Hyderabad: As Artificial Intelligence makes it harder to tell real content from AI-generated media, technology companies are developing tools to improve transparency. One such tool is SynthID, an invisible watermarking technology developed by Google DeepMind.
Need for easy identification of AI content
The tool adds hidden digital markers to AI-generated text, images, audio and video. These markers help identify content created using AI.
Google introduced SynthID in 2023 for AI-generated images. Since then, it has expanded to other forms of media and is now being adopted beyond Google’s own products.
Google’s SynthID partnership
Google has partnered with several AI companies to expand the use of SynthID beyond its own products.
The companies include Nvidia, OpenAI, Kakao and ElevenLabs. Nvidia plans to use SynthID in its Cosmos foundation models, while OpenAI will combine SynthID watermarking with C2PA metadata to improve content verification. The partnerships are part of a wider effort to help users identify AI-generated content and increase transparency across AI platforms.
How does SynthID work?
We asked AI chatbots, including Google Gemini and ChatGPT, how SynthID detects AI-generated elements in content.
The chatbots said that SynthID tags AI-generated media with an invisible digital fingerprint at the moment of creation.
The technology embeds tiny signals into image pixels, video frames and audio frequencies. It makes subtle adjustments to word choices to create a hidden statistical pattern for text. They can survive editing and compression because these watermarks are embedded directly into the content rather than its metadata.
How effective is SynthID?
SynthID can help identify content generated by AI systems that support the technology.
The watermark is embedded directly into the content and can often survive common edits, compression and format changes. This makes it more reliable than traditional metadata-based labels, which can be easily removed.
The tool can detect content generated by other AI models using SynthID.
Limited adoption across AI models
SynthID is not a perfect solution. Its effectiveness depends on AI developers adopting the technology in their models and platforms.
Many popular AI tools do not currently use SynthID or any watermarking system at all.
Content generated by such models cannot be detected through SynthID.
Can watermarks be removed?
There are reports that state watermarking methods can be weakened or removed through extensive editing, cropping, re-encoding or by passing content through other AI systems.
There are several videos, tutorials and websites online that teach techniques for removing or weakening watermarks embedded in AI-generated content. Companies continue to improve watermarking technology but no system is completely resistant to manipulation.
Another challenge is that AI-generated content often spreads across multiple platforms. The content may be altered, reposted or combined with other media elements. Detecting the original watermark can become more difficult for the tool.
Do AI companies need to share user data?
Even if multiple AI companies adopt SynthID, detection does not necessarily require them to share users’ data or prompts.
However, they can share a common watermarking standard that allows verification tools to recognise AI-generated content while keeping the underlying generation data private.
For example, if Google Gemini generates an image with a SynthID watermark, OpenAI may be able to detect the presence of the watermark without accessing the user’s prompt or any data stored by Google.
The tool would only analyse the watermark embedded in the content itself.
It is also important for users to know which AI tool generated a piece of content so that they can use the appropriate verification tool.
Understanding limitations
The expansion of SynthID is likely to help researchers, journalists and fact-checkers identify AI-generated content more effectively. However, the technology still has limitations. Its success will depend on wider adoption across AI platforms and continued improvements to make watermarks more robust and easier to verify.
