What if you see an image or video that’s been viewed and spread online millions of times and you can’t believe it’s real, only to find out it was created by an AI? That’s frustrating. Artificial intelligence is creating images, video, audio, and written content at an incredible pace, making it extremely difficult to distinguish between human-generated and machine-generated material. Generative AI tools can now generate lifelike videos, realistic images, and text.
So while this innovation has introduced opportunities for creativity and productivity all in one place, it has also created misinformation, deepfakes, and content that lacks authenticity. So, to address this issue, Google DeepMind developed SynthID, a watermarking technology that invisiblely stamps AI-generated content for later detection and verification. SynthID helps the public, journalists, and even social media platforms point out whether digital media has been created or modified using AI tools, thereby increasing transparency.
Read more: ChatGPT creator Sam Altman reveals Google’s biggest weaknesses
Read more: ChatGPT wants to remember everything – OpenAI’s Sam Altman has a shocking new goal
What is SynthID and how does it work?
SynthID is essentially a digital watermarking system developed by Google DeepMind and integrated into many generative AI tools. But unlike visible watermarks placed on images and videos for branded content, like Gemini adds to the corners of images, or Grok and Meta AI add to the corners of images, SynthID embeds imperceptible markers directly into the structure of AI-generated content. These are invisible to the human eye and do not affect quality in any way, but they can be detected by special tools designed to scan them. In fact, you can even use Gemini itself. All you have to do is go to Gemini, upload an image, and ask Gemini something like, “Was this image created by AI?” If so, Gemini is more likely to catch it.
SynthID’s primary purpose is a technical response to a large-scale problem on the Internet where it is often impossible to know whether media was created by AI or completely modified. Traditional watermarks are either visible and easily removed, or not visible at all. However, the SynthID method embeds the watermark itself during generation. In other words, watermarks are part of the digital creation of content and remain within the content even after common file modifications such as cropping, compression, and filtering.
Additionally, SynthID works with multiple types of media. Especially for images and videos, watermarks are distributed across pixels, so they don’t degrade image quality and remain detectable by analysis tools like Gemini. When it comes to text, SynthID uses subtle statistical symbols during the generation process to adjust the probabilities of words and tokens in ways that can later be recognized by detection systems, but the text can be easily modified. Additionally, audio content generated through AI models can undergo inaudible watermarks even after common transformations such as changing formats, adding noise, and compression, none of which are 100% foolproof, of course.
Moreover, this technology is already being applied on a large scale. Google reports that since its rollout, billions of pieces of content have been embedded with SynthID watermarks across its platforms, around 10 billion pieces of content to be exact. This includes images generated by Google’s models such as Gemini and hosted through services such as the Vertex AI suite. In addition to watermarks, Google has launched a specific SynthID Detector portal where users can upload media and check if it contains a SynthID watermark. The detector can also point out which parts of the content may be watermarked and pinpoint where AI involvement has occurred.
Why SynthID matters in the age of AI
This is clearly necessary, as synthetic content leads to both practical and ethical issues. On the one hand, generative AI is a powerful creative tool that improves productivity in many industries, and that’s great. But on the other hand, the ease with which realistic deepfakes, edited media, and fabricated news can be created poses real risks, as we have seen in many instances of hate-mongering. So, without an effective way to authenticate content, viewers can be misled by false visuals or misleading stories. Traditional digital literacy and skepticism no longer work.
SynthID addresses these issues by providing a scalable and automated way to tag synthetic content as it is created. Because watermarks are embedded during generation, they cannot be easily added or removed later without compromising the integrity of the content.
Importantly, SynthID is not a standalone solution to all AI reliability issues. That’s because it relies on creators using AI tools that support the watermark system in the first place. Content generated by systems that do not have SynthID embedded cannot be verified in this way. This means that content created with third-party AI tools and platforms will not include the SynthID watermark unless you choose to adopt that technology.
