Watermarks, provenance, and AI media
SynthID is Google DeepMind's invisible watermark for AI-generated images, video, and audio. This page covers what the official detector verifies, what its results actually prove, and where a pixel-level AI image checker answers the question it cannot.
Last updated:
Watermarks, provenance, and AI media
Upload a photo and get an AI likelihood score in seconds. No signup.
Drop an image here, or click to upload
PNG, JPEG, or WebP. Paste works too.
Upload an image on the left. Results will show AI likelihood, generator matches, and next steps here.
SynthID verifies a signal that was deliberately embedded at creation. That makes it strong where it applies and silent everywhere else — so the useful skill is knowing which check answers the question you actually have.
Which models it covers, which three media types it accepts, and what each of its two result states does and does not establish — taken from Google's own documentation rather than inferred from marketing copy.
The most useful idea on this page: a watermark answers whether a known watermarked tool made the file, while a detector answers whether it looks synthetic. Knowing which question you are asking is most of the work.
The checker above reads statistical and structural cues left by generative models. It needs no watermark, no metadata, and no publication history, so it stays usable on stripped screenshots and non-partner output.
This site does not decode SynthID watermarks — only Google's detector can do that, and there is no public API for it. The value here is an honest division of labour: the watermark check when it applies, the pixel check when it does not.
A SynthID detector reads an invisible watermark that Google DeepMind embeds directly into AI-generated content. Unlike a file's metadata block or a visible logo, the watermark is woven into the content itself — into the pixels of an image or video, and into the frequency components of audio. Google's own description is that this makes the watermark travel with the content wherever it goes, through much of the resaving and resharing that destroys every other provenance signal.
The public detector is hosted by Google at synthid.com. You upload an image, video, or audio file and it reports whether a SynthID watermark is present. It is free to use, and Google's terms for the portal prohibit automating it or treating the frontend as a batch verification API.
This is the part most people get wrong: SynthID is not a general AI detector. Google states plainly that it can only detect media from the companies that have adopted SynthID — Google itself plus NVIDIA, OpenAI, and Kakao, with Apple announced as a future partner. An image from a generator outside that list returns no watermark however obviously synthetic it is.
The mechanism explains the reliability difference. Because SynthID is an intentional signal inserted at generation time, it does not have to infer anything from ambiguous visual evidence. A general AI detector, including the one on this page, reads passive artifacts — texture regularity, noise behaviour, edge consistency — and estimates a probability. The watermark check is closer to reading a signature; the pixel check is closer to a judgement call, and Google itself frames general detectors as broader in scope but sometimes less reliable.
That is why the two belong together rather than in competition. A watermark check is decisive when the file carries one. A pixel check still works when the file came from a generator that does not watermark, when heavy editing degraded the signal, or when the file's origin is simply unknown.
Google's own guidance is explicit: upload the best version available. Watermarks degrade under heavy compression and repeated editing, so a full-resolution original gives the detector the most to work with. A screenshot pasted through a chat app is the worst case.
The portal reports either that a watermark was detected — meaning the file was likely created or edited by a SynthID-enabled model — or that no confident signal was found. There is no percentage to weigh. The second state means the check could not confirm, not that the image is authentic.
Google lists the reasons a watermarked image can return no match: it came from a model without SynthID, or heavy modification degraded the signal. Google also notes the detector can very rarely flag content that is not watermarked. Absence of a watermark is absence of evidence, never evidence of absence.
If the file carries no watermark — because it was stripped, or because it came from a non-partner generator — a visual-signal check is the remaining automated option. Upload the same file above for an AI likelihood score and signal notes, then weigh both results together.
EXIF blocks and Content Credentials are routinely stripped by screenshots, chat apps, and social platforms. SynthID is embedded in the pixels, so it persists through many of the ordinary operations that destroy every other provenance signal.
A watermark match tells you a SynthID-enabled model touched the file. It does not tell you intent, authorship, or whether the surrounding claim is true. A generator score tells you how synthetic the pixels look. Neither replaces the other.
SynthID only marks output from Google, NVIDIA, OpenAI, and Kakao models. A great deal of the internet's AI imagery predates those adoptions or sits outside those pipelines, which is why a watermark check alone leaves the majority of cases unresolved.
A watermark match is close to conclusive; a watermark miss still leaves you needing a visual read. Running both takes under a minute and covers the provenance question and the pixel question at once.
Google draws the distinction itself. SynthID verifies media by detecting an intentional signal embedded during content creation, which gives it high reliability and resilience but limits it to identifying content from SynthID-integrated models. General AI detectors analyse passive artifacts and patterns to estimate the likelihood of AI generation when no embedded signal is present: broader in scope, and in Google's words, sometimes less reliable.
That trade-off is structural rather than a matter of engineering quality. A watermark reader is looking for something deliberately placed there, so a match is close to binary. A detector is inferring intent from evidence the generator left behind by accident, and every act of compression, cropping, filtering, or re-encoding blurs that evidence further.
It follows that the two tools fail in opposite directions. SynthID produces confident answers on a narrow slice of the world. General detectors produce qualified answers on almost everything. A verification workflow that understands both failure modes is far more useful than one that trusts either number on its own.
Ordinary provenance signals live in the file wrapper. EXIF camera records, editing history, and C2PA Content Credentials are all attached to the container rather than to the picture, so a screenshot — which renders pixels and writes a new file — discards them entirely. That is why check-the-metadata has been unreliable advice for years.
