AI-Generated Text Detection

If you've worked with patents before, skip this part and go to the next heading. Everyone else needs a short explanation, because most of this article is built on patent language.

Companies aren't obligated to answer our questions. Not Google, not OpenAI, not anyone. Whatever they say is entirely at their own discretion, and they can change their story whenever they want. That's why quoting a tweet or a podcast never counts as solid ground.

But there's one exception. When a company wants to own a method for itself, it's forced to describe that method precisely and hand it to a government office. If the description is vague, the patent doesn't get granted. Which makes a patent one of the few places where a company is speaking in its own interest and is still forced to be precise.

That's also the difference with a journal paper. A journal paper is something a researcher wanted to publish. A patent is something a company was forced to publish to get something in return. The two are nothing alike in motive, which is why a patent reveals things about what a team actually thought and built that usually don't get said this plainly anywhere else.

And Why They're Hard to Read

A patent is written in legal language, not technical language, and it's deliberately stretched to keep the scope of ownership as wide as possible. A patent has two pieces. The description, which can run to a hundred pages and can contain any claim at all, and the claims, which come at the end of the document and are the only part that carries legal weight. The whole analytical task is figuring out which sentence sits inside an independent claim and which one is just talk in the description.

You'll see a real example of this later in the article. For Originality.ai's patent, I had to compare independent claim 1 against dependent claim 3 to work out that the AI-detection step has no method attached to it at all. If I'd only read the abstract, I would have concluded the exact opposite.

If This Is Your First Time Opening a Patent

1. Number and title. The last letter of the number matters. A "B" means granted. An "A" means it's only a published application, not yet granted.
2. Current assignee. Different from the inventor. The inventor is a person, the assignee is a company, and the assignee may have changed since filing.
3. Priority date. More important than the publication date, because it shows when the idea was actually filed, not when it became public.
4. Legal status. Granted means accepted. Abandoned means the owner stopped pursuing it. Expired means the maintenance fee wasn't paid. These three mean completely different things.
5. Claims. An independent claim is complete and stands on its own. A dependent claim (dependent claim) refers back to another claim and only narrows it further. Anything not in the independent claim isn't guaranteed ownership.
6. Citations. Backward citations show where the idea came from; forward citations show who built on it later. This is where you can trace the actual trajectory of a technology.
7. Continuations. If a patent has a continuation, the owner kept pursuing it even after the grant. That can be a sign the company took it seriously.

Let's start where everyone assumes the answer lives. The General Guidelines for Search Quality Rating, dated September 11, 2025. Google's own human raters use this document to score pages, and it's published publicly.

The only place in this document that touches content signals is section 4.6.7. Four signals are listed, and all four are about rewritten content, not AI style. First, that the text only contains generic, well-known information. Second, high overlap with an established source like Wikipedia. Third, that it reads like a summary of one specific page with no added value.

And the fourth is the only textual signal Google attributes to AI anywhere in this entire document. Leftover assistant phrasing in the published text: a sentence starting with something like "as an AI language model." That's it. No coefficient, no word list, no sentence structure.

You'll find the second signal in section 4.5.3, under the heading Deception. There, Google talks about a fabricated author profile: AI-generated content published with a generated photo and a misleading author bio designed to make it look human-written. Again, this is a signal outside the text itself, on the author page.

There's one more line that basically gives away the whole policy. In section 4.6.5, Google tells raters that even if you're not sure a specific page was produced with a generative tool, if after seeing several pages from that site you suspect scaled content abuse (scaled content abuse), give it the lowest score. Google is telling its own human raters not to go looking for AI detection: look for the pattern at the site level, not in a sentence's style.

And Two Documents Google Published Outside of Search

Everything up to this point has been policy: something that directly affects ranking. There are two more documents worth setting alongside it, with one important difference: these two aren't policy, they're research. Google published them as a research organization, not as ranking guidance, and that changes the level of the claim. Policy tells you what gets scored; research tells you what was measured. But because both bear directly on the question of AI text detection and both officially come from Google, they belong here.

The first is the paper by Ippolito et al. at ACL 2020, pages 1808 to 1822. It's listed in Google's own official research publications, so it's the company's own output, not independent academic work.

