Skip to content

SynthID Detector (Google DeepMind)

FreeFirst-line

Vendor: synthid.google | Type: Non-detector | Cost: Free

TL;DR

Google DeepMind's portal for detecting SynthID watermarks embedded in content generated by Google's own AI products (Gemini, Imagen, Veo). Used by Thai PBS on the Anutin Charnvirakul image in February 2026. A clean non-detector provenance signal when the suspect content was generated through a Google product.

What it does

SynthID Detector accepts an image, audio file, or text input and checks for the presence of an imperceptible SynthID watermark that Google embeds at generation time across its production AI products. The output is a watermark-presence verdict: yes, this content was generated through a SynthID-instrumented Google model; no, no SynthID watermark detected. The portal is currently in pilot / limited access; the underlying text watermarking library is published as open source under Apache 2.0.

The tool is a non-detector provenance signal under Architectural Anchor 1: it does not estimate "is this AI-generated" probabilistically the way a detector does, it answers a definite watermark-presence question. When a SynthID watermark is present, the provenance signal is unambiguous within its scope. When no watermark is present, the signal is "this content was not generated through a SynthID-aware Google model." That does NOT mean the content is human-made. It means the question routes elsewhere (other provenance via Content Credentials Verify, or a detector pass via Hive).

When to use it

  • A suspect image looks like it might have been generated through Gemini, Imagen, or Veo (Google's production AI stack) and you want a clean provenance signal, not a detector probability.
  • A Thai or regional case mirrors the Thai PBS Anutin Charnvirakul workflow where SynthID provided the verifiable signal that classical detection alone would not have produced.
  • You want to demonstrate the toolkit's preference for provenance signals over detector signals when both are available.
  • You are training fact-checkers on the difference between watermark-based provenance (definite within scope) and probability-based detection (one weak signal).

Limitations

Limitations

  • Only works on content generated by SynthID-instrumented models, currently Google production (Gemini, Imagen, Veo). Useless for Llama, GPT-4o, Claude, DeepSeek output.
  • Watermark breaks under heavy paraphrase or translation.
  • Reverse-engineered by aloshdenny/reverse-SynthID with claimed 90% detector, meaning the watermark is not unforgeable in principle; the toolkit excluded the reverse-engineered tool from the shortlist because reframing watermark verification as detection violates Anchor 1.
  • In pilot / limited access status; portal availability and coverage are subject to Google's roadmap.

Privacy and threat model

The SynthID portal uploads the input content to Google DeepMind's infrastructure for watermark reading. Google is a US-jurisdiction vendor; the upload routes through Google Cloud's policy. For routine provenance checks on already-public content, the upload is low-risk because the content is already public; the verdict (yes/no SynthID watermark) is not derived from any pixel-level forensic operation that would expose source-identifying detail.

For source-identifying material, the upload chain itself is the relevant concern; classify accordingly before use. Note that SynthID is by design a verifier-side signal; if you suspect content was generated through a non-Google model, SynthID will not return a useful signal and there is no operational reason to route the file through it.

Country and platform applicability

  • Indonesia: no documented use recorded; tool applies in principle for any Gemini- or Imagen-generated content.
  • Laos: no documented use; language-agnostic.
  • Malaysia: no documented use; tool applies in principle.
  • Philippines: no documented Rappler / VERA Files deployment recorded.
  • Sri Lanka: no documented use; tool is language-agnostic.
  • Thailand: documented use by Thai PBS on the Anutin Charnvirakul / Mauerberger image (February 2026), accessible via Google Lens at FLT. This is the toolkit's clearest provenance-first worked example.

Documented case base is concentrated in Thailand because the Anutin / SynthID case is the cleanest provenance-first verification in publicly available regional documentation. Coverage across the other five focus countries is structurally limited by SynthID's scope (Google production only).

Platform applicability: works on any SynthID-watermarked content regardless of source platform.

How to access

The SynthID Detector portal is accessible through Google's pilot programme; current entry is via deepmind.google/synthid and the linked verifier interface. Pilot / limited-access status as of May 2026; access conditions may shift. Google Lens also surfaces SynthID watermark detection on supported content (this is the path Thai PBS used in the Anutin case).

The SynthID text watermarking library is published as open source under Apache 2.0 on GitHub for organisations integrating SynthID verification into their own pipelines.

Cost (current as of 2026-05)

Free, with limited-access status during the pilot. Apache 2.0 for the text library. No paid tier documented; no quota for ordinary verifier use.

Quickstart

  1. Open the SynthID portal (currently deepmind.google/synthid) in your browser.
  2. Upload the image, audio file, or text input.
  3. Wait for the watermark-read result.
  4. If a SynthID watermark is detected, record this as a definite provenance signal: the content was generated through a SynthID-instrumented Google model.
  5. If no watermark is detected, the content either was not generated through Google's stack or had its watermark broken by heavy paraphrase / translation. Route to Content Credentials Verify for C2PA provenance, then to Hive for a detector signal if no provenance signal is available.
  6. On Thai PBS-style cases, Google Lens is an alternative entry: right-click the image inside Google Lens and look for the SynthID watermark indicator.

In the toolkit's workflow

Provenance pillar non-detector tool per Architectural Anchor 1. Sits in 1A.4 First-Line Triage provenance / watermark as the alternative to Content Credentials Verify. SynthID covers the Google generation stack; CCV covers the C2PA / Adobe / hardware stack. Together they cover the two main provenance ecosystems relevant to fact-check verification.

Standard combinations:

  • With Content Credentials Verify as the C2PA-side provenance check; the two together cover Google-generated and C2PA-instrumented content respectively.
  • With C2PA Conformance Explorer at 1A.4 when the manifest itself is under question.
  • With Hive at 1A.1 when neither provenance check returns a hit and a detector pass becomes operationally necessary. In the Anutin / Mauerberger case, the SynthID hit resolved the question without invoking the detector.

This card produces a non-detector signal: the watermark verdict is a provenance signal that does not depend on probability scoring. Per Anchor 2, a SynthID hit combined with a Hive verdict is two signal classes (provenance + detector); the SynthID hit alone is one signal class and is unambiguous within its scope.

Decision-tree references: T4 provenance triage – SynthID is the vendor-watermark check at T4.7 inside the Google-generation branch.

Override notes

Override notes

  • SEA Google-only disclaimer (sea-google-only-disclaimer): the tool's coverage is bounded to content generated through Google's production stack (Gemini, Imagen, Veo). Most SEA content reaching fact-checkers will not have been generated through Google's models; the Limitations and How to access sections name this scope explicitly; broader coverage is not implied.

  • Non-detector declaration (g2-non-detector-declaration): the tool's signal class is non-detector (provenance). Every watermark verdict is a provenance signal, never a detector signal. The card and the workflow section name this classification.

Sources