With the rise of generative artificial intelligence tools in recent years, whether for images, videos, or music, internet users often question the authenticity of what they see and hear online. As one of the leading providers of generative AI, Google is now working to make it easier for people to identify content produced by its technology. Through the new SynthID detector, anyone can check images, videos, or audio files to determine if they were generated by artificial intelligence, though there are some important limitations.
Google launched SynthID in 2023, embedding imperceptible labels in AI-generated content, but until now, regular users have not been able to actually check these labels. In 2025, Google released a limited version of the SynthID detector for certain researchers and journalists, and now this tool is available to everyone. While this is good news for those looking to verify whether content was generated by AI, it is not foolproof. Since the launch of SynthID, Google's Gemini AI model has embedded watermarks in all content it creates.
The SynthID detector provides users with a simple "yes" or "no" answer by scanning the watermark to determine if it has been found.
Google's SynthID detector offers a straightforward content verification method
Using this new tool is very simple; users just need to visit SynthID.com, upload an image, video, or audio file, and then wait a few seconds to receive Google's results. This represents a significant improvement over Google's previous process. Previously, users could ask Gemini itself whether a particular piece of content was generated by AI, but for those who do not trust generative AI, it was difficult to fully trust Gemini's answers, making it an unacceptable option for those trying to avoid this technology.
The SynthID detector also allows users to scan for watermarks in content created by Google SynthID partners, which was not possible before. This means the tool can determine whether media was created by AI models from OpenAI, NVIDIA, and Kakao. In addition to Gemini, Apple is also involved with SynthID, but the capability to detect its creations is still under development. Since the SynthID detector only checks for watermarks and does not analyze media for patterns indicating AI usage, the chance of false positives should be minimized.
However, this simplicity also brings its own limitations, the most obvious being that it cannot detect media created by AI models that do not use SynthID, such as those from Meta.
Limitations of the SynthID detector compared to other tools
The SynthID detector is also limited to telling users whether a label was found in the uploaded file, without providing any information about which parts were generated by AI. Google states that its SynthID signature will persist even if the AI-generated content is edited. While there is no guarantee that some malicious forgers won't eventually find ways to bypass this, the company appears to be taking steps to prevent this from happening. According to reports, a Google spokesperson told Ars Technica that users are limited to about 10 checks per day to prevent them from exploiting the detector's output to find ways to remove SynthID.
Users must also log in with a Google, OpenAI, or Apple account to use the SynthID detector.
The SynthID detector is not the only available generative AI detector, but despite its limitations, it may be one of the most reliable. SynthID partner OpenAI has its own detection tool that can identify AI-verified metadata as well as SynthID watermarks. Meta has its own watermark called Content Seal, embedded in its AI creations, and provides a web-based detection tool, although a 2026 report by Reuters found that it sometimes fails to accurately identify these watermarks.
Companies like Anthropic and Microsoft have also begun watermarking AI media, but consumer detection tools are not yet available. The drawback of all these tools is that they are designed to detect specific AI markers, without a reliable detector that can comprehensively identify generative AI. Many researchers have developed their own AI identification tools, but these are not publicly accessible. For general-purpose AI detectors that are publicly available, their results are unreliable and prone to mislabeling human-created works as AI content.