Their core finding is that humans and machines look at different cues. Sampling methods used in text generation have been optimized mainly to fool humans, and that same optimization introduces statistical irregularities that make the text easy for an automated system to catch.

In plain terms, it's an inverse relationship. The more natural a text reads to a person, the easier it is for a machine to catch.

The second document belongs to DeepMind, and its whole direction is the opposite. The paper by Dathathri et al., in Nature volume 634, pages 818 to 823, 2024, describes a method called tournament sampling.

The logic is that instead of hunting for a fingerprint after the fact, they embed a signal into the text during generation itself: a watermark. It ran in a live experiment with close to 20 million Gemini users, and the reference code has been open-sourced.

It has two limitations worth stating explicitly. First, it only recognizes output from Google's own models and does nothing for text from any other model. Second, the public verification tool is gated behind a waitlist for journalists and researchers, and that restriction covers text too.

But what matters for our purposes is the choice itself. When the lab that builds the models chooses watermarking over pattern detection as its solution, that could itself count as a kind of institutional admission. That's my own reading, though, not something DeepMind itself has said.

Now let's get to where actual numbers exist. The largest grammatical study in this field comes out of Carnegie Mellon University and was published in PNAS in February 2025. They built two parallel corpora, roughly 66,000 and 77,000 text pieces, six models spanning four Llama variants and two GPT variants, and compared human and machine text using Douglas Biber's set of 66 linguistic features.

The overall direction of these numbers says one thing. Model text is noun-heavy and information-dense. Present participial clauses appear 5.3 times as often as in human text, clauses opening with "that" as the subject 2.6 times, nominalizations 2.1 times, phrasal coordination 1.9 times, and the agentless passive at roughly half the human rate.

Vocabulary tells the same story. The same paper lists its highest-frequency words. Relative to human text, the word "camaraderie" appeared 162 times as often, "tapestry" 155 times, and "intricate" 119 times. The second of those showed up in 23% of outputs, and "amidst" in 27%.

But That's Not the Paper's Most Important Finding

The most important finding is that base Llama models write almost like humans, and these patterns only show up after the instruction-tuning stage: what we call "AI tone" isn't a byproduct of being a machine, it's a byproduct of the instruction-tuning stage specifically.

A second point nobody quotes: when they tested those same trained classifiers on a different corpus and genre (scientific abstracts), the accuracy of the models trained on base Llama output dropped to 50%. Exactly random guessing. The pattern works inside its own domain and not outside it.

The authors themselves wrote in the discussion that their goal was never to build a new detection tool. The patterns were there to understand the difference, not to hand down a verdict.

And the Layer Commercial Tools Renamed

These grammatical patterns are one side of the coin. Another study answered the same question from a different angle, not through grammar, but through distributional statistics. The paper by Muñoz-Ortiz et al., in Artificial Intelligence Review, volume 57, issue 10, 2024, compared six models across three families and landed on distributional patterns rather than grammatical ones.

This is exactly what commercial tools rebranded as perplexity (perplexity) and burstiness (burstiness) and sell as their own proprietary method. The root of it is a public academic paper.

Two More Studies, at a Much Larger Scale

Everything above operates at the sentence and style level. Two more studies exist at a much larger scale, tracing word-level fingerprints across millions of documents instead of writing patterns. A July 2025 paper in Science Advances examined more than 15 million PubMed medical abstracts from 2010 through 2024 and found 379 style words with sharply elevated frequency. Their estimate is that at least 13.5% of 2024 abstracts were processed through a language model, and up to 40% in some subsets.

And a 2024 ICML paper examined academic peer reviews. In ICLR 2024 reviews, the probability of the word "commendable" appearing in a sentence rose 9.8x, "meticulous" 34.7x, and "intricate" 11.2x. Their estimate was that between 6.5% and 16.9% of review text had been substantially modified by a language model.

This is the new territory. I went to the patents because, unlike a blog post, a patent's claims have to be precise and defensible or it doesn't get granted. These patents trace the full path.

