Reading time: 10 minutes
Google does not say in its official guidance that it can detect AI content to demote pages. It talks about fighting spam and about quality, whoever or whatever writes the text. An AI detector, meanwhile, errs even on human writing. You can humanize AI text and change that score, but not what Google’s policy asks.
In short
- Google says it finds spam through patterns and signals, whoever or whatever writes the text.
- SynthID, from Google DeepMind, marks what Google’s tools generate. It does not read text from any source.
- In academic studies, detectors flagged human writing as AI and failed on paraphrased text.
- To humanize AI text changes the detector score. It adds no fact, experience or value for the reader.
Does Google detect AI content in the pages it ranks?
Google does not say it can detect AI content to rank pages. In its FAQ on AI content, it says it analyzes patterns to find spam, however someone produces the content. In other words, the stated target is spam, not who wrote the text.
Here is the answer in the Google Search Central FAQ on how Search tells whether AI is used to spam results:
“We have a variety of systems, including SpamBrain, that analyze patterns and signals to help us identify spam content, however it is produced.”
Google Search Central (2023)
On the same page, Google says its focus is on the quality of content, not on how someone produces it. So when someone claims that Google can detect AI content and punish it, the claim goes beyond the documentation. On the other hand, the documentation does not explain how those systems work inside.
The target is clear. Under Google’s spam policies, scaled content abuse means generating many pages to manipulate rankings, not to help readers. That holds however someone creates the content. So mass-produced human text with no value breaks the same rule as AI text.

What is SynthID, and where does Google detect AI content with it?
SynthID is Google DeepMind’s digital watermark for AI-generated content. Google’s tools add it when they generate the content, in images, audio, text and video. So Google can detect AI content with SynthID only where the mark exists, not in any text on the web.
According to the SynthID page, the mark is invisible to people, but the system can detect it. For text, it applies to what the Gemini app and web experience generate. To do that, SynthID adjusts the probability score of each word the model picks. In 2025, Google launched the SynthID Detector, a portal that verifies whether a file carries the mark.
The SynthID Text documentation shows the limits for text. Google says the mark survives cropping, a few changed words and mild paraphrasing. However, it works less well on factual answers. Detector confidence also drops a lot when someone fully rewrites or translates the text.
Google open-sourced SynthID Text for developers. Even so, the mark lets whoever generates the text check if it came from their own model. So SynthID says nothing about text without the mark, such as text from a writer or from another AI tool. Also, Google does not say it uses SynthID for ranking. The Search Central pages on AI content, spam and helpful content do not mention it.
Are AI detectors reliable?
Not on their own, to decide whether a text is human or machine. Studies with published methods show two errors. An AI detector flags human text as machine text, and it lets rewritten AI text pass. Even a company that built a detector admitted those limits.
OpenAI is the clearest example. In January 2023, it launched an AI text classifier and withdrew it in July of the same year, because of low accuracy. On a set of English texts from the company itself, it caught 26% of AI-written text. At the same time, it labeled 9% of human text as AI. OpenAI said it should not be the main decision-making tool.
Independent studies point the same way:
- Weber-Wulff and others (2023): in a test of 14 tools, including Turnitin and PlagiarismCheck, all scored below 80% accuracy on ChatGPT texts. Only 5 scored over 70%.
- RAID (2024): the benchmark from the University of Pennsylvania and partners tested 12 detectors on more than 6 million texts. Many detectors, commercial and open, claim 99% accuracy or more. Yet the 12 tested failed under attacks, new models and other generation settings.
Put simply, an AI detector measures statistical patterns in text, such as how predictable the words are. It does not measure whether content is useful, correct or original. The studies use detectors from 2023 and 2024, so the results cannot be stretched to every tool today. Still, the idea that Google can detect AI content from style alone has little support in these studies.
Why does an AI detector flag human writing as machine text?
According to the authors, it happens because text with a limited vocabulary is more predictable. The detector reads that as a machine signal. In a Stanford study, seven detectors flagged 61.22% of 91 human-written TOEFL essays as AI, on average. TOEFL is the English exam for non-native speakers. Essays by US 8th graders came out almost error-free.
In all, at least one detector flagged 89 of the 91 essays as AI. So the authors urge caution when using detectors in grading and teaching. The samples were all English writing, so the study says nothing about other languages.
