SynthID on images: what Google’s check shows and what it can’t prove
Now that anyone can run Google’s SynthID Detector, here is what each result can support, what the AI Act requires, and how a real image is proven.
Updated on
A SynthID check can confirm that an image went through an AI model that applies Google’s invisible AI watermark, but it cannot prove that an image is real. When the check finds nothing, the image may be authentic, it may come from a generator that doesn’t use SynthID, or the signal may have faded after repeated edits: Google’s own Gemini help page says unmarked content “could have been created by other AI systems.” Proof of authenticity has to come from somewhere else. TrueScreen certifies photos and videos at the source, at the moment of capture, with forensic acquisition and an official digital seal and qualified timestamp, recognized internationally.
Key takeaways
- A positive SynthID result means a Google or partner model created or edited the image; it says nothing about who posted it, when, or whether the caption is true (Google Gemini help page).
- A negative result is not proof of authenticity: Google says unmarked content “could have been created by other AI systems” and the watermark may not survive many alterations.
- SynthID Detector recognizes Google and partner content (OpenAI, NVIDIA, Kakao, Apple soon), while the Gemini app only recognizes content from Google AI tools (Google, October 7, 2026).
- AI Act Article 50 makes providers mark generated content from August 2, 2026, with systems already on the market given until December 2, 2026 for marking under the Digital Omnibus on AI.
- Proof that an image is real comes from certification at the source; watermark checks, metadata and AI analysis are documented clues, not conclusions.
62%
correct answers when people tried to tell real and AI-generated images apart
Roca et al., Microsoft AI for Good Lab, arXiv 2507.18640, 2025
Google marks the images its AI models generate with SynthID, an invisible watermark built by Google DeepMind. Since October 7, 2026, anyone can check an image for it in the SynthID Detector portal, on top of the checks built into Gemini, Search and Chrome, which Google says handle more than a million requests a day.
The catch is that the check only works in one direction. When it finds the watermark, it tells you something specific. When it finds nothing, it tells you very little, and a negative result on a suspicious photo can feel like a clean bill of health when it isn’t one.
So what can you actually conclude from a SynthID check, and what do you need when you have to show that an image is real? Proof of authenticity has to be built at the source; for images that already exist, a watermark check is one documented element of a broader analysis, never the verdict.
What is the SynthID AI image watermark and where can you check it?
SynthID is a watermark Google writes into the image itself, and the places where you can look for it don’t all see the same content. Google DeepMind embeds it in images, video, audio and text produced by Google models and by partners that have adopted it. It launched in beta on August 29, 2023 for a small group of Vertex AI customers using Imagen, and by October 2026 Google said it had marked more than 180 billion images and videos and 240,000 years of audio. Text watermarks raise questions of their own, covered in our analysis of what the Claude text watermark proves and what it cannot.
A watermark in the pixels, not in the metadata
SynthID lives in the pixel values of the image, where the eye can’t see it. Two models trained together do the work: one embeds the watermark at generation, the other looks for it during a check. Because the mark sits in the picture and not in a metadata field, stripping the metadata doesn’t remove it.
Google’s SynthID product page says the image watermark is “designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression,” and Gowal and colleagues describe the system in the SynthID-Image technical paper (arXiv 2510.09263, October 2025). Google is candid about the ceiling, though: its 2023 launch post said “SynthID isn’t foolproof against extreme image manipulations.”
SynthID Detector, Gemini, Search and Chrome: where to run a SynthID check
Each place you can run a SynthID check recognizes a different slice of content.
- SynthID Detector: announced at Google I/O on May 20, 2025 and first offered through a waitlist, the portal highlights the parts of the content most likely to carry the watermark. Since October 7, 2026 it has been open to everyone, worldwide, in English only, and it recognizes content from Google and from partners OpenAI, NVIDIA and Kakao, with Apple coming soon.