In 2012, "machine text" meant scraping, aggregation, and machine translation, and the patent's method is identifying and quantifying various aspects of a work, without specifying which aspects. This patent also has a continuation, meaning Amazon filed a follow-on after the first grant.

Its content is worth an SEO's time. It defines two things. First, a list of domains, URLs, and IPs previously identified as sources of AI content. Second, a list of patterns. And its figures show how a search results page could change to display the AI-content percentage of each result.

The only methodological explanation anywhere in the patent's description is that the distinctions can range from common repetitive patterns typical of AI output to the absence of emotional undertones usually present in human writing. The patent names no threshold, no feature, and no number.

Legally, this SEO step sits inside the independent claim, not a dependent one, meaning it was necessary to obtain the patent. The AI-detection step alone wasn't specific enough to be patentable against prior art.

The Five Core Documents in This SpaceLegal status and priority date

One more document here takes a completely different path. Patent US11853708B1, granted in 2023, whose inventor is a private individual. Instead of analyzing the text, it selects relatively uncommon words and phrases from it and asks whoever claims authorship to demonstrate they understand those same words. It's measuring the author, not the text.

The largest independent multi-tool test in this field is a December 2023 paper in the International Journal for Educational Integrity. Fourteen tools, 54 documents with known ground truth, 756 tests in total. Breaking results down by document type is not something any tool puts on its own website.

Human English text: 96% accuracy. Raw AI text: 74%. That same text with a light human edit: 42%. And that same text after a machine paraphrase: 26%. No tool in the entire test reached 80% accuracy, and only five cleared 70%.

One finding matters enormously for a Persian-speaking audience: human text originally written in another language and then machine-translated into English dropped accuracy by 20%. Machine translation leaves an AI fingerprint on human-written text.

The false positive rate across tools ranged from 0% to 50%, and the false negative rate from 8% to 100%. Meaning one of the tools failed to catch a single piece of AI text.

And the One Tool Left Out of This Test

One detail here circles right back to this section's core question: why a tool's number isn't evidence. In that same 14-tool paper, the tool GLTR was excluded from testing, because it doesn't issue a verdict at all; it only shows each word's predictability, color-coded. The one tool left out of the test was precisely the one that didn't hand the user a definitive number or percentage.

Alongside it are two other open-source tools that work zero-shot: DetectGPT and its faster successor. Neither gives a final number either, and their methods are inspectable, unlike commercial tools that just show you a percentage.

Bias Against Non-Native Writers

A July 2023 Stanford paper in the journal Patterns. Seven commercial tools tested against 91 TOEFL essays. More than 61% of those human-written essays were flagged as AI-generated, while accuracy on essays by American eighth-graders was close to perfect.

The authors explained the cause themselves. Non-native writers' text has lower perplexity (perplexity) because it has less vocabulary variety, and that's precisely what the tool reads as a machine signal. When they asked a model to enrich the vocabulary of those same essays, the AI label dropped.

A 2025 paper in ACL Findings built a dataset of roughly 14,700 samples and tested twelve current tools against it. The result: tools consistently flag human text that's only been lightly AI-edited, and they can't separate degrees of involvement at all, which matches what most writers actually do.

And finally, the RAID benchmark at ACL 2024. More than 6 million samples, eleven models, eight domains, and eleven adversarial attacks. The first line of their abstract goes straight at this same issue: many commercial tools claim 99% accuracy or higher, but few of them have been evaluated on a shared dataset. When they were, they were easily fooled by adversarial attacks and unseen models.

This is where the whole picture changes for us, and nobody in the Persian-language market has talked about it.

Copyleaks lists thirty languages for AI detection in its own official documentation. is there. Turkish is there. Hindi is there. Persian is not on that list. When you paste in Persian text, the model is working on a language it was never trained for, and it still hands you back a confident number.

GPTZero officially fully supports English, German, Portuguese, French, and Spanish, and recently added Arabic, Korean, Japanese, Chinese, and Italian. Persian is not on their official list.

The only large commercial tool with official Persian support is Originality.ai. Their multilingual model, released October 28, 2025, reports these numbers for Persian: 98.91% accuracy, 1.18% false positive, 1.00% false negative.