Does it help to humanize AI text?
It fixes the detector score, not what Google’s policy asks. To humanize AI text is to rewrite sentences so they avoid the patterns a detector looks for. But Google judges value for the reader. So a text with no new fact still has no new fact after the rewrite.
Studies show that you can humanize AI text to fool a detector with little effort. In the same Stanford study, a request to ChatGPT to rewrite with literary language cut detection of 31 AI-generated essays. The rate fell from up to 100% to up to 13%, depending on the detector. Likewise, Krishna and others built a paraphraser that cut DetectGPT’s accuracy from 70.3% to 4.6%. The false positive rate stayed fixed at 1%, and the meaning of the text barely changed.
The market sells both sides. Undetectable AI, for example, sells a detector and a humanizer that promises text scoring as human on every major detector. It also lists a text watermark remover. Grammarly, which also offers a detector and a humanizer, says it does not design its humanizer to bypass detectors.
On Google, though, you can humanize AI text and still leave the content as it was. The spam policies give an example of scaled content abuse: copying content and transforming it with synonyms, translation or other obfuscation techniques. So running many pages with no value through a humanizer does not take them out of that rule. Publishers who work in bulk will find the risks of scaled content in our guide on automated blogs with AI.
If Google does not detect AI content by origin, what does it assess?
Google assesses usefulness, originality and signals of experience, expertise, authoritativeness and trust. In its FAQ, it says content that is useful, original and shows those signals might do well. So the origin weighs less than what the text gives the reader.
On the same page, Google sums it up: using AI does not give content any special gains. On the other hand, it says not all use of AI is spam. Value is what counts. What the policies and the studies say about rankings is in our guide on whether Google penalizes AI content.
In practice, the signals Google describes are in our E-E-A-T guide: a real author, own experience, sources and transparency. The structure that makes a text useful and citable is in our guide on how to write a blog post.
The same goes for AI answers. The tactics with the most evidence for being cited are sources, expert quotes and statistics, as the guide to what GEO is shows. An AI detector measures none of these signals.
How do you answer a client who doubts the text?
Show the process, not another detector’s score. Explain that Google does not say it punishes origin, that detectors err on human text and that the score changes from tool to tool. Then show what the text has of its own: sources, data, review and an author.
Imagine that a client ran the article through a detector and saw “80% AI”. A clear answer has four parts:
- What Google says: it looks for spam and quality, whoever or whatever writes the text.
- What the detector measures: statistical writing patterns, with errors documented in academic studies.
- What the text has: sourced facts, data or own experience, human review and a real author.
- What not to do: humanize AI text only to lower the score.
If the text has none of this, the client’s doubt has grounds, and the detector is not the problem. In that case, the way forward is to rewrite with sources and experience, as our guide to content marketing in the AI era shows. The step-by-step of review, fact-checking and brand voice is in our guide on the editorial process with AI.
Frequently asked questions
No. SynthID recognizes only its own watermark, which the tool applies when it generates the text, as in the Gemini app. A text without that mark, from a writer or from another AI tool, is out of its reach. So SynthID cannot say whether AI made any given text on the web.
It can. According to the SynthID Text documentation, detector confidence drops a lot when someone fully rewrites or translates the text. In academic studies, automatic paraphrases cut the accuracy of well-known detectors. But fooling the detector does not change what Google assesses: value, originality and trust.
Not as a verdict. An AI detector errs on human text and on rewritten text, according to the studies cited here. Even OpenAI said its classifier should not decide alone. Before publishing, the check that counts most is another one: sourced facts, something of your own in the text, a real author and human review.
Main sources
- Google Search Central, Google Search’s guidance about AI-generated content and FAQ (2023).
- Google Search Central, spam policies for Google web search (2026).
- Google DeepMind, SynthID page (2026).
- Google AI for Developers, SynthID Text documentation (2025).
- OpenAI, AI text classifier and note on its withdrawal (2023).
- Liang and others, GPT detectors and their bias against non-native English writers (2023).
- Weber-Wulff and others, test of 14 tools that detect AI-generated text (2023).
- Dugan and others, RAID, a benchmark for machine-generated text detectors (2024).



