Google SynthID Detector: Portal Verdicts and File Limits

What Is Google’s SynthID Detector at synthid.com?

Google’s SynthID detector is an online check service at synthid.com built to find hidden marks placed by AI models. Google DeepMind first showed SynthID for pictures on August 29, 2023. That early tool ran on Vertex AI with Imagen. It used two models trained side by side: one model added the watermark to pixels, and the other model scanned for the mark. It gave three confidence levels: detected, possibly detected, and not detected.

Google expanded the technology across several years of testing. On May 20, 2025, Google announced the portal at Google I/O and opened it to early testers such as journalists and media professionals. The same post said NVIDIA was watermarking videos from its Cosmos preview service with SynthID and that over 10 billion pieces of content had been watermarked. On October 7, 2026, Google launched global public access in English.

The service acts as a watermark decoder. It is not a broad tool to tell human work apart from synthetic output. The main page states this fact plainly: ‘This is not a general AI detector’.

The tool does not look for image flaws or unnatural speech cadence. It does not guess if human hands made the file. Instead, it looks only for the hidden SynthID pattern woven into pixels or audio waves by partner models.

During the global launch in October 2026, Google gave new figures for the system: over 180 billion images and videos and over 240,000 years of audio carry the mark, and built-in verification features in Google Search, the Gemini app and Chrome handle over 1 million requests a day. If you want background first, see our SynthID remover overview or our guide to how SynthID works across modalities.

The synthid.com site runs as a single-page web app. The interface gives users five main paths to visit:

• The root path / provides the drop zone to upload files.
• The /image-detection route displays analysis results for pictures.
• The /audio-detection route processes spoken and musical sound tracks.
• The /video-detection route inspects video streams and embedded sound.
• The /faq and /terms pages explain service rules and legal terms.

If you open a detection page without picking a file, you see a message that says ‘No media selected’. The full web layout uses English (en-US) and lets users pick system, light, or dark visual modes.

Portal Requirements: Sign-In Options, File Formats, and Limits

The synthid.com portal asks users to log in, accepts one file per check up to 100 MiB, and reviews images, audio, and video. If you try to upload a file before logging in, you see an alert: ‘Please sign in before detection’. The portal gives three login choices: Google, Apple, and ChatGPT using OpenAI OIDC. Logging in with Apple is just an identity option. It does not mean Apple currently embeds SynthID marks into its own media.

The client checks file sizes in your browser. It sets a strict ceiling of 104,857,600 bytes (100 MiB) for every kind of file. If a file is too heavy, the portal stops it right away with the message ‘File is too large’.

Empty files show ‘File is empty’. Files with unlisted extensions show ‘Unsupported file type’. The file picker takes one file at a time.

Image checks cover common formats like JPEG, PNG, and WebP, alongside newer formats such as AVIF, HEIC, and HEIF. For audio, the site takes lossy files like MP3 and AAC, plus high-fidelity FLAC and WAV streams. For video, users can upload MP4, MOV, or WebM clips. High-bitrate 4K video clips or long audio recordings may exceed the 100 MiB cap. In those cases, you must trim or compress the file before uploading.

The public portal does not accept text files. Google DeepMind published its SynthID Text method in Nature 2024, and the team shared reference code on GitHub in google-deepmind/synthid-text. The synthid.com portal itself handles only image, video and audio files.

In browser code examined on October 8, 2026, the partner list named four groups: Google AI, NVIDIA, OpenAI, and Kakao. If a marked file comes from a partner without a custom label, the site shows the fallback label ‘AI from a SynthID partner’. Google’s launch post says Apple is coming soon, but Apple was not in the live partner list on October 8, 2026. NVIDIA’s Cosmos watermarking was announced at I/O in May 2025, and OpenAI announced its SynthID adoption for images in May 2026.

How to Read the SynthID Detector Verdicts and Unsure Codes

The SynthID detector client returns three main verdicts: a positive detection result, a negative outcome, and an unsure status. If a server drop occurs, the client displays a fourth state: Try Again. In client scripts reviewed during October 2026 testing, the UI shows ‘Google AI’ headers by default, while tracking distinct event tags in the background.

When a video file has a mark, the portal tells you which track carries it. The header shows (Video), (Audio), or (Audio & Video).

If the portal cannot make a clear call, it returns an unsure result. The client code shows nine distinct reasons for this outcome. The tool swaps the media type into the message:

• Code 0: ‘Not enough information to detect SynthID’.
• Code 1: ‘The uploaded media was empty’.
• Code 2: ‘The [modality] resolution is too low’.
• Code 3: ‘Not enough details to watermark’.
• Code 4: ‘The [modality] has less than 10 frames per second’.
• Code 5: ‘The [modality] is too low quality’.
• Code 6: ‘We suspect the image has been tampered with or Google AI was only used for a very small part’.
• Code 7: ‘The [modality] is too long to be processed’.
• Code 8: ‘The [modality] is too short to be processed’.

