000 URLs cited in Google AI Overviews in July 2025, found 3.6% pure AI, 8.6% pure human and 87.8% mixed. The two firms used different detectors and different definitions, yet they agree on one point: pages that are mostly machine output rarely win the top slots.

Detector accuracy depends on who ran the test

Accuracy claims swing wildly because the test conditions swing wildly. Read every number in this section with the tester’s name attached.

The low point came from the company that started the boom. OpenAI launched its own text classifier on January 31, 2023, and published its numbers openly: it correctly flagged 26% of AI-written text as “likely AI-written” and wrongly labeled 9% of human text as AI. By July 20, 2023, OpenAI had pulled the tool, citing its low rate of accuracy.

Academic testing that year was just as harsh. Debora Weber-Wulff and seven co-authors tested 14 detection tools in 2023, including Turnitin and PlagiarismCheck, and concluded the tools were “neither accurate nor reliable.” They also found a lean toward calling text human, and that simple obfuscation cut performance sharply.

Adversarial testing remains the weak spot. The RAID benchmark, published in 2024, built more than 6 million generated texts from 11 models across 8 domains, applied 11 adversarial attacks and ran 12 detectors against them. The authors found current detectors are easily fooled by paraphrasing, changed sampling settings, repetition penalties and models they had not seen in training.

The numbers improved once commercial vendors matured. Brian Jabarian and Alex Imas at Chicago Booth tested GPTZero, Originality.ai, Pangram and the open-source RoBERTa model on about 2,000 human passages paired with output from four large language models. Their September 2025 paper, summarized by Chicago Booth Review, found all three commercial tools kept false positives below 1%, with Pangram near 0%. False negatives told another story: Originality.ai missed between 10% and 40% of AI text depending on the model. All three commercial tools lost accuracy on passages under 50 words, and RoBERTa performed close to random guessing. The authors proposed a “policy cap,” where an institution fixes a maximum false-positive rate such as 0.5% and compares tools under that one standard.

GPTZero later re-ran the same benchmark itself and reported 99.5% accuracy in January 2026. Treat that as a vendor claim, published on the vendor’s own blog, with a downloadable prediction file you can audit.

On clean, unedited model output, the best commercial detectors in 2026 are good. On short, edited or paraphrased copy, which describes most PR writing, the error bars widen fast.

False positives fall hardest on non-native writers

The most cited fairness study in AI content detection statistics came out of Stanford in 2023. A team led by James Zou ran 91 TOEFL essays, written by real students before ChatGPT existed, through seven popular detectors. As Stanford HAI reported, the detectors labeled 61.22% of those essays as AI-generated on average. Eighty-nine of the 91 essays, or 97%, were flagged by at least one detector. Eighteen essays were flagged by all seven.

Detectors reward surprise. Writers working in a second language pick common words and constructions, which reads as low surprise, which reads as machine. The same pattern hits heavily edited corporate prose and anything run through a grammar tool until every rough edge is gone.

Scale turns a small error rate into a large number of people. Turnitin said at launch that its detector had a false positive rate under 1%. Vanderbilt University did the math on its own volume: the school submitted 75,000 papers to Turnitin in 2022, so a 1% error rate would have mislabeled about 750 student papers. Vanderbilt disabled the detector in August 2023, citing that risk, the bias against non-native speakers and Turnitin’s refusal to explain how the tool works.

Turnitin’s own usage data shows how much text flows through these systems. By March 2024 it had reviewed more than 200 million papers. Over 22 million, about 11%, had at least 20% AI writing. More than 6 million, about 3%, were at least 80% AI.

For PR, the lesson carries straight over. A founder who writes in English as a second language, or a client whose bio was polished by three rounds of editing software, can trip a detector without a single sentence coming from a model. If an editor runs your pitch through a checker, you will never see the score. You will only see the silence.

What does Google actually do with AI content?

Google does not demote content for being AI-generated. It demotes content for being unhelpful and mass-produced, and it has said so in writing more than once.

In February 2023, Google’s Search Central team published guidance stating that it rewards “high-quality content, however it is produced,” as 9to5Google reported at the time. The same guidance said using automation with the primary purpose of manipulating rankings breaks its policies.

Google sharpened that line in March 2024. Its core update that month added a spam policy called scaled content abuse, aimed at pages produced in bulk to game rankings “no matter whether content is produced through automation, human efforts, or some combination.” Google first projected the update would cut low-quality, unoriginal results by 40%, then reported a 45% reduction after rollout finished on April 19, 2024.

The detector that matters to Google is a usefulness detector, not an authorship detector. The Graphite finding that 86% of top-ranking articles are human-written fits that reading.