But treat that number with skepticism. It's their own benchmark, on their own dataset of 127,150 samples that they built themselves, the exact setup the RAID benchmark warned about. No independent third-party evaluation of this tool's Persian performance exists.

The Only Academic Benchmark That Includes Persian

A genuine independent evaluation does exist, though not of Originality.ai (of a separate model entirely). A Technical University of Munich paper at the NLP4PI workshop, EMNLP 2024. Four languages: English, Turkish, Hungarian, and Persian, with generators including BLOOMZ, Llama, Mistral, and GPT-4.

In the out-of-domain test, Persian scored highest of all. For GPT-4 text, for instance, the F1 score was 0.99, compared with 0.45 for English and 0.34 for Turkish. But the authors wrote their own explanation for this. It's probably down to Persian texts being shorter. One of the models also couldn't even be trained on Persian news data and they had to bring in a separate model, and most of the length outliers came from Persian outputs too.

In plain terms, that 0.99 doesn't mean Persian is inherently easier to detect. My own reading is that it means 2024-generation models wrote clumsier Persian, and that clumsiness was easier to spot. Today's models write more fluent Persian, and that benchmark no longer holds in the same way.

One more supporting document. A paper indexed as arXiv 2401.12070 states explicitly that in low-resource languages, the false-positive rate stays low, but machine text tends to get classified as human, so recall (recall) is weak.

A Persian Model You Can Run Yourself

That same TU Munich team released the classifiers built for that paper under an open license on HuggingFace. Its Persian version is built on XLM-RoBERTa and has 125 million parameters. Its training dataset is public too.

But it has one fundamental advantage over commercial tools. Here you can actually see the weights, the training data, and the method. There, all you get is a number, with no explanation of where it came from. And as far as I've seen, nobody in the Persian market has introduced this model.

After all this talk of grammar and tools, it's worth stepping back and looking at the bigger picture. Two large studies have been run on the web itself.

Ahrefs counted 900,000 new pages in April 2025. 2.5% fully AI-generated, 25.8% fully human, and 71.7% hybrid: the dominant state of the web isn't one side or the other, it's mixed.

A second study looked at 65,000 URLs from a public web corpus. The share of predominantly AI-generated articles in May 2025 was reported at around 51.7%.

But the number that's directly useful to an SEO is the third one. According to that same study, 86% of articles that rank in Google are human-written, and 82% of the articles chatbots cite are also human-written.

So machine output has gone up in volume, but its share of results hasn't gone up with it.

There's a common assumption that big platforms haven't done anything yet and machine content is spreading freely everywhere. That's not true. YouTube, Google, LinkedIn, Amazon, and Spotify all have systems in place, and all of them are documented and public. But when you read their documents one by one, a shared pattern emerges that differs from what the market thinks.

None of these platforms are trying to figure out whether a text was written by a person or a machine. What they target is behavior and scale, not the tool.

Five Platforms, Five Public DocumentsWhat each one targets

YouTube, the most explicit document in this space

On July 15, 2025, YouTube renamed its "repetitive content" policy to "inauthentic content" and clarified that the policy covers templated, mass-produced content too. Media outlets headlined this as YouTube demonetizing AI content, but YouTube's own text doesn't say that at all.

Its second document is even clearer. In March 2024, YouTube made disclosure of altered or synthetic content mandatory, but only for realistic content that a viewer might mistake for reality. And in that same announcement, it drew the line explicitly.

LinkedIn, the newest and closest document to our work

On May 20, 2026, LinkedIn announced it had launched a system to detect low-quality machine content. Laura Lorenzetti, VP and Global Editor-in-Chief at LinkedIn, explained it in a post titled Keeping conversations real.

Read the details closely, because this has a direct effect on anyone publishing on LinkedIn. The post isn't removed, but it's cut from being suggested outside your followers' network, meaning it stays visible to your own followers and not to anyone else. Three things are targeted. Generic posts with no clear point of view, bot comments or comments that just repeat the post itself, and automation tools that produce at scale.

The method is disclosed too. A human editorial team manually labeled thousands of posts, and the model was trained on that same data. LinkedIn claims that in early testing, it correctly identified generic machine content 94% of the time.

