Google does not punish AI writing, and it does not give it a pass

Two answers circulate, and both are wrong

Ask this question online and you get one of two confident replies.

The first says Google penalizes AI content, so keep the machine away from anything you publish. The second says Google does not care how content is made. So publish as fast as the tool will write.

The first is wrong on the policy. The second is wrong on the evidence.

The useful version sits between them, and the evidence for it is public. Google runs no AI penalty. Google also gives AI-assisted pages no advantage and no protection. Pages made by pointing a model at a keyword list perform measurably worse. Both halves are documented, and the second half is the one nobody selling an AI writing tool mentions.

What Google's policy says, in its own words

Google published its position on AI-generated content in February 2023 and has not moved off it. Appropriate use of AI or automation is not against its guidelines. The company pointed out that automation had long produced useful content. A decade earlier the worry was mass-produced human writing. Nobody proposed banning that either.

The rule that does apply arrived in March 2024, when Google renamed its old auto-generated-content policy scaled content abuse. The policy covers generating many pages to manipulate rankings rather than help users. It applies no matter how the pages were made. Its list of examples names generative AI first. One qualifier carries the whole rule: generating many pages without adding value for users.

Google's guidance for creators using generative AI says the same thing in shorter form. It calls the tools useful for researching a topic and for adding structure to original content. It then warns that generating many pages without adding value may break the policy.

On May 15, 2026, Google clarified that its spam policies cover generative AI responses in Search. The same rules now govern what appears inside AI Overviews and AI Mode. Google added no new rules. It extended the existing ones to a new surface.

So the policy question has a short answer. Volume aimed at rankings is the violation. The tool is not.

What 331,000 pages show

Policy is what a company says. The more useful question is what its systems do. In July 2026, Ahrefs published the largest recent look at that. The team pulled a million pages from the top 10 positions across 100,000 search results. It ran the ones with enough text through its own AI content detector and reported the distribution.

Fully AI-written pages rank at the top. Around 5% of pages in positions one through three were classified as entirely AI-generated. Pages that were at least 80% AI came to 9%. Every position in the top 10 held a meaningful share of heavily AI-written pages. If a filter existed that blocked AI text from ranking, those pages could not be there.

The same study found the other half of the answer.

Pages low or moderate in AI content drew two to three times the search impressions. Pages rated high or very high drew the smaller share. Indexing showed the same slope. About 49% of low-AI pages were indexed, against 40% of the highest-rated AI pages.

No cliff appears anywhere in the data, and no page gets a free pass either. Performance slides gradually as AI content rises. The likeliest reason is the one Ahrefs offers. Quality tends to fall as AI use rises, and Google measures quality.

One caution on all of this. AI detectors work on probability. Ahrefs says openly that its detector will not match whatever Google does. Google may do nothing of the kind. The numbers are useful as a large-scale trend and useless as a verdict on any single article.

What sinks AI pages is what sinks any page

Read the failure patterns and the mystery goes out of the subject. Basic AI output repeats what the top results already say. It adds no information to the page. It skips links, images, and firsthand detail. It writes in a flat academic register, and it slips in errors nobody caught.

Every one of those problems would sink an article a person wrote badly. They cluster around AI use because most people take the first draft and publish it.

That is also why the failure looks so dramatic when a site publishes at volume. Google crawls a burst of new pages and finds little reason to keep them. The pages come back out of the index over the following months. The traffic chart climbs and then falls off a shelf, which is why practitioners call the shape Mount AI.

The instinct behind the fear is sound. Something does go wrong with mass-produced AI content. The cause is the mass production, and the tool only made it cheap.

The enforcement that does exist

Manual actions are real, and they land on scale.

In April 2026, Glenn Gabe documented a site with a directory of 850,000 AI-generated local news pages. The section ranked in Top Stories and Google News. Then it took a manual action for scaled content abuse, and the whole directory left the index. The knock-on effect is worth noting. Its visibility inside ChatGPT fell alongside its Google visibility, because assistants lean on the same web.

Now put your own publishing next to that. A firm publishing two or three considered articles a month sits in another category. Each of those articles carries its author's judgment. The scaled content abuse policy was not written about you. Worrying about it is worrying about someone else's problem.

Your risk is different and quieter. It is publishing a page that reads as competent and adds nothing. That costs you the reader rather than the ranking. Whether anyone can tell a machine helped matters far less than whether the page was worth their time.

What this means for a practice publishing a few pages a month

Human input decides the outcome, and it goes in before the draft rather than after.

A model can organize, draft, and tighten. What it cannot supply is the material that makes a page worth ranking. That material is the figures from your own matters. It is the situations you have handled, and the judgment you would give a client who asked. Feed those in and the output has something the rest of the web does not. Skip them and the output is an average of pages that already exist. That is the definition of adding no value.

Google's framework points at the same input. Its four quality signals now lead with experience, and experience is the one input a content operation cannot manufacture.

Three habits carry most of the weight. Publish under a byline that leads somewhere. Check every claim, because a model states a wrong figure as fluently as a right one. And publish on the schedule your material supports, rather than the schedule the tool makes possible.

That last one is the whole safety margin. Sites that got hurt were not hurt for using AI. They were hurt for publishing more pages than they had things to say.

The page still has to earn its result

Google sorts competent from useful without knowing who typed the words. Whether it can tell is beside the point.

What decides the page is what it adds to a result that already holds 10 competent answers. When the addition is a figure, a case, or a judgment only you have, publish it. How it got written does not matter. When there is no addition, no amount of editing saves the page. No penalty is needed to bury it.