On October 7, 2026, watermark detection stopped being a specialist tool and became a consumer feature. Google opened its SynthID Detector to everyone, globally, at synthid.com — the portal that had been gated to journalists, media professionals, and researchers since it was announced at I/O in May 2025. Upload an image, a video, or an audio file and it tells you whether Google's invisible SynthID watermark is present, now reading not just Google's own models but partner output from OpenAI, NVIDIA, and Kakao, with Apple flagged as coming soon. That single change reframes the whole subject for creators. For most of the AI era, "can anyone tell my content is AI?" was a question that required a researcher's access or a pile of circumstantial visual tells. Now a viewer, an editor, or a platform can get a provenance answer in one upload — and Google says it has already watermarked more than 180 billion images and videos and 240,000 years of audio, with verification features in Search, the Gemini app, and Chrome handling over a million checks a day. This guide is about the detection side specifically — not the taxonomy of watermark types, which is covered separately — because the detector going public is what turns an abstract provenance debate into a concrete one every creator now has to reason about. It covers what SynthID Detector genuinely checks, the four things it pointedly cannot do (Google itself warns it "is not a general AI detector"), how detection works differently across image, video, and audio, why text remains the blind spot, the parallel C2PA Content Credentials verification path that OpenAI and others also read, and the reliability gap that keeps getting misread in both directions. The last third is the part that matters most: once checking provenance is a one-click public feature, the smart creator stops asking whether their AI media is detectable — it increasingly is — and starts treating honest disclosure at publish time as the only durable move.
On October 7, 2026, Google DeepMind opened its SynthID Detector to the public — available globally, in English, at synthid.com. The portal is not new; it was announced at Google I/O in May 2025 and spent the intervening year gated to a narrow group of journalists, media professionals, and researchers behind a waitlist. What changed is access. Checking whether a piece of media carries Google's invisible AI watermark went from a credential you had to apply for to a consumer feature anyone can use in one upload.
That access shift is the entire reason this is worth a guide of its own. The mechanics of watermarking — the four different things that wear the word, how each behaves across media, what an edit does to them — are a separate subject, mapped in full in the companion guide on AI-generated content watermarks. This guide is about the other side of the glass: detection. For most of the AI era, "can anyone actually tell my content is AI?" was a hypothetical, because reading the provenance signal required either researcher access or a trained eye hunting for visual tells. Now a viewer, an editor, a brand's legal team, or a platform can get a real answer by dragging a file into a web page. The question stopped being hypothetical, and that reframes what a creator should be doing about it.
The scale behind the tool is the part that makes it more than a novelty. Google says it has watermarked more than 180 billion images and videos and 240,000 years of audio since SynthID launched in 2023, and that its verification features already live in Search, the Gemini app, and Chrome, together handling over a million provenance checks a day. Detection is not a lab demo being wheeled out for a press cycle; it is infrastructure that is already woven into the surfaces where people view and share media. The original announcement is written up in the news piece on SynthID Detector; what follows is the practitioner's read on it.
SynthID Detector reads one specific thing: whether a file carries Google DeepMind's SynthID watermark, an invisible signal the generating model embeds directly in the content. You upload an image (JPG, PNG, WEBP, HEIC, and similar), a video (MP4, MOV, WEBM), or an audio file (WAV, MP3, OGG, FLAC, AAC, M4A), and the tool reports whether a SynthID mark is present and, where possible, highlights the portions of the media most likely to be watermarked. Because the mark lives in the pixels, frames, or waveform rather than in an attached label, it survives the ordinary handling — cropping, screenshotting, re-encoding, compression — that strips a visible badge instantly.
The genuinely new capability in the October 2026 release is cross-model coverage. SynthID Detector no longer reads only Google's own output. Google has brought partners into the scheme, so the tool now also detects SynthID marks on content from OpenAI, NVIDIA, and Kakao, with Apple named as coming soon. That matters because a single-vendor detector is close to useless in a world where creators mix tools — a verifier that only recognized Gemini output would miss most of what crosses a feed. Cross-model detection is the difference between a branded gimmick and something a platform could plausibly lean on.
Google is unusually candid about the tool's limits, and reading those limits correctly is more important than celebrating the launch. The headline caveat, from Google's own FAQ, is blunt: "This is not a general AI detector." Four specific blind spots follow from that.
SynthID Detector can identify media from companies participating in the scheme — Google and its named partners — and nothing else. Content from any generator that does not embed SynthID will return clean no matter how synthetic it is. So a negative result is not evidence of a human origin; it is evidence that this particular watermark, from this particular set of tools, is absent. Treating "no SynthID found" as "this is real" is the single most dangerous misreading of the tool.