- The Gemini app: since November 20, 2025 you can upload an image and ask whether it was made with AI. Per the help page, “Gemini can currently only recognize content created by Google AI tools,” one file at a time, up to 100 MB.
- Search and Chrome: both have the check built in.
A picture from a partner’s model may therefore register in the portal and come back unmarked in Gemini, so always note which SynthID checker you used.
What a positive SynthID result tells you
A positive result answers one question about origin and leaves every question about context open. A positive result from SynthID Detector tells you that the image passed through a Google or partner model that marks its output with SynthID. It doesn’t tell you who published it, when, or whether the caption is true. Google doesn’t publish an error rate for the check, and its help page warns that “in some cases it isn’t clear if the image or video was created or edited by Google AI or not.” TrueScreen answers the other question, how an image was captured, by certifying it at the source.
What it tells you: the image went through a SynthID model
A positive result means a model that applies SynthID generated the image or edited it. Google’s wording is “created or edited,” and the difference matters: a positive doesn’t mean the whole scene is synthetic, so look at which areas the portal highlights before you judge the entire picture.
It’s a strong clue, but not beyond challenge. Saberi and colleagues (arXiv 2310.00076) found an unavoidable trade-off between missed detections and false alarms, and showed that watermarks can be spoofed so that real images appear marked as AI. Any AI watermark detector faces that trade-off, so a positive result belongs in the file as evidence to weigh.
What it doesn’t tell you: who posted it, when, and whether the caption is true
The watermark says nothing about the account that shared the image, the date it first appeared or the story attached to it, which in a dispute are usually the questions that matter. A real photo of a flood from years ago, reposted as yesterday’s news, carries no watermark at all: the check comes back negative and the caption is still false.
Why a missing SynthID watermark proves nothing
“No watermark found” describes the check, not the image. A negative SynthID result doesn’t prove that an image is authentic: it only means the check didn’t find the watermark. The image may be real, it may come from a generator that doesn’t use SynthID, or the signal may have weakened after edits. Google’s Gemini help page says that without a watermark the content “could have been created by other AI systems,” and that “there’s still a chance that after many alterations the watermark won’t be detected.”
| SynthID check result | What you can conclude | What you can’t conclude |
|---|---|---|
| Watermark found | The image was generated or edited with a Google or partner model that uses SynthID | Who published it, when, or whether the context and caption are true |
| Watermark not found | The check didn’t find the signal | That the image is authentic: it may come from a generator that doesn’t use SynthID, or the signal may have weakened after many edits |
| Inconclusive | The signal can’t be read with confidence, a case Google anticipates (not enough detail, or an edit too small to assess) | Anything, in either direction |
| Check not possible (for example, a file over Gemini’s 100 MB limit) | Nothing, because no check took place | Anything |
TrueScreen treats a missing marking as missing information, never as evidence of tampering or of authenticity.
Images from generators that don’t use SynthID
SynthID only sees content from Google and its partners. An image from any other generator comes back negative every time, however synthetic it is.
Your own eyes won’t close the gap. A Microsoft AI for Good Lab study (Roca et al., arXiv 2507.18640) of about 287,000 judgments from more than 12,500 people found that participants told real and generated images apart 62% of the time, barely better than chance. For the other signals and what each one proves, see our guide on how to tell if an image is AI generated.
Edits that weaken the signal
Google says the watermark may not survive “many alterations,” and academic research goes further: Saberi and colleagues found that light watermarks can be erased with minimal changes. University of Waterloo researchers, who presented their UnMarker work at the IEEE Symposium on Security and Privacy 2025, put it bluntly: “Watermarking is being promoted as this perfect solution, but we’ve shown that this technology is breakable.”
How often that happens to SynthID is disputed: when IEEE Spectrum covered the research in August 2025, Google DeepMind contested the rate attributed to SynthID and said it had measured one “significantly lower.” Whatever the true figure, an invisible watermark can be degraded, so its absence can’t carry the weight of proof.
