AI Content Detection: What Google Actually Checks

A detector asks whether a model wrote it. Google asks whether it is worth reading. Only one of those is the ranking question.

Almost every page ranking for "AI content detection" is selling you something: a detector subscription, or a "humaniser" that rewrites text until the detector stops complaining. Both are answering a question Google has never asked.

This page is the correction. It covers what Google's own documentation says about AI-assisted content, how detectors work and why they fail, what actually gets sites into trouble, and the editorial process we use on AI-assisted drafts. If you want the broader picture of how search is changing, start with our AI search guide — this post handles one narrow, badly-answered question inside it.

One boundary first, so you know what you're reading. What is AI SEO covers using SEO to get found in AI engines. LLM visibility covers being cited by them. This page is about the opposite direction: whether content produced with AI gets detected or punished by Google.

Does Google detect AI content?

Google has never published a claim that it identifies AI-written text, and it does not operate a public detector. What it publishes instead are spam policies and content guidance, and both are written around purpose and value, not authorship.

The relevant policy is scaled content abuse. Google's spam policies documentation defines it as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users," and adds that the practice "is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created."

Read that last clause slowly. It cuts both ways: a thousand AI-spun pages are spam, and so are a thousand outsourced human-written pages with nothing in them. The tool is not the offence.

The same page lists, as one example of the abuse, "using generative AI tools or other similar tools to generate many pages without adding value for users" — note that the condition is without adding value, not generated with AI.

Google's guidance on creating helpful, reliable, people-first content puts the same test in one sentence: "If you use automation, including AI-generation, to produce content for the primary purpose of manipulating search rankings, that's a violation of our spam policies." The trigger is the primary purpose. Not the keyboard.

The claims people make vs what Google documents

Here is the gap between what circulates on LinkedIn and what is actually written down.

How AI detectors actually work

Most detectors score two statistical properties of text. Perplexity is roughly how surprising each next word is to a language model — predictable, low-perplexity text gets flagged. Burstiness is how much sentence length and rhythm vary — human writing tends to lurch between long and short sentences, model output tends to be even.

Nothing in either measurement detects a machine. They detect predictability. Which is a problem, because plain, clear, well-edited writing is predictable by design — and so is the English of a competent writer for whom English is a second or third language.

Why detector scores are unreliable in both directions

Take the most honest datapoint available: OpenAI's own detector. It shipped an AI Text Classifier in 2023 and then withdrew it, with the note "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy." At launch it correctly identified 26% of AI-written English text, while labelling human-written text as AI 9% of the time. The company that built the model could not reliably detect the model.

The false-positive side is worse than it sounds, and it lands squarely on India. A peer-reviewed study in Patterns by Liang and colleagues, GPT detectors are biased against non-native English writers, found detectors "consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified." Across seven detectors, TOEFL essays written by non-native speakers drew an average false-positive rate of 61.22%, while US school essays were classified near-perfectly.

If your content team writes in Indian English, a detector is measuring the wrong thing about them. Firing a writer over a score, or forcing them to rewrite clean copy into something knottier, is a decision made on noise.

Evasion is the other half. The same study found that "simple prompting strategies can not only mitigate this bias but also effectively bypass GPT detectors." A detector that can be defeated by asking the model to write differently cannot be the thing standing between your site and a penalty — which is also why the "humanise your AI content" tool category exists, and why it is solving a manufactured problem.

So: false positives on honest human writing, false negatives on anything lightly disguised. A score from that instrument is not evidence of anything you should act on.

What actually gets penalised

The pattern Google describes is recognisable, and it has almost nothing to do with which tool produced the first draft.

• Volume without value. Hundreds of near-identical pages targeting keyword variants, each one saying the same thing with a different noun swapped in.
• Nothing first-hand. Pages assembled entirely from what already ranks, adding no data, no experience, no test, no opinion. A summary of page one is not a contribution to page one.
• Fabricated facts and citations. This is the real, specific risk of unedited AI output. Models invent plausible statistics, misattribute studies, and cite sources that do not say what the sentence claims. Published unchecked, that is a trust failure your readers can catch.
• Unedited bulk publishing. Thin output at scale, shipped without a human deciding whether each page deserved to exist.