SynthID is durable against mild handling, but it is not indestructible. Aggressive editing, heavy compression, or layered transformations can degrade the watermark below the threshold where the detector can read it. A file can be genuinely AI-generated, genuinely SynthID-marked at creation, and still come back as "not detected" after enough processing. Detection is probabilistic at the margins, not a binary guarantee.
A detector reads how a file was made, not whether what it depicts is real. A positive SynthID result tells you a participating AI tool was involved; it says nothing about whether a photograph documents a true event, and a negative result on a real photo says nothing about whether the scene was staged. Provenance and veracity are different questions, and the tool only answers the first.
This is the conflation that trips up almost everyone. Watermark detection reads a signal the model deliberately planted and can confirm — close to decisive when it fires. A conventional AI detector, the kind that scans an essay or a blog post, guesses from statistical style cues with no planted signal to read, and is notorious for false-positiving on human writing. SynthID Detector is the first kind, not the second. The limits of the second kind are a whole topic on their own, covered in AI content detection.
Detection is only ever as good as the watermark it is reading, and watermarks are embedded differently per medium. Images, video, and audio share a useful property: a generator can perturb pixels, frames, or the waveform below the threshold of human perception and leave a robust signal a detector can recover. That is why SynthID Detector supports those three media and why their marks survive normal editing — there is perceptual headroom to hide a durable signal in. In practice, image detection is the most mature, video detection reads a pattern spread across frames, and audio detection reads information tucked into frequency ranges of the waveform.
Text is the conspicuous absence from the detector's upload options, and that is not an oversight — it is the hard case. Prose has no imperceptible layer to hide a signal in; the only durable watermark is a statistical bias in the model's word choices, a technique covered in AI text watermarking. Even where that watermark exists — Google's SynthID-Text was published in Nature in 2024 and open-sourced, and Anthropic began watermarking Claude's text in 2026 — reading it is a separate, key-based statistical test, not a drag-and-drop portal, and most text models embed nothing at all. The upshot: an AI image is far more likely to be detectable than an AI paragraph, and anyone assuming "AI content" is uniformly checkable has it backward.
SynthID is not the only provenance signal a creator's media might carry, and detection increasingly means checking more than one. C2PA Content Credentials are a signed cryptographic record of how a file was made — which tool, which edits — attached to the file's metadata rather than hidden in the content. You verify them with a Content Credentials inspector, which reports the issuer, the generating tool, and the recorded actions. The strength is detail; the weakness is fragility, because the metadata wrapper is lost the moment a tool strips it — a screenshot, a careless re-export, a CMS that discards it on upload. The definitional treatment is in the C2PA glossary entry.
The two signals are complementary, and the better verification tools now read both. OpenAI has been previewing a public verification tool that checks whether an image came from its models by looking for both Content Credentials and a SynthID-style watermark, and Google's own verification surfaces in Search and Chrome read C2PA alongside SynthID. The practical lesson for a creator is that provenance is layered: your published image may carry an invisible in-content watermark, a metadata credential, or both, and a determined checker can read whichever survived. A clean visible file is not a clean provenance file.
The public launch has produced two opposite errors, and both are worth naming. The first is over-trust: reading a "watermark not detected" result as proof that content is human-made, or that your own AI content is safely untraceable. It proves neither. The detector only sees participating models, only reads marks that survived editing, and says nothing about non-SynthID tools. A negative is an absence of one specific signal, not a verdict.
The second error is over-fear: assuming every piece of AI-assisted work is now instantly exposed and penalized. Also false. Detection is uneven — strong for image and video from major tools, weak-to-nonexistent for text, defeatable by heavy editing — and detection is not the same as punishment. Being identified as AI-made is only a problem when it collides with an obligation you ignored: a platform label you skipped, a disclosure the law required, a likeness you used without consent. The signal is neutral; the consequences come from what you did or didn't disclose around it. The sane posture sits between the two errors: assume your image and video output is checkable, don't assume your text is, and never confuse a clean result with either safety or honesty.
The strategic shift is small to state and large in consequence. For years the operative question was "can anyone tell this is AI?" The public detector has largely retired that question for visual media — increasingly, yes, anyone can. So the energy a creator might have spent hiding provenance is now wasted energy, and worse, it is the energy most likely to blow up: scrubbing a watermark to pass AI work off as human is both harder than it looks and exactly the deceptive act platforms in 2026 are built to catch. The durable move is the opposite of concealment. Keep the provenance, disclose where it's required, and make content good enough that the label is a non-issue.
Concretely, that means three habits. First, label at publish, per platform — YouTube, Meta, and TikTok each have their own AI-disclosure rules and their own detection, and the EU AI Act's Article 50 transparency obligations, in effect since August 2, 2026, land on you as the publisher of synthetic media, not only on the model maker. Second, treat a real, identifiable person in generated media as a separate and stricter obligation — likeness detection and consent rules apply whether or not a watermark is present. Third, make disclosure a step in the act of publishing rather than an afterthought, so the decision is made by a human who can see each platform's requirement, on every asset, before it ships. The reliable failure mode is not being detected; it is being caught having said nothing.