The false negative in a dispute
Picture a dispute in which one side offers a negative SynthID result as proof that a photo is genuine. If the image came from a generator outside Google’s program, or went through heavy editing, that result is a false negative propping up exactly the wrong conclusion: a watermark check can add evidence to a case, but it can’t clear an image.
A careful report doesn’t say “the image is authentic.” It says “no SynthID watermark detected on [date], using [tool], on the file as received,” which is accurate, repeatable and no stronger than the test.
AI watermarks, declared markings and the EU AI Act: who must mark what
The AI Act makes the companies behind generators mark their output, but a marking only helps you if it survives the trip to your screen. Two kinds of marking are in play: invisible watermarks inside the content and declared labels in the file’s metadata. The law requires machine-readable marking; it doesn’t make an unmarked image trustworthy, which is why labelling AI content is not enough to prove what is authentic.
Article 50: what providers must do, and by when
Article 50 of the AI Act, Regulation (EU) 2024/1689, requires providers of AI systems that generate synthetic audio, images, video or text to mark the output in a machine-readable format, detectable as artificially generated or manipulated, from August 2, 2026, according to the European Commission’s AI Act Service Desk. Under Article 50(4), deployers of an AI system that generates or manipulates a deepfake must also disclose that the content has been artificially generated or manipulated.
Systems already on the market got more time. Under the Digital Omnibus on AI, which entered into force in July 2026, providers of generative systems placed on the market before August 2, 2026 have until December 2, 2026 to comply with Article 50(2). The extension covers marking only. The Commission also published a final, voluntary Code of Practice on marking and labelling AI-generated content on June 10, 2026, and our insight on AI Act Article 50 and the labelling of synthetic content covers the obligations in detail. These duties fall on providers and deployers of AI systems; TrueScreen helps organizations document the provenance of their content.
Declared markings in metadata: IPTC, EXIF and XMP
Not every AI marking is a watermark. The IPTC Digital Source Type vocabulary lets a file declare its origin: `trainedAlgorithmicMedia` means media created by a trained AI model, and `compositeWithTrainedAlgorithmicMedia` means content retouched or extended with a generative model, as in inpainting or outpainting. Generators can also write their name into EXIF or XMP fields, the same metadata that can help establish the date of a photo.
Markings like these are declared, not embedded, so they vanish when the metadata is removed, and a file without them tells you nothing either way. For an image found on a web page, Media Inspector in the TrueScreen Forensic Browser reads what the file declares about its origin, such as the IPTC digital source type and generator tags in EXIF and XMP metadata.
Proving an image is real: what you need when you have to prove something
Proof that an image is real comes from how it was captured, not from what a detector fails to find. Detection tools, SynthID included, can only add clues about content whose history is unknown, which is the heart of the choice between deepfake detection and source certification. When you’re the one who has to prove something, document that history from the first moment and treat every later check as supporting material.
The agency that has to show its event photos are real
An agency covers a conference, and the client later asks whether some shots were generated. A SynthID check can’t settle it, because genuine photos and images from generators outside the program come back unmarked alike. The client needs a record of how each photo was taken, the same logic behind a newsroom verification workflow.
If the photographer shoots with the TrueScreen app, each photo captured in real time is certified as a forensic acquisition. Photos imported from the phone’s gallery are sealed with a certain date from the moment of import, but they aren’t certified as a forensic acquisition. One honest limit applies throughout: a certification proves how content was captured, not what stood in front of the camera, so a staged scene is certified as an authentic recording of a staged scene.
The expert witness facing a negative SynthID result
An insurer sends an expert a photo of property damage, and the SynthID check comes back negative. The expert has to explain why that doesn’t close the question, starting with Google’s own caveat that unmarked content may come from other AI systems. A file of unknown origin produces an opinion; a file sealed at capture produces a result you can defend.
The expert evaluates and interprets, and TrueScreen doesn’t replace that judgment. What changes is the material on the desk: a photo certified at the source arrives with its fingerprint, a fixed date and time and a seal that anyone can verify.
