Google is opening its SynthID Detector to the public. Since October 7, the company says the tool has been available worldwide in English, checking images, videos, and audio files for invisible AI watermarks. That makes a resource previously aimed primarily at media professionals easier to access: Anyone receiving a suspicious clip can look for a technical clue to its origin instead of relying solely on odd-looking hands or unnatural voices.
Key takeaways
- SynthID Detector checks uploaded media files for SynthID watermarks. It is available at synthid.com.
- Alongside its own tools, Google names OpenAI, Nvidia, and Kakao as partners. Apple is expected to follow; that does not promise Apple support today.
- A detection is a clue to AI generation or editing. A negative result confirms neither a file’s authenticity nor the truth of its caption.
- Google recommends using the highest-quality file possible. Origin, date, and context still require separate investigation.
What the tool looks for in an upload
SynthID marks content when it is generated. The identifier is embedded directly in the image, video, or audio signal and is designed to be imperceptible during normal viewing or listening. The detector subsequently looks for that signal. This differs from a general assessment of whether a scene looks artificial or a voice sounds unusual.
According to Google’s technical description, the marks are designed to survive common edits such as cropping, filters, or lossy compression. That is an engineering goal with limits. In its help documentation for the related Gemini verification feature, Google notes that watermarks may become undetectable after many changes. The quality and processing history of the file therefore remain relevant.
Public access broadens the tool’s usefulness beyond a small group of investigators. A newsroom can run an additional check on a submitted clip; a teacher can question the origin of an illustrative image. A video forwarded in a family chat can also be checked more deliberately. The value is an additional, technically grounded clue that can guide further research.
Google puts the existing total at more than 180 billion watermarked images and videos. That is a company figure describing adoption, rather than an independent reliability test. Similarly, naming a partner does not mean that every older piece of content or every product from that company can be identified. What matters is whether the specific file carries a supported watermark.
How to approach checking a file
The starting point is the SynthID Detector website. Its interface includes a drop area and a file-selection button. Before use, a dialog displays terms of service and privacy information. Anyone wishing to use the service should read those notices and then select the image, video, or audio file to check. This description is not based on our own file test.
The page recommends the highest possible file quality for accurate detection. Our practical inference: If the original file is available, it is a better starting point than a screen recording that has been forwarded repeatedly. It also helps to record who supplied the file and what claim accompanied it. Otherwise, it is easy to lose track of what the check was supposed to establish.
After the upload, the question is the reported watermark, rather than a comprehensive assessment of the event. A positive result should therefore be documented alongside the exact version of the file. If a different version or a new crop circulates later, the earlier check cannot automatically be applied to it. The underlying material for the investigation has changed.
The upload itself is also a deliberate transfer to an online service. According to the displayed privacy notice, Google deletes original media files immediately after returning the result. A digital signature of the file is linked to the user identifier for abuse prevention and deleted after 24 hours. These statements describe Google’s announced processing; they are not the result of a deletion audit we performed.
Why detection and authenticity are different questions
An AI watermark can indicate that an image was generated entirely or modified using AI. That alone does not establish which claim about the image is false. A harmless illustration can be clearly marked. Conversely, a genuine photograph can mislead when paired with an invented location or an incorrect date, without any AI being involved.
A missing SynthID detection therefore answers only a narrow question: The mark being sought was not identified in this file. Other AI tools, missing marks, or modified files remain possible explanations. Such a result cannot clear a viral clip of suspicion. One possible clue to its origin therefore remains unresolved.
We explained the same underlying issue when covering textGrain and watermarks in ChatGPT text: A detector checks a specific marking system. The new SynthID access, by contrast, concerns media files and an expanding partner network. Sharing the term watermark does not create universal detection of all AI output.
Content Credentials are a complementary technology. These provenance records can contain information about a file’s creation and editing history; their specification comes from the Coalition for Content Provenance and Authenticity, or C2PA. Their approach differs from embedding an invisible signal in media content. For readers, combining them is useful: A watermark, documented provenance, and a reverse image search can answer different parts of the same question.
More verifiable origins, rather than instant verdicts
Opening access is a concrete advance because more people can use an additional verification method. Its useful role is at the beginning of an investigation: What clues exist, who first published the file, and does the claimed timing fit? If a technical detection makes those questions more precise, the tool is already helping. Whether a scene actually happened remains a question of evidence about that particular scene.