Google launches public AI detector scanning for watermarks from Gemini, OpenAI, NVIDIA and Kakao models

Google has made its SynthID Detector available to the general public, allowing users to verify whether digital media contains embedded AI labels. According to Engadget, this tool scans images, videos, and audio files for imperceptible watermarks added by participating generative AI systems. The release expands access beyond the limited group of researchers and journalists who tested the service in 2025.

How does the detection tool function?

The detector operates by scanning uploaded files for specific watermarks rather than analyzing visual or audio patterns for signs of AI generation. Google introduced the underlying SynthID technology in 2023 to add these labels to content. As reported by Engadget, the current tool checks for the presence of these embedded marks and returns a simple yes or no result. This approach aims to eliminate false positives that can occur with pattern-based analysis.

Which AI models are covered?

The detector can identify watermarks from Google’s Gemini model, which has embedded labels in its creations since the technology's launch. According to Engadget, the tool also scans for marks from Google’s SynthID partners, including OpenAI, NVIDIA, and Kakao. Apple participates in the SynthID program, but Engadget notes that support for identifying Apple’s creations in the detector is still in development.

What are the technical limitations?

Because the tool relies on the presence of a specific watermark, it cannot detect media generated by AI models that do not use SynthID. Engadget highlights that this means content created with tools from Meta, for example, will not be flagged. The system is designed to verify the presence of a label, not to determine if content is AI-generated through independent analysis, which limits its scope to participating vendors.

Questions to ask your suppliers

Security teams should ask vendors whether their AI-generated outputs include verifiable provenance labels compatible with industry standards. Organizations need to determine if suppliers are participating in watermarking initiatives like SynthID. Risk managers should verify that third-party content providers have clear policies regarding AI disclosure. Internal communications teams must be informed about which vendor outputs can be technically verified and which rely on contractual assurances alone.

Background: Generative AI

Generative AI creates original text, code, or images by predicting the next likely element in a sequence based on training data. It does not "think" or "know" things. Instead, it calculates probabilities. This distinction matters because the system can produce plausible but entirely fabricated information with high confidence.

Read the full guide: Generative AI Defined: Mechanics, Origins, and Operational Reality

What to do and how to stay safe: Google

• Verify critical media using the detector to check for known watermarks from participating vendors.
• Maintain manual review processes for content from non-participating providers like Meta.
• Update content verification policies to reflect that watermark absence does not prove human origin.
• Monitor vendor announcements for expanded support of additional AI models in detection tools.

Step-by-step guide: AI Usage Policies at Work: How to Prevent Data Leakage

General security guidance from the TechSec Feed newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Can the detector identify all AI-generated content?

No, it only detects content with SynthID watermarks from participating vendors like Google, OpenAI, NVIDIA, and Kakao.

Does the tool analyze image patterns for AI signs?

No, it scans for embedded watermarks to provide a yes or no result on label presence.

Is Apple content currently detectable?

Apple participates in SynthID, but Engadget reports that detector support for its creations is still in development.