A single AI-assisted post, fact-checked and edited by someone who knows the subject, sits nowhere near any of that. A hundred unchecked ones a month sit inside all of it.

The E-E-A-T angle: experience is the part that can't be generated

Google's self-assessment questions ask, among other things, whether content shows "first-hand expertise and a depth of knowledge (for example, expertise that comes from having actually used a product or service)" — and whether you are "producing lots of content on many different topics in hopes that some of it might perform well."

A language model has read about your topic. It has not run the campaign, seen the account, or watched the thing fail in month two. Experience — the first E — is the one input no model can supply, which makes it the one thing worth spending your editing hours on. Everything else on the page can be assembled; that cannot.

Practically, that means the value you add to an AI draft is the part a reader could not have got from asking a chatbot themselves. If your published page and a chatbot answer are interchangeable, the page has no reason to exist — a point that applies to your whole content marketing strategy, not just to AI-assisted posts.

An editorial process for AI-assisted drafts that survives scrutiny

This is the process, in order. It is also roughly what we run inside our content marketing work — every claim on every page gets checked before it ships.

1. Decide the page deserves to exist before you generate anything. A real query, a real gap, a real reader. If the only argument for the page is that the keyword has volume, stop here — this is the step that separates a content programme from scaled content abuse.
2. Set the point of view yourself. What are you claiming, and what would you cut if the page had to be half as long? A draft written to a stated position reads as an argument; a draft written to an outline reads as a summary.
3. Verify every factual claim against a primary source, and delete what you can't verify. Vendor documentation, the platform's own help pages, the actual study. If a number cannot be sourced this week, write around it rather than guessing — an unverifiable statistic is worse than no statistic.
4. Check every citation actually says what the sentence claims. Open the link. Models hallucinate sources that exist but do not support the point, which is harder to catch than an invented URL.
5. Add what only you have. A number from an account you ran, a mistake you made, a screenshot, a process you use, a considered opinion. If nothing in this step is possible, the page has no reason to be published under your name.
6. Cut the model's tics. Paired-clause padding, throat-clearing openers, three-item lists where two facts exist, and conclusions that restate the introduction. You are editing for a reader's time, not for a detector.
7. Have someone who knows the subject read it. A specialist catches the confidently wrong sentence in seconds — the one a generalist and a detector both wave through.
8. Publish with a named author and a review date. Accountability is a trust signal for readers first, and it happens to be the thing E-E-A-T describes.

Note what is not in that list: running the draft through a detector. There is no step where a score changes a decision, because there is no score that would tell you anything about the page's quality.

Do you have to disclose that you used AI?

Google does not require a label. Its gen-AI guidance suggests you "consider adding information on how your content was created in a way that makes sense for your audience," and the helpful-content questions ask whether automation is "self-evident to visitors through disclosures or in other ways." That is guidance about reader context, not a ranking requirement, and no documentation ties a disclosure to rankings either way.

The reader-trust argument is separate and, in our view, stronger. Nobody is owed a per-paragraph provenance log. But a reader is owed an accurate page, and telling them how you work — what you check, who reviews it — is worth more than a badge saying a human typed it. Note also that some Google surfaces do have specific labelling rules: the same guidance sets out metadata and labelling requirements for AI-generated images and product data in commerce contexts. Editorial blog content is not in that bucket.

Where we stand

We use AI in production. It helps with research, structure, first drafts and the unglamorous parts of a content programme, and pretending otherwise would be its own honesty problem.

What we do not do is publish unverified output. Every factual claim on a Digital Hangover page is checked against a primary source before it ships, every page is edited by someone who knows the subject, and every claim we can source, we link. That is why our posts cite documentation instead of asserting — including this one, where we quoted Google's spam policy and helpful-content pages rather than telling you what Google "says." If a claim cannot be sourced, it does not go on the page.

Content retainers in India typically run ₹25,000–₹1,50,000 a month depending on volume, subject difficulty and how much original research is involved — and the verification work described above is the part that makes the difference between the ends of that range, far more than word count does. The same editorial standard applies to the technical pages covered in our SEO guide.