SynthID Detector goes public, but won’t catch Meta AI

Google opened its SynthID Detector to the general public this week, allowing anyone to upload an image, video or audio file and check for embedded watermarks. The tool scans exclusively for a specific digital signature, not visual artifacts or text patterns, meaning it will never flag content generated by platforms outside its network.

The rollout expands access that was previously restricted to researchers and journalists in 2025. Users navigate to SynthID.com, upload a file and receive a binary result. To run a scan, individuals must be logged into a Google, OpenAI or Apple account. According to a report from Ars Technica, a Google spokesperson said daily usage is capped at roughly 10 checks per user to prevent bad actors from reverse-engineering the watermark through repeated testing.

Originally introduced in 2023, SynthID embeds an imperceptible label directly into media produced by the Gemini AI model. The new detector now also recognizes watermarks applied by partner companies, including OpenAI, NVIDIA and Kakao. Apple participates in the program, though support for identifying its creations within the detector remains under development.

Because the system looks only for a cryptographic marker rather than analyzing the content itself, it avoids the false-positive problem that plagues heuristic AI detectors. That architecture creates a massive blind spot. Any media generated by models that do not implement SynthID — most notably Meta’s generative tools — will return a negative result regardless of its origins. The tool also cannot isolate which specific portion of a composite file carries the watermark.

Fragmented detection infrastructure has become the industry standard. OpenAI operates a separate tool capable of reading both metadata and SynthID watermarks. Meta applies its own Content Seal watermark and hosts a web-based checker, though a 2026 Reuters investigation found the platform frequently failed to identify its own tags. Anthropic and Microsoft have started embedding watermarks in their outputs but offer no consumer-facing verification interfaces.

Independent researchers have built cross-platform spotting tools, yet none are available to the public. General-purpose detectors currently on the open market remain prone to misclassifying human-created work as synthetic. Google maintains that its SynthID signatures survive post-generation edits, though the company acknowledges no guarantee against future circumvention methods.