-
Certify what you produce at the source
If you create the images, certify them as you take them. Forensic acquisition at the source fixes the file, its date and its context before anyone can question them, and no later check can recreate that record.
-
Capture images you find online together with the page, before it changes
For an image published by a counterparty or circulating online, capture it with a forensic method together with the page where it appears. Posts get edited and deleted, and a certified capture keeps the context a downloaded file loses.
-
Record SynthID, markings and AI analysis results as clues, not conclusions
Run the SynthID check, read the declared markings and, if needed, run an AI analysis. Record each result with the tool, the date and the exact file tested, and write them up as clues that add to the evidence, never as conclusions about authenticity.
How to prove an image is authentic with TrueScreen
TrueScreen, the Data Authenticity Platform, proves an image is authentic by certifying it at the moment it is created: forensic acquisition at the source captures the content in a controlled environment where no person, software or system can alter it while it is being captured, and an official digital seal and qualified timestamp, recognized internationally, make any later change detectable. A watermark marks what is generated; proof of what is real is built at the source with TrueScreen‘s patented methodology. There are two paths, depending on where the image comes from.
- Images you capture: with the TrueScreen app, photos and videos are certified as they are taken, as explained in how to certify a photo. Each file gets a SHA-256 fingerprint, a qualified timestamp and the digital seal, so any change after capture is detectable. Content certified with TrueScreen has full legal and evidentiary value worldwide, while the weight it carries in a given case stays with the court.
- Images that already exist: for an image found on a web page or published by a counterparty, the Forensic Browser captures and certifies the page, the browsing session and the media on it. Its Media Inspector lists the images, videos and audio on the page you visited, shows each file’s SHA-256 fingerprint and the markings the file declares about its origin, and runs the AI analysis. The deepfake and AI-generated content detection covers photos, video and audio, not text; it is available on every TrueScreen plan and is reachable today in the Forensic Browser, in Media Inspector.
Detection adds a signal; certification at the source establishes what happened. The AI analysis gives a probabilistic reading and doesn’t certify authenticity, so the decision stays with whoever uses the result. If the content can still be captured, capture and certify it rather than leaning on analysis alone.
Say a photo agency covers a street protest with the TrueScreen app, and two days later an edited version of one shot spreads on social media. A SynthID check on the viral copy finds nothing and says nothing about the original. The agency hands over the certification report instead: the file’s fingerprint, the date and time fixed by the qualified timestamp, and the digital seal, so anyone can compare the fingerprint with the original file. That is digital provenance in day-to-day work.
Conclusion
SynthID is worth running, as long as you read it for what it is: a one-way signal. A positive result tells you a Google or partner model touched the image; a negative result tells you the check found nothing, which is not the same as finding the image genuine. When you have to prove that an image is real, the answer is in how it was captured: certify what you produce at the source, capture what you find online together with its page, and record every watermark check, metadata reading and AI analysis as a clue that adds to the evidence.
FAQ: SynthID image checks
What is SynthID and who uses it?
SynthID is the invisible watermark Google DeepMind introduced in 2023. Google models embed it in the images, video and audio they generate without changing how they look, and partners including OpenAI, NVIDIA and Kakao use it too, with Apple announced. Google says more than 180 billion images and videos have been marked. You can check for it in the SynthID Detector portal, the Gemini app, Search and Chrome.
If SynthID finds no watermark, is the image real?
Not necessarily. A negative result only means the check didn’t find the signal. The image may be authentic, it may come from a generator that doesn’t use SynthID, or the watermark may have weakened after many edits. Google’s Gemini help page says unmarked content “could have been created by other AI systems.” A negative check is missing information, not proof that a photo is real.
Does ChatGPT use SynthID, and can Gemini detect it?
OpenAI is one of the partners whose content the SynthID Detector portal recognizes, according to Google’s October 7, 2026 announcement. The Gemini app is narrower: Google’s help page says it “can currently only recognize content created by Google AI tools.” An image from a partner model may register in the portal but not in Gemini, and an image from a generator outside the program always comes back negative.