SynthID works differently. It is embedded into the pixels of an image and into the frequency components of audio, which is why Google describes it as travelling with the content rather than with the file. The transformations that most commonly destroy metadata — screenshots, re-saving, format conversion, mild cropping — are the ones SynthID was designed to survive.
This does not make it permanent. Google is direct that no signal is completely robust, and that someone running many transformations may find one that escapes detection. The practical reading is that SynthID raises the cost of laundering provenance rather than eliminating it, which is a meaningful improvement over metadata that vanishes on the first screenshot.
The most consequential mistake with watermark detection is treating not detected as not AI. Google lists two ordinary explanations for a miss — the content came from a model without SynthID, or heavy modification degraded the watermark — and neither says anything about whether the image is synthetic.
The coverage gap is the larger of the two. SynthID marks output from Google, NVIDIA, OpenAI, and Kakao models. Every other generator, every older model released before those adoptions, and every locally run open-weights pipeline produces images with no watermark to find. In those cases the detector is not failing; it is being asked a question it was never built to answer.
Google also notes the inverse risk, that the detector can in very rare cases flag content that was not watermarked. Combined with the ordinary degradation cases, the discipline becomes clear: a positive result is strong evidence, a negative result is neutral, and neither replaces looking at the image and its context.
Once the watermark check has run, the remaining question is almost always about appearance: does this file carry the statistical fingerprints of generative output? That is what a visual-signal detector measures, and it is the only automated read available when no embedded signal exists.
The checker on this page works on JPG, PNG, and WebP files up to 10 MB and returns an AI likelihood score with notes on which families of cues contributed. It does not need the file to be published anywhere, so it also covers the case where reverse image search returns nothing because the image has never appeared online before.
Treat its output as a triage signal rather than a verdict, exactly as you would treat a watermark miss. Above roughly 70% is grounds to verify the source before publishing or paying. The 40-70% band is genuinely inconclusive. Below 40% means no strong synthetic cues were found, which is not a certificate of authenticity.
The full sequence — watermark check, pixel check, reverse search, source verification, and human review for anything consequential — is laid out in our guide to detecting AI-generated images. No single step in that chain is decisive on its own, and the value comes from the chain rather than from any one link.
Common questions about using an AI image checker for photos, art, and synthetic media.
It is Google's tool for reading the SynthID watermark — an invisible signal Google DeepMind embeds into AI-generated images, video, and audio at the moment of creation. You upload a file and it reports whether that watermark is present. It is hosted by Google at synthid.com and is separate from general AI detection tools, because it verifies an embedded signal rather than judging how synthetic an image looks.
No, and no third-party tool can. Decoding a SynthID watermark requires Google's own detector, and Google publishes no API for it. What this site offers is the complementary check: a pixel-level AI likelihood score that works on files SynthID cannot cover, including non-partner generators, heavily edited images, and screenshots with no watermark left to read.
No. SynthID only marks output from models whose makers adopted it, including Google, NVIDIA, OpenAI, and Kakao. Images from any other generator carry no SynthID watermark regardless of how clearly AI-generated they are. Google states this limitation directly on its own detector page.
It means the detector found no confident SynthID signal, not that the image is real. Google names the two common causes: the file came from a model without SynthID, or heavy editing and compression degraded the watermark beyond detection. Treat a negative result as inconclusive and pair it with a visual check and source verification.
Google DeepMind pioneered SynthID and applies it across its own generative products including Gemini and Imagen. NVIDIA, OpenAI, and Kakao have adopted it, and Google has announced Apple as a forthcoming partner. The list grows over time, which is why it is worth checking the official page rather than relying on a description written months ago.
Much less than you would expect, and far less than it removes metadata. Because the watermark lives in the pixels rather than in a file header, it survives many screenshots, crops, colour shifts, and lossy re-saves. Google describes SynthID as robust against typical transformations while acknowledging that no watermark is infallible under determined repeated processing.
No. They are complementary provenance systems. Content Credentials are signed metadata attached to a file, which means they can be stripped and their absence proves nothing. SynthID is embedded in the content itself, which makes it harder to remove but limits it to models that support it. Google's own position is that these should work as layers rather than as one universal standard.
Three: images, video, and audio. For video and audio the detector can identify which segments carry the watermark rather than returning a single verdict for the whole file. SynthID text watermarking exists separately in Google's developer tooling, but the public detector portal covers the three media types above.
Yes, it costs nothing to use. Google does gate the portal behind an additional terms-of-service screen, and those terms prohibit automated submission, scraping, and quota evasion, so it is built for individual manual checks rather than bulk verification. That is worth knowing before you plan a review pipeline around it.
Run the watermark check first when you suspect the file came from Google, NVIDIA, OpenAI, or Kakao output, and run it on the highest-quality version you can obtain. A positive result is the strongest single signal available, because it confirms an embedded mark rather than an inference.
When the watermark check returns nothing, the case is not closed. It is simply a different question. Upload the same file to the checker above for a pixel-level read. The score reflects statistical and structural cues left by generators, so it applies to files that never carried a watermark in the first place.
Keep the two results in separate columns of your thinking. Provenance tells you where a file probably came from. Appearance tells you how synthetic it looks. A verification decision that uses only one of them is guessing with extra confidence, and Google's own documentation recommends gathering multiple points of data before deciding.
For anything consequential, add the steps that no automated check covers: reverse-search the image for prior publication, ask the sender for the original camera file, and route high-stakes cases to a human reviewer before acting.