And what LinkedIn did with its own tool two months later

On July 30, 2026, LinkedIn added an option to every post's menu letting users flag content as generic machine content. Being flagged doesn't get a post removed, but it feeds the same classifier and reduces distribution outside your own network.

But the more significant move came the next day. LinkedIn completely removed its own writing tool, sold to paying users under names like Write with AI and Improve this post, and replaced it with a proofreader that only fixes grammatical errors and doesn't touch tone. Meaning the same company that, until the day before, was selling its users a text-generation button removed that button itself.

Another number in that same announcement shows the scale. LinkedIn said it blocks hundreds of thousands of attempts to post automated comments every day.

Amazon and Spotify, the same pattern in two other industries

In September 2023 Amazon did two things at once. It made disclosure of AI-generated content mandatory at publishing, and it lowered the daily cap on publishing new titles. Amazon draws a distinction between AI-generated content and AI-assisted content. If you wrote it yourself and the tool only edited or brainstormed, no disclosure is required at all.

On September 25, 2025, Spotify announced it had removed more than 75 million spam tracks in the previous twelve months and put three new policies in place. A ban on impersonating an artist's voice, a spam filter for mass uploads and metadata manipulation, and support for an industry-standard disclosure in music credits that specifies exactly where AI was used, for example only in the mix or only on one instrument.

Spotify itself explicitly wrote an important limitation. Because this disclosure depends on the artist's own declaration, the absence of a label on a track doesn't mean AI wasn't involved.

The shared pattern across all five platforms

1. None of them banned tool use. All five state explicitly that the tool itself isn't a problem.
2. What's targeted is scale and behavior. Mass uploads, templated repetition, automation, impersonation.
3. Wherever they did go after the content itself, the standard was deception, not authorship. YouTube only subjects realistic content to disclosure.
4. The usual penalty isn't removal, it's reduced distribution. LinkedIn keeps the post up and just stops suggesting it. Spotify's filter also operates at the recommendation level.
5. And wherever they wanted precision, they went to disclosure by the creator, not detection. Amazon added a checkbox, Spotify a metadata standard, YouTube a disclosure toggle in Studio.

If you've read this far, the logic of the checklist should already be clear. We need to move from style detection to claim verification. The reasoning is simple. Machine content doesn't actually fail on style, it fails on untraceable detail. And that's roughly the same philosophy Google itself gives its own raters, look at behavior, not style.

1. Look for a falsifiable claim. Machine text is full of sentences that are correct but empty. If you read ten paragraphs and there's nothing you could prove wrong, that itself is a signal.
2. Isolate first-hand experience. A sentence only someone who actually did the work could write, like a number from their own dashboard or a mistake they made, is almost never present in machine text.
3. Check for internal contradiction. Long machine text usually says one thing early on that doesn't line up with something later.
4. Open up every number, date, and study name. If the text cites a study, go check that the study exists and that the number inside it matches.
5. Trace the author. Do they have a real name, do they write elsewhere, is there a way to contact them, do they have a track record. That's the same logic Google itself follows in section 4.5.3 of its guide, just from the reverse angle.
6. If you're going to run a tool, run at least two and treat the number as a lead, not a verdict. And if the text is in Persian, first check whether that tool supports Persian at all.

The fourteen-tool study reached the same conclusion. It wrote that these tools' reports provide no evidence and can't serve as the sole basis for an accusation, and that anyone accused on that basis alone has no way to defend themselves.

I always write this section, because without it this article would just become one more unsourced claim itself.

First, these documents don't prove Google penalizes AI content. Google's official documentation never claims to detect AI text anywhere, and it states explicitly that using a generative tool by itself doesn't determine the score. What Google's policy actually targets is mass page production with no added value, regardless of how it was made.

Second, they don't prove detection is impossible. They only show that, as of today, no detection method has been precise enough to stand alone as a record, and no tool has reached reliable accuracy in independent testing.

Third, and most important, none of these documents render a verdict about one specific document, Persian or otherwise. The PubMed paper needed 15 million abstracts to reach its figure. The PNAS paper works within its own corpus and drops to random guessing outside it. These are population-measurement tools, not tools for judging a single document.