The reasons are short and the portal does not explain them further. What can be read from the wording: code 4 is a hard rule for video, so a clip under 10 frames per second gets an unsure status rather than a negative result; code 6 covers files where Google AI touched only a small part, or where the detector suspects tampering; codes 7 and 8 mean the file is outside the length the service processes. Google does not publish the resolution or length thresholds behind codes 2, 7 and 8. To learn more about video checks, read our guide on SynthID video detection rules.

What Does a Negative ‘Pass’ Result Actually Mean?

A negative verdict on the web portal means only that the tool did not spot a confident watermark. It does not prove that a human made the file: the synthid.com FAQ says a not-detected result “does not totally rule out AI generation or manipulation”. When the site says ‘Not made with Google AI’, three different real-world situations can explain the outcome:

1. Human creation: The file was shot on a physical camera, recorded with a real microphone, or made without generative AI.
2. AI without SynthID: The file came from an AI model that does not use SynthID, or was made before a partner joined the program.
3. Weakened watermark signal: The file had a mark, but heavy edits, harsh compression, or filters wore the signal down.

Google DeepMind’s SynthID overview says the watermark is ‘designed to stand up to’ modifications like cropping, adding filters, changing frame rates and lossy compression. The synthid.com FAQ adds that the mark is ‘not infallible’, so heavy changes can degrade it until the tool can no longer spot it.

A negative scan does not rule out synthetic media. Many image and audio generators are not SynthID partners and do not embed it, so their output has nothing for the detector to find. That negative output is accurate, but it does not mean the file is camera-original.

The portal FAQ also notes that false positives can happen on unmarked files, though they are rare. Because of this, Google tells users not to rely on SynthID as a single proof point. To understand how images change under edits, read our guide on removing SynthID from image files.

Terms of Service, Privacy Policies, and Quota Limits

Google sets firm anti-abuse rules on synthid.com to stop bot scans and adversarial tests. The Terms of Service page adds strict forensic rules on top of the standard Google terms:

• No adversarial testing: You may not run black-box tests or try edit chains to weaken, break, or bypass the detector.
• No automated scraping: You may not send script calls, scrape pages, or use the web front as an unofficial API.
• No quota circumvention: You may not make extra accounts or swap user details to slip past daily caps.

Breaking these rules can lead to immediate suspension of access, according to the Terms.

Privacy terms on synthid.com describe two things. Uploaded files are processed in real time and deleted as soon as the result is returned. To enforce daily quotas and defend against automated attacks, Google stores a digital signature of the media, mapped to your user identifier, and deletes it after 24 hours.

When you log in with a third-party account such as Apple, Google stores the identifiers that provider shares, including email address, user ID and provider details, to manage the session, prevent abuse and prevent duplicate accounts.

The interface does not show a public counter with remaining checks. When you hit a ceiling, the portal shows a ‘Quota Exceeded’ message that says how many hours to wait. The client has two kinds of quota message: one for a whole media type, and one for ‘images similar to this one’, which suggests repeat uploads of near-identical files are limited separately.

How synthid.com Compares to Gemini, Search, and C2PA Checks

The standalone synthid.com portal differs from Google Gemini and Google Search in partner scope, upload allowances, and metadata handling. Google runs watermark checks across several apps, but the tools follow different rules.

Google says its built-in verification features in Search, the Gemini app and Chrome handle over 1 million requests a day. Google’s October 2026 post does not say whether those surfaces check partner watermarks or only Google’s own.

The Gemini web app and mobile tools also carry built-in AI checks described in Google Gemini Help. But Gemini works under tighter limits than synthid.com:

• Partner coverage: Gemini checks only for Google AI marks, while synthid.com scans for Google AI, NVIDIA, OpenAI, and Kakao marks.
• File length caps: Gemini caps videos at 90 seconds and sound clips at 1 hour.
• Daily usage caps: Gemini allows roughly 10 images, 10 videos (up to 5 minutes total), and 10 audio clips (up to 3 hours total) per rolling day.
• Uncertainty notices: Gemini uses unique notes like ‘Not enough details to watermark’ and ‘Likely too small an edit’.
• Size limit: both cap a file at about 100 megabytes (Gemini says 100 MB; the synthid.com client enforces 100 MiB).

A key difference exists between SynthID pixel marks and C2PA metadata tags. SynthID weaves tiny changes straight into pixels, video frames, or audio frequencies. C2PA attaches cryptographically signed data packets to file headers. Gemini checks both channels at once. In contrast, synthid.com looks only at SynthID marks.

Standard cleaning tools remove C2PA headers with ease, but they leave pixel-level SynthID marks in place. To see how these two systems compare side by side, read our SynthID versus C2PA metadata comparison.

Using the SynthID detector provides transparency into whether media originates from supported AI systems. While Google continues expanding its detection portal and partner network, creators requiring independent signal processing can monitor upcoming capabilities. Early access opens on synthidpass.com first.