That has a direct bearing on earned media. Of the 1,016 publications in the Instant Press catalog, 96% are Google-indexed, so a placement’s value rides on Google treating the article as helpful. The traits Google rewards (specific names, dates, quotes and numbers) are the same traits a detector reads as human. Quality and low detection risk point the same way.

Watermarks and the EU AI Act move detection upstream

Statistical detectors guess after the fact. Watermarks try to settle the question at the moment of creation.

Google announced its SynthID Detector portal at I/O on May 20, 2025, and said SynthID had already been used to watermark more than 10 billion pieces of content across images, audio, video and text from tools such as Gemini, Imagen, Lyria and Veo. Google has also open-sourced its text watermarking so other model builders can adopt it. The limit is obvious: SynthID only confirms output from tools that applied SynthID. Text from a model without a watermark, or text rewritten heavily by a person, carries no signal to find.

Regulation is pushing more providers toward marking. Article 50 of the EU AI Act took effect on August 2, 2026, according to Cooley’s analysis. Providers of generative AI systems must embed machine-readable markings and offer a way to detect them, with existing systems given until December 2, 2026. Deployers that publish AI-generated text to inform the public on matters of public interest must disclose it, unless the text has gone through substantive human editorial review with a person taking editorial responsibility. Fines reach €15 million or 3% of worldwide annual turnover, whichever is higher.

That editorial review exception is the single most useful clause in the law for communications teams. A named human who checks facts, rewrites and signs off is both the legal safe harbor in Europe and the practical defense against a false flag everywhere else.

Score your PR content with the TRACE Risk Score

No public study measures detection risk for press releases or contributed articles specifically, so we built a model from the factors the research above keeps pointing to. We call it the TRACE Risk Score. It rates a single piece of content on five factors, each scored 0, 1 or 2, for a total between 0 and 10. A higher number means a higher chance that a detector, an editor or an answer engine treats the piece as machine output.

Template phrasing captures the low-surprise wording the Stanford study shows detectors key on. Raw-draft ratio reflects the Graphite and Ahrefs finding that mostly machine-written pages rarely rank or get cited. Attribution gaps and context vacuum track Google’s helpfulness standard and the Article 50 editorial review exception. Exposure venue reflects a simple truth: the higher the stakes at the receiving desk, the more likely someone runs a check.

Read the total this way. A score of 0 to 3 is low risk and ready to send. A score of 4 to 6 is medium risk: fix the two highest factors before pitching. A score of 7 to 10 means rewrite the piece with a human in the chair, because the problem is not detection, it is that the content has nothing only you could have said.

Here is how it plays out on a typical funding announcement. A draft built from a prompt, lightly edited, with a quote from “the CEO” that sounds like every other quote, scores 2 on template phrasing, 2 on raw-draft ratio, 1 on attribution, 2 on context and 1 on exposure for a trade outlet. That is an 8. Add the investor’s name and a real line from the term sheet call, the revenue figure the founder approved, and a rewritten lede in the founder’s own words, and the same release drops to a 3.

Lower your detection risk before a pitch goes out

Start with the person, not the tool. Put a named writer or editor on every piece and have them rewrite the opening, the quotes and the closing by hand. Those three spots carry the most template phrasing in machine drafts, and they are where editors decide in seconds whether to keep reading.

Feed the draft facts a model cannot invent. Dates of real events, figures your finance team signed off on, the name of the customer who agreed to be quoted, a detail from the founder’s first year. Specificity lowers your TRACE score on two factors at once and makes the piece more useful to the outlet that runs it.

Watch the short copy. The Chicago Booth paper showed accuracy dropping on passages under 50 words, and short text is where PR lives: headlines, bios, subheads, boilerplate. Write those yourself, because they get read more than anything else and they are the hardest place for a detector to give you the benefit of the doubt.

Do not run your content through a humanizer to dodge a checker. The RAID results show those tricks can fool today’s detectors, but they make prose worse, they do nothing for Google’s helpfulness test, and they forfeit the editorial review defense under Article 50 if you publish into the EU.

Match the venue to the score. Instant Press standard placements run 1 to 3 days and top-tier outlets 1 to 4 weeks, and the top-tier desks are where a medium TRACE score is most likely to cost you. If you want to know where your current coverage stands with AI answer engines before your next release, run the free AEO rating, then compare packages on our pricing page or browse more research on the Instant Press blog.

Every one of these AI content detection statistics points toward the same habit, which is keeping a human accountable for every sentence that carries your name.

As watermarks spread and detectors sharpen through 2027, the brands that keep earning coverage will be the ones whose content never needed to pass the test in the first place.