Google SynthID watermark detector

The portal takes one file per check after sign-in with a Google or Apple account: images in JPG, JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, TIF or GIF; video in MP4, MOV or WEBM; audio in WAV, MP3, OGG, FLAC, AAC or M4A (synthid.com, read 9 November 2026). Its additional terms forbid using the frontend "as an automated API to run bulk media verification queries" and creating multiple accounts to "circumvent the service's rate-limiting restrictions", so the API field on this card is marked unavailable and the daily quota is stated without a number. The Gemini app states its quota: "approximately 10 image checks, 10 video checks (up to 5 minutes total), and 10 audio checks (up to 3 hours total) within any 24-hour rolling window", on files of 100 MB or less, with video under 90 seconds and audio under 1 hour (Gemini Apps Help, read 9 November 2026).

There is no paid plan. Both surfaces are free with a Google sign-in, and SynthID Text is open source, so the plans field is empty and the table below records the two public surfaces, read 9 November 2026.

Retention is stated in hours. The portal's privacy notice reads: "Media files you upload for verification are processed in real-time. Once the detection results are returned to you, the original media files are immediately deleted", and, to enforce quotas, "we securely store a digital signature of your media. This signature is mapped to your user identifier and permanently deleted after 24 hours" (synthid.com, read 9 November 2026). The card's retention rank field takes the 24-hour signature as 1 day, the longest period the notice states, with the immediate deletion of the file itself quoted beside it. The portal "is hosted by Google and is governed by the Google Terms of Service", which fixes ownership for the card.

Google publishes no accuracy percentage for the SynthID Detector. The portal's FAQ says watermark verification "provides high reliability and resilience, though it is limited to identifying content generated by SynthID-integrated models", and on errors: "In very rare cases SynthID can also falsely trigger on non-watermarked content" (synthid.com, read 9 November 2026). No rate accompanies either sentence. The Gemini help page states the reading rule for a negative: "If a SynthID watermark isn't detected, it means the image or video wasn't created or edited by Google AI, but it could have been created by other AI systems" (Gemini Apps Help, read 9 November 2026).

The vendor's technical account is a Nature paper by Google DeepMind authors, Dathathri et al., "Scalable watermarking for identifying large language model outputs", Nature 634, 818 to 823, published 23 October 2024 (doi.org/10.1038/s41586-024-08025-4, read 9 November 2026). The card treats it as a vendor paper, not a study record: its competing interests statement reads "Work funded and performed by Google DeepMind", with patent applications filed "on behalf of DeepMind Technologies Limited". The paper reports detectability as "the true-positive rate (TPR) for a fixed false-positive rate (FPR) of x%", so its 1% false-positive setting is a threshold chosen on the authors' own unwatermarked texts, not a measured error rate on outside data. It also states that "generative watermarks are weakened by edits to the text, such as through LLM paraphrasing", that watermarking "require[s] coordination between actors running the LLM text generation services", and that quality was checked across "nearly 20 million Gemini responses" with thumbs-up rates differing by 0.01%.

The developer documentation carries the robustness caveat in plain terms: SynthID Text watermarks survive "cropping pieces of text, modifying a few words, or mild paraphrasing", but "Detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten, or translated to another language" (ai.google.dev, read 9 November 2026). A watermark can only be found where one was put, which is the sentence the whole card rests on.