Google opened its SynthID Detector to everyone on 7 October 2026. Until that day the tool sat behind a waitlist for journalists, media professionals and researchers. Anyone can now upload a file and ask whether it carries Google’s AI watermark.
The detector accepts images, video and audio. It does not accept text. Google watermarks text as well — its own SynthID page says so plainly — but the one modality Google’s content policy guidance actually tells publishers to review before publishing is the modality the new public tool will not take.
That asymmetry is the part that matters if you run a website. Everything below comes from documents read live on 8 October 2026, each with its own date stamp attached.
What Google said on 7 October 2026
Pushmeet Kohli, Vice President of Science and Strategic Initiatives at Google DeepMind, announced the expansion in a post on blog.google dated 7 October 2026. The post says the company is “expanding access to everyone – with the tool available globally in English starting today.”
It describes the capability in one sentence: “anyone can easily check if an image, video, or audio file was made with AI from Google or our partners.” The partners are named — “including OpenAI, NVIDIA, Kakao, and soon, Apple.”
Two scale figures appear. Google says it has “watermarked over 180 billion images and videos, along with 240,000 years of audio content”, and that built-in verification already runs inside “Search, the Gemini app, and Chrome, which now regularly handle over 1 million requests daily.” Neither is an accuracy measurement: they count watermarked items and requests served. Google published no accuracy rate, no false-positive rate and no sample size with the announcement.
The post states no limitation of any kind. Asked explicitly, on two separate reads, whether it names anything the tool cannot detect, the answer both times was no. The post does not mention text.
Google’s own SynthID page is where text appears
The gap closes and reopens on Google DeepMind’s SynthID page, read on 8 October 2026. That page says SynthID “embeds digital watermarks directly into AI-generated images, audio, text or video”, and that “We’ve expanded SynthID to watermarking and identifying text generated by the Gemini app and web experience.”
The same page describes the detector as checking “an image, video, or audio file.” Text is absent from that description. Asked as a yes/no question on a second read, the page does not say the detector can check text, and the word does not appear there at all.
The page also states what the watermark survives, and the statements are modality-specific. For images and video the watermark is designed to “stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression.” For audio it “can’t be altered by common modifications like adding noise, MP3 compression, or changing the speed of the track.” For text the page says the watermark “doesn’t affect the quality of the output” and gives no robustness statement at all — nothing about paraphrasing, rewriting or editing.
What the tool will not tell you
The detector’s own site carries a disclaimer sharper than anything in the announcement. The page at synthid.com could not be read by automated fetch during this run — it returned metadata and no body text — so the wording here comes from two independent reports of 7 October 2026, which do not render it identically.
Search Engine Journal renders it as: “This is not a general AI detector. We can only detect media from the above companies, who have adopted SynthID technology.” MacRumors, in Juli Clover’s report of the same day, renders the second clause differently: the site “is not a general AI detector” and “can only detect media that uses the SynthID technology.” The two agree on the first sentence and diverge on the second, which is worth stating rather than smoothing over.
Both reports agree on the practical shape. The partner list is closed, so a file from any model outside it returns nothing. Search Engine Journal also reports the result wording — “SynthID was detected” and “This media was made or edited with Google AI” — which does not separate a file generated from scratch from one lightly retouched, plus an FAQ line that “no signal is completely robust.”
What Google’s documentation for site owners says
Four Google documents govern what a publisher is supposed to do about AI-made material. All four were read live on 8 October 2026, and every absence below was confirmed on a second read asking explicitly, string by string, for a yes or no.
| Google document | Last updated | What it contains on AI provenance |
|---|---|---|
| Guidance on using generative AI content | 2026-10-01 UTC | Requires manual review before publishing. Contains no “SynthID”, no “watermark”, no “provenance”, no “C2PA”, no “disclose”. |
| Image SEO documentation | 2026-03-02 UTC | Contains none of “SynthID”, “C2PA”, “Content Credentials”, “IPTC”, “watermark”, “AI-generated” or “provenance”. |
| Search Central documentation changelog | read 8 Oct 2026 | Exactly one October 2026 entry, dated 1 October, about the generative AI guide. Nothing about the detector. |
| blog.google announcement | 7 Oct 2026 | Launches the tool. States no limitation and does not mention text. |
The generative AI guide, stamped 2026-10-01 UTC, is the one that asks publishers for work. It says: “It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing.” It extends that review beyond the body copy — “This review also applies to metadata like <title> elements, meta description elements, structured data,” and “alternate texts for images, which can appear in Search results.”
The same guide contains exactly one labelling requirement, and it is not a Search requirement. Under its ecommerce paragraph it states that “AI-generated product data such as title and description attributes must be specified separately and labeled as AI-generated.” That obligation is placed under Google Merchant Center’s policies. Asked directly, the page does not extend it to Google web Search.
So the documented position on 8 October 2026 is: Google requires AI labelling in Merchant Center, requires manual review everywhere, documents no provenance signal for web Search at all, and ships a public detector that cannot read the format most of that guidance is about.
What has actually been measured about text watermarks
Google has published no robustness figure for its text watermark. OpenAI published one two days earlier for a different system, and the numbers show why text provenance is harder than image provenance.