Kompozy is an AI content generation and multi-platform publishing engine, and the public detector changes its role in one specific way: it removes the fantasy that you can quietly ship AI media and not be checked. Once provenance is a one-click feature your audience, an editor, or a platform can run on any post, the only sustainable workflow is one that bakes honest disclosure into publishing — which is precisely the seam Kompozy sits on. From a single source it produces across 18 formats — persona and avatar video, clips, carousels, images, blogs, newsletters, text — and then publishes to each destination individually rather than blasting one identical post everywhere. Because every destination is handled on its own, the caption and any AI-disclosure label are set per platform, and the per-post review step under Autopilot is the single human checkpoint where a person confirms that anything synthetic carries the label a given platform expects before it goes live.
Be clear on the boundary, because it is easy to wish for more than any publishing tool does: Kompozy is not a detector and not a provenance service. It will not read a SynthID mark, inspect a C2PA credential, or embed a watermark for you — those stay with the generating models and the verification portals like SynthID Detector. What it does is operationalize the response to a world where detection is public and trending toward universal: it assumes your media is checkable, puts the disclosure decision in one governed place at publish time, and carries the mark a brand actually wants — its own, through a Persona Brief that owns the voice and brand-exact HyperFrames that stamp identity into each asset. That is the constructive inverse of scrubbing a generator's watermark. For a solo creator, Starter ($199/mo, 5,500 credits) covers it; a brand publishing across every channel fits Pro ($499/mo, 18,000 credits); agencies running it for clients use custom Enterprise.
Watermark detection went public on October 7, 2026, and that is the fact that reorganizes the whole subject. SynthID Detector reads an invisible, durable, now cross-model watermark in images, video, and audio — but it is not a general AI detector, a clean result is not proof of a human origin, heavy editing can erase the signal, and text is largely outside its reach. C2PA Content Credentials are a second, metadata-based signal that a growing set of verifiers read alongside it. For a creator, the honest conclusion is to stop treating detectability as a question worth beating. Assume your visual AI output can be checked, disclose where a platform or the law asks, build that disclosure into the moment of publishing, and spend the saved effort making content you would gladly put your name on.
It is the act of checking whether a file carries a watermark that identifies it as AI-generated — reading a signal the generating model deliberately planted, rather than guessing from style. The most prominent tool is Google's SynthID Detector, which reads Google DeepMind's invisible SynthID mark in images, video, and audio; C2PA Content Credentials verification is a parallel path that reads a cryptographic provenance record in a file's metadata. Both answer "was this made with a participating AI tool?" not "does this look AI-written?"
Yes. Google opened SynthID Detector to everyone globally, in English, on October 7, 2026, at synthid.com. It had been limited to journalists, media professionals, and researchers through a waitlist since it was announced at Google I/O in May 2025. You upload an image, video, or audio file and it reports whether a SynthID watermark is present, now covering Google's models plus partners OpenAI, NVIDIA, and Kakao, with Apple support said to be coming.
No, and Google is explicit about this — its FAQ states the tool "is not a general AI detector." It can only identify media from companies that have adopted SynthID, so content from a model that doesn't use SynthID will come back clean even if it is fully AI-generated. A "watermark not detected" result means no SynthID signal was found, not that the content is human-made — heavy editing can also weaken the watermark below the detection threshold.
A watermark detector reads a signal the model intentionally embedded and can confirm with a key, so a positive result is close to decisive for content that model marked. A conventional AI detector — the kind schools and platforms use on essays — estimates whether content is AI-made from statistical style cues, with no planted signal to read, so it is probabilistic and can false-positive on human work. Watermark detection is verification; style-based detection is inference.
Assume your AI-generated media is detectable, because it increasingly is — the question "can anyone tell?" has largely been answered, and a viewer or platform can now check in one upload. The durable move is to stop trying to scrub or outrun provenance and instead disclose honestly where a platform or the law requires it, and compete on quality. Practically, that means building the disclosure decision into your publishing step, per platform, rather than remembering it after the fact.
AI-generated media watermark detection is checking whether a file carries a watermark that marks it as AI-made. As of October 7, 2026, Google's SynthID Detector is public at synthid.com, reading invisible SynthID marks in images, video, and audio from Google, OpenAI, NVIDIA, and Kakao. Crucially, it is not a general AI detector — a clean result means no SynthID signal was found, not that the content is human-made. C2PA Content Credentials verification is a parallel path that reads provenance from file metadata.
Get started → · ← All guides · Compare Kompozy vs other tools