AI content detection: does Google actually check?

You've drafted a post with AI, it reads well, and now you're hovering over publish wondering whether Google can tell. Short answer: what Google checks for is not what you're worried about. The worry is a fingerprint — some scan that flags machine-written sentences. What Google's policies and its enforcement history actually target is a pattern of publishing. Those are different things, and the difference decides what you should do.

What Google's written policy says

Google's spam policy names the offense precisely: scaled content abuse, defined as 'many pages generated for the primary purpose of manipulating search rankings and not helping users.' Its guidance on generative AI adds that using such tools 'to generate many pages without adding value for users may violate' that policy.

Read those twice and notice what's absent: any rule against AI as an authorship method. The policy language is about volume, purpose, and value. A single useful AI-assisted article violates nothing in it; a thousand pages of templated filler violates it whether a model wrote them or an intern did. Google's position on AI content generally has been consistent on this point: quality is judged on the page, not on how the page was typed.

What enforcement actually looked like

Policies are cheap; the March 2024 update showed the enforcement. Google announced it expected the update to 'reduce low-quality, unoriginal content in search results by 40%' — and when the rollout finished that April, reported the real figure came in at 45%.

During that update, the detection company Originality.ai — a vendor with an obvious interest, though the numbers were reported and scrutinized by the search press — tracked 49,345 sites and found 837 of them deindexed entirely: not demoted, removed. Every deindexed site they examined showed signs of AI-generated content, and half were 90 to 100 percent AI-written.

Here's the detail that answers this article's question. The sites that vanished weren't blogs that used AI to help write good posts. They were sites running one identifiable playbook: hundreds or thousands of pages, published at a pace no human reviewed, targeting queries wholesale. You don't need a text-fingerprint detector to find that — publishing velocity, site structure, and the absence of anything a reader would stay for are all visible signals. Google was hunting a farming pattern, and AI text was the tractor, not the crime.

So can anyone detect AI text reliably?

The commercial detectors — the tools that score your text '87% likely AI' — deserve their own skepticism, because their track record is messier than their marketing.

• The vendors' own numbers and independent results diverge. Detection companies publish studies showing their tools scoring in the high nineties; the same tools tested by outsiders, on text lightly edited or from newer models, come back notably worse. When the seller of a detector is also the main source of its accuracy studies, price that in.
• OpenAI built a detector for its own models' output and shut it down within months because its accuracy was too low to stand behind. That's the company with the most inside knowledge of the text, declining to certify detection.
• The false positives land on real people. A Stanford-led study found detectors consistently misclassified writing by non-native English speakers as AI-generated — plain, correct, human-written prose, flagged for being plain. Any tool with that failure mode will also flag clear, simple business writing.

None of this means detectors detect nothing — obvious, unedited model output does get caught. It means a probability score from a tool with known false positives, sold by the party reporting its accuracy, is not something Google would or could hang a penalty on. And nothing in Google's public statements claims they do: there is no published Google figure for AI-detection accuracy, no confirmed case of a site penalized over authorship alone.

There is no confirmed case of a site penalized for AI authorship alone. Every documented removal involved scale without review.

What this means for how you publish

The practical rules fall out of the enforcement pattern, and they're cheap to follow:

• Volume with review beats volume without it, at any ratio of AI involvement. Everything documented about the 2024 wave points at unreviewed scale, which is the same reason most AI content fails even without penalties.
• Put something on each page that no model could have produced: your price, your photo, your result from last month. It's the strongest ranking argument and the clearest not-a-content-farm signal at once.
• Skip the 'humanizer' tools that rewrite AI text to fool detectors. You'd be paying to evade a check Google isn't running, usually making the prose blander in the process, while doing nothing about the thing Google does police.
• Spend the freed-up worry on accuracy instead: models state false things confidently, and a wrong number on your pricing page costs you customers regardless of what wrote it.

The answer, in two sentences

Google has never claimed to detect AI text as such, and the detectors that do claim it are unreliable enough that its own maker retired one. What Google demonstrably detects — and in 2024, mass-removed — is scaled publishing of pages nobody reviewed and nobody needs; stay out of that pattern and the question of what wrote your sentences stops mattering.