Google Opens SynthID Detector to Everyone: What It Can and Cannot Catch

Google has opened SynthID Detector to the general public, giving anyone with a Google, OpenAI or Apple account the ability to check whether an image, video or audio file carries an invisible AI watermark. The tool, available at SynthID.com, returns a simple yes-or-no answer within seconds of an upload.

The launch closes a gap that has existed since 2023. Google introduced SynthID — a system that embeds an imperceptible label into AI-generated content — two years ago, but ordinary users had no way to check for that label. In 2025, Google released a limited version of SynthID Detector for selected researchers and journalists. That restricted trial is now over, and the tool is available to everyone.

The stakes are straightforward. Generative AI tools now produce images, videos and songs at a volume that makes it difficult for any internet user to judge whether what they see or hear is real. Google, which operates one of the largest generative AI businesses through its Gemini model family, is now supplying a verification instrument alongside the generation technology itself.

How does the detector actually work?

The mechanism is watermark matching, not pattern analysis. Since SynthID launched, Google's Gemini model has embedded a watermark in everything it creates. SynthID Detector scans an uploaded file for that watermark and reports whether one was found.

The workflow is minimal:

• Go to SynthID.com.
• Upload an image, video or audio file.
• Wait a few seconds for Google to return the result.

The new tool improves on Google's previous verification path in a concrete way. Before now, users who suspected a piece of content was AI-generated had to ask Gemini itself — an approach that asked skeptics to trust a generative AI model's self-reporting, and that was a non-starter for people deliberately avoiding the technology altogether.

SynthID Detector also extends beyond Google's own output. The tool can scan for watermarks embedded by Google's SynthID partners, which the previous process could not do. That means it can identify media created with AI models from OpenAI, NVIDIA and Kakao, in addition to Gemini. Apple participates in SynthID as well, but detector support for Apple's creations remains in development.

How reliable is the answer?

The design trades coverage for precision. Because SynthID Detector checks for a specific watermark rather than analyzing media for statistical patterns that indicate AI use, it should eliminate the chance of false positives — the failure mode that has plagued general-purpose AI detectors, which frequently flag human-created work as AI content.

That simplicity imposes limits of its own. The most obvious one: the tool cannot detect media made with AI models that do not participate in SynthID, such as Meta's tools. The detector is also strictly binary in its reporting. It tells users whether it finds a label in the uploaded file, and provides no information about which parts of that file were AI-generated.

Google says its SynthID signatures persist even when AI generations are edited after creation. The company acknowledges no guarantee against circumvention, but it has built friction into the system. A Google spokesperson told Ars Technica that users are limited to roughly 10 checks per day, a cap designed to prevent anyone from using the detector's output to reverse-engineer and strip the SynthID watermark. The login requirement — Google, OpenAI or Apple account — adds a second layer of accountability.

Where does SynthID sit among the competition?

SynthID Detector is not the only generative AI detector on the market, but its watermark-first architecture likely makes it one of the most reliable. The competitive picture, as documented in the source material:

• OpenAI, a SynthID partner, runs its own detector tool that can pick out AI-confirming metadata as well as SynthID watermarks.
• Meta embeds its own watermark, called Content Seal, into its AI creations and offers a web-based detector tool. A 2026 report from Reuters found that Meta's detector sometimes failed to accurately spot its own seals.
• Anthropic and Microsoft have begun watermarking AI media but do not currently offer consumer tools to detect those marks.

The structural weakness cuts across every one of these tools. Each detects a specific AI marker, and no reliable detector exists that can identify generative AI output across the board. Researchers have built AI-spotting tools of their own, but those remain unavailable to the public. The general-purpose AI detectors that are openly accessible produce unreliable results and are prone to flagging human-created work as AI content.

What does the release change?

For the first time since SynthID's 2023 debut, watermark verification is a public utility rather than a privilege extended to researchers and journalists. A user who encounters a suspicious image can, in seconds and at no cost, determine whether it came from Gemini, OpenAI, NVIDIA or Kakao models — provided the creator's toolchain embedded the mark and the roughly 10-check daily limit covers the user's needs.

The release does not resolve the detection problem. Media generated with non-participating models such as Meta's remains invisible to SynthID Detector, and the tool's yes-or-no output offers no granularity on partial manipulation. Apple support is still pending. What the launch establishes is a working baseline: a low-false-positive public checker backed by the major model vendors, positioned against a field of unreliable generic detectors. Whether SynthID's watermarks survive sustained adversarial pressure — the scenario Google's rate limits are built to slow — will determine how long that baseline holds.