Fourth, the figures on machine content's share of the web aren't definitive either. Both studies were run by companies with a stake in the outcome, and both were counted with the same tools whose error rates this article documented.

And one last point. Every grammatical figure in this article belongs to the 2024 generation of models. Today's models aren't the same ones. Every pattern you read here has an expiration date, and I think that date is close.

From here on, it's not documented. Everything you've read up to this point was tied to a patent, an official document, or a peer-reviewed paper. This section is my own personal opinion, and it's deliberately written apart from the rest.

Let me be clear about one thing first. I'm not against tools. I use them myself, and part of the source-gathering for this very article was done with a tool. Anyone who tells you they don't use tools is, in my view, either lying or working slower than they need to.

My point is something else. The question the market asks is wrong at its foundation. "Was this text written by AI or not" has no answer you can rely on, and this entire article demonstrated exactly that. The right question is, if a sentence in this text turns out to be wrong, who's accountable for it.

These two questions are completely different. The first is about the tool, and no tool gives you a definitive answer to it. The second is about the person, and its answer is always knowable.

Why I think chasing a percentage is a trap

When a tool tells you seventy-three percent, you've gotten a number and nothing else. You don't know where it came from, you can't show it to anyone, and you have no way to prove it wrong if it is wrong. The fourteen-tool study wrote exactly this, that anyone accused on the basis of these numbers has no way to defend themselves.

Another layer stings more for us specifically. The Stanford paper showed that more than 61% of non-native writers' texts were mistakenly flagged as machine-written: the very system that's supposed to deliver fairness is stacked against us from the start. If these numbers become the standard of judgment in the Iranian market tomorrow, the one who loses the most is the Iranian writer, not the spammer.

Now that LinkedIn has put the report button directly in users' hands, I don't think this risk is purely theoretical anymore. Flagging no longer even needs a tool, it just takes someone's prose reading as unfamiliar to someone else.

So when I say don't chase the tool, it's not out of naivety. It's because in our hands, this tool works against us.

What I do myself

In the reportage campaigns I run for clients, I read every piece of content before it goes live. But my standard isn't any tool's number. I look at three things.

1. Whether the claim in the text is verifiable or not. If there isn't a single sentence in it that could be proven wrong, the text isn't worth much even if a person wrote it.
2. Whether there's something in it that only someone who actually did the work could write. A number from a dashboard, a mistake they made, a detail that's in no public source.
3. Whether there's a traceable person standing behind the text or not.

All three of these standards are independent of the tool. A text that has all three, I don't care what tool it was typed with. A text that doesn't, I reject even if it was written entirely by hand.

As it happens, the platform documents say exactly the same thing. None of them banned the tool, all of them went after scale, behavior, and deception. LinkedIn wrote explicitly that the ultimate value comes from the person behind the tool. That's the same thing I've been saying in my classes for years, just now a major platform has written it too.

What I think happens next

I don't think the path of style detection goes anywhere. Even the few examples we've seen in this article show that some players are already heading toward other paths. DeepMind went for watermarking. Amazon and Spotify went for disclosure by the creator. LinkedIn removed its own writing tool and replaced it with a proofreader. These examples show, at minimum, that the practical answer isn't just finding the text's linguistic pattern.

What becomes more valuable is exactly what a tool can't produce. First-hand experience, a real number from a real project, and a mistake you actually made and are willing to write about. No model can be asked for these, because they're in no training data.

One last line. If you're worried your content will read as machine-written, don't spend your time polishing sentences. Add something to the text that nobody but you could have written. That one sentence does more than any cleanup tool.

Everything you read in this article came from one of these documents. None of it was quoted from an SEO blog.

• Official documentDocumentation published by the companies themselves, meaning rater guidelines, Search Central, and official platform announcements
• PatentA granted patent or a published application at the US Patent and Trademark Office
• ResearchA peer-reviewed paper or a reputable conference paper with public access
• Industry researchResearch published by a commercial company that itself has a stake in this same market

Research grounded in documentation. Written by Shahram Rahbari