Does SynthID give false positives, and does a positive check count as proof?
Google doesn’t publish an error rate, and research on AI watermarks (Saberi et al., arXiv 2310.00076) shows they can be spoofed so that real images appear marked. A positive result is a technical clue that the image passed through a model using SynthID. It doesn’t show who shared it, when, or whether the caption is true, so in a dispute it should be documented and weighed by an expert alongside the other evidence.
How can I prove a photo is authentic?
Certify it at the source. TrueScreen certifies a photo while it is being taken with the TrueScreen app: forensic acquisition prevents alteration during capture, and an official digital seal and qualified timestamp, recognized internationally, make any later change detectable. Content certified this way has full legal and evidentiary value worldwide. A watermark check can only suggest where an image came from; a certification records how it was captured.
Sources and verification
Every figure and principle cited here links to its source, listed with type and date. Links checked on the publication date.
| Source | Type | Date | What it supports |
|---|---|---|---|
| Google DeepMind, Identifying AI-generated images with SynthID | Document | 29/08/2023 | Beta launch, pixel watermark, not foolproof against extreme manipulations |
| Google DeepMind, SynthID product page | Document | 2026 | Declared robustness to cropping, filters, frame rate changes and lossy compression |
| Gowal et al., SynthID-Image, arXiv 2510.09263 | Survey | 10/10/2025 | Technical paper, more than ten billion images and frames marked |
| Google, SynthID Detector announcement | Document | 20/05/2025 | Launch of the SynthID Detector portal at Google I/O |
| Google, AI image verification in the Gemini app | Document | 20/11/2025 | SynthID check available in the Gemini app |
| Google, SynthID Detector open to everyone | Document | 07/10/2026 | Global opening in English, partners OpenAI, NVIDIA, Kakao and Apple, over 1 million daily checks |
| Google, SynthID among TIME’s best inventions | Survey | 15/10/2026 | More than 180 billion images and videos and 240,000 years of audio marked |
| Google, Gemini Apps Help: check if content was made with Google AI | Document | consulted 17/10/2026 | Gemini recognizes only Google AI content, negative and inconclusive results, file limits |
| Saberi et al., Robustness of AI-Image Detectors, arXiv 2310.00076 | Survey | 2023, rev. 2024 | Trade-off between misses and false alarms, watermark removal and spoofing |
| University of Waterloo, UnMarker (IEEE S&P 2025) | Survey | 23/07/2025 | Research showing watermarks can be broken |
| IEEE Spectrum, AI watermark remover | Document | 07/08/2025 | Google DeepMind disputes the removal rate attributed to SynthID |
| Roca et al., Microsoft AI for Good Lab, arXiv 2507.18640 | Survey | 2025 | 62% human accuracy in telling real and generated images apart |
| AI Act, Regulation (EU) 2024/1689, Article 50 | Law | 02/08/2026 | Marking duty for providers and disclosure duty for deployers of deepfakes |
| Digital Omnibus on AI (European Commission) | Law | 27/07/2026 | Marking deadline of December 2, 2026 for generative systems already on the market |
| European Commission, Code of Practice on marking and labelling AI-generated content | Document | 10/06/2026 | Final voluntary code of practice |
| IPTC Digital Source Type vocabulary | Standard | 2026 | trainedAlgorithmicMedia and compositeWithTrainedAlgorithmicMedia declared markings |
Prove your images are real from the moment you take them
Certify photos and videos at the source with the TrueScreen app, and capture images you find online with the Forensic Browser, with an official digital seal and qualified timestamp, recognized internationally.
TrueScreen editorial team
This section is edited by the TrueScreen editorial team, which brings together expertise in digital forensics, the law of digital evidence and regulatory compliance. Every article is checked against primary sources: legislation, published rulings, technical standards and official documentation, always cited in the text.
