Deepfakes are currently one of the most critical issues we are facing in the world. AI technologies have today made it easier for anyone to create manipulated video, audio and images.
Deepfake tools can swap faces of people in seconds and show events that never happened in real life. Thousands of celebrities, politicians and even ordinary people have faced this situation. The scariest part is that this content can look so realistic that it is difficult to detect.
In such scenarios, AI watermarks can help identify AI-generated content. It can provide information about the origin of a video and help determine if the content was created using AI. However, AI watermarks alone are not enough.
Metadata can be lost when files are edited, converted into another format, or shared across different platforms. Even screenshots can break the link to the original content. Therefore, AI watermarks won’t stop deepfakes unless social media platforms and editing tools themselves can stop breaking them.
ALSO READ: Top Deepfake Statistics 2025
Content Credentials and AI Watermarks Have Their Limits
Digital or AI-generated content can be primarily detected through two methods: Content Credentials and Invisible Watermarks. Both methods have different ways to identify digital content.
Content Credentials uses the Coalition for Content Provenance and Authenticity (C2PA) standard. Its advanced authenticity detection tools can record information regarding where the content originally came from and its editing history.
Invisible Watermarks, on the other hand, work differently. This method adds an invisible signal directly into the digital media, such as an image, audio clip or video clip. Based on the technology and the edits made to the file, the invisible signal may still remain detectable when the metadata disappears.
Both Content Credentials and Invisible Watermarks are used in OpenAI for supported images created using its tools. Regardless of using both methods, OpenAI has acknowledged that neither of the methods is 100% reliable.
Metadata is capable of getting lost during file handling or too much editing. This can change or weaken the detectable watermark in the content. Therefore, these methods cannot guarantee a perfect solution for deepfake content detection. Instead, they can be considered a useful method to identify the content but are not completely reliable.
ALSO READ: OpenAI Introduces textGrain Watermarks for ChatGPT and Codex in the EU
Why Platforms Should Retain Watermark Information
AI watermarks alone are not enough. Platforms must also preserve watermark information when content is uploaded and shared. In order to truly identify deepfakes, the platform needs to build systems that can preserve information that shows where the files came from.
Because when a deepfake image is created, it may have information about the origin or watermarks of the AI generator. However, such information can be easily dropped with each image shared.
Like, a user can download a deepfake image, upload it to different social media platforms, a third person can take a screenshot and share it in different groups. The metadata will disappear with each step. This makes the watermark extremely difficult to detect, which will eventually make it harder to recognize the original source of the image.
This makes it important for platforms to create more effective systems that can keep the metadata protected regardless of the sharing stages.
For example, social media platforms could keep valid Content Credentials when they edit or resize images. They could also check for watermarks and show the results to users. If no watermark is found, the platform should not assume that the content is real or authentic.
However, these steps might not help identify each and every deepfake. But it can still help journalists and fact-checkers identify suspicious content more easily.
How Common Standards Can Improve Deepfake Detection
One of the major issues with AI watermarks and deepfake detection tools is that they do not always use the same platform or follow the same signals. A verification system can successfully detect a watermark on a deepfake image created by a company. However, it may not recognize content created by a completely different AI tool. These limitations can be a huge problem when it comes to identifying deepfake content.
This is why sharing the same standard can help solve this problem. C2PA provides a common way to record where digital content came from and how it has changed. If more companies adopt this standard, different platforms may find it easier to recognise deepfake content.
But just adopting a standard is not enough. Companies need to make sure the systems work properly. They need to focus on what happens to the metadata when files are uploaded to a different platform. Along with the changes when an image is downloaded, edited or converted to another format.
Platforms should focus on these areas, test these processes and explain which signals they can detect. Additionally, they should also provide users with simple methods to check where content came from without any heavy technical knowledge.
The purpose is to make sure information about a deepfake content origin can travel across different platforms. Instead of the information being available only on the platform where the content was originally created.
AI Watermarks Alone Cannot Solve the Deepfake Problem
AI watermarks can be a useful way to identify deepfake content. However, the usefulness depends on the technology itself. AI watermarks should remain intact regardless of the editing and how many times the file is shared to be truly effective. Some content might not have any watermarks at all. AI developers need to work on detection methods to ensure watermarks are detected regardless of which platform was used.
Apart from watermarks and detection tools. Social media and editing platforms should also take responsibility for preserving information about a piece of content’s origin. It’s important to know that not a single method can stop all deepfakes. But AI watermarks and detection tools can certainly help identify manipulated content and take necessary action against.
Right now, the main challenge is not just to add a watermark to an AI-generated file. Instead, it’s to make sure that particular watermark remains in the file after the file has been shared. Until this issue is resolved, watermarks will remain useful but will not be enough to solve the deepfake problem.