OpenAI’s EU text provenance post, dated 5 October 2026, reports detection at a 1% false-positive target. On psychology content, detection was about 80% for 200-token passages and about 95% for 400-token passages. On mathematics it was “substantially lower”, with no figure given. On 400-token passages, replacing 10% of words with synonyms cut detection from roughly 92% to 66%; replacing 25% cut it to about 17%. Asked directly, the post publishes no sample sizes and no confidence intervals.
The same post contains the sentence that bears hardest on everything above: “The absence of a detected watermark does not prove human authorship.” It also says that text from OpenAI tools “may be too short, edited, or translated for detection to work reliably”, and that detector access is “initially be limited to approved researchers and expert organizations” rather than public.
One sourcing note. PPC Land’s report of 6 October 2026 renders a stronger line — that rewriting or translating text “can completely remove the watermark” — which does not appear in OpenAI’s post in that form. The primary is the hedged sentence quoted above. Where a secondary report is firmer than the document it describes, the document wins: the same discipline the course applies to figures a chatbot invents.
What to do this week
Four things, all checkable, none of them speculative.
- Check your own AI-made images. Upload three images your team generated with Gemini, Imagen or any partner model. You then know what a third party sees when they check your media, before a client or a journalist does.
- Do not buy an AI text detector on the strength of this launch. The tool does not read text. If a vendor cites SynthID while selling text detection, that is a claim to check against Google’s own page, not to take on trust.
- Fix the metadata gap first. The 1 October guidance extends manual review to title elements, meta descriptions, structured data and image alt text — the fields most often generated in bulk and least often reviewed. Pull fifty URLs and read the alt text you shipped.
- If you sell through Merchant Center, read the labelling rule. AI-generated title and description attributes must be specified separately and labelled as AI-generated. That is a documented requirement with a documented surface, unlike anything in web Search.
None of this changes a ranking. It changes what you can verify about your own material, which is the measurement half of AI for SEO.
Where the industry genuinely disagrees
Three questions sit under this story that the available evidence does not settle, and this site does not pretend otherwise.
Does the absence of a watermark tell you anything? One camp treats a clean result as evidence of human authorship, which is how a provenance tool becomes a detection tool in practice. The other says a closed partner list and an untested robustness profile make absence meaningless. The strongest evidence belongs to the second camp and comes from a vendor about its own system: OpenAI says absence “does not prove human authorship.” Google publishes no equivalent statement either way, and nobody has measured how often a watermark survives an ordinary publishing pipeline.
Should a publisher preserve or strip provenance signals in media it publishes? One camp says preserve: the mark is designed to survive cropping and compression anyway, Google already runs verification inside Search, and removing it reads as concealment if it is ever noticed. The other says a signal whose consequence is undocumented should not be shipped deliberately — Google’s image SEO documentation says nothing about provenance, so nobody can tell you what a detected watermark does to image eligibility. No published study separates the two.
Where does scaled content abuse begin? Unchanged by this launch, and still the question most readers want answered. Google’s policy targets purpose rather than production method, and no public threshold defines scale. The 1 October guidance adds a review obligation and no number. The course treats the boundary in Grey Hat, Black Hat and Parasite SEO.
Txema Hermoso’s opinion
I will defend one claim from this story and no more: the modality Google tells publishers to check is the modality its public detector does not accept. That is documentary, it is reproducible from the two Google pages in ten minutes, and it is the part of the week’s coverage that mostly went unwritten.
What I will not say is that this launch makes AI content riskier or safer to publish. It does neither: it changes no policy and adds no threshold. Nor will I tell you a clean detector result means a human wrote something. The only party to publish a figure on that question says the opposite about its own system, and Google has published nothing.
The useful instinct is the one that applies to every provenance claim: ask what the tool was built to prove, not what you would like it to prove. SynthID was built so Google can recognise Google’s own output. Treating it as a verdict on authorship is a category error, and it is the error the next year of vendor marketing will be built on. The same caution applies to confident claims about how AI search picks what to cite.
What is still unknown
Google has published no accuracy rate, no false-positive rate and no sample size for the detector, no robustness figure for the text watermark and no statement about paraphrasing. It has not said whether a detected watermark affects anything in Search, and its image documentation, last touched on 2 March 2026, does not mention the subject. There is no date for text support in the detector, and none for Apple beyond “soon”.
Sources
- Google DeepMind, post by Pushmeet Kohli, Vice President of Science and Strategic Initiatives, blog.google, 7 October 2026.
- Google DeepMind, SynthID, read 8 October 2026. The page publishes no date stamp.
- Google, Google Search’s guidance on using generative AI content on your website, Last updated 2026-10-01 UTC, read 8 October 2026.
- Google, image SEO documentation, Last updated 2026-03-02 UTC, read twice on 8 October 2026.
- Google, Search Central documentation changelog, read 8 October 2026.
- OpenAI, EU text provenance, 5 October 2026. No sample sizes or confidence intervals published.
- Search Engine Journal, “Google Opens Its SynthID Watermark Detector To The Public”, 7 October 2026.
- MacRumors, Juli Clover, 7 October 2026.
- PPC Land, Luis Rijo, 6 October 2026, on OpenAI’s detection figures.
- Search Engine Journal, Matt G. Southern, 5 October 2026, on OpenAI’s EU text watermarking.