Microsoft presented the results of a large-scale study on detecting AI-generated and altered digital content. The artificial intelligence security team tested more than 60 combinations of methods, from digital watermarks and enhanced metadata to “mathematical fingerprints” that can establish a file’s source. The goal is to create a technical foundation for transparent disclosure of where materials come from.
The study showed that there is no universal method of protection. Some methods remain robust even against attempts to hide AI origins through editing or masking, while others may work incorrectly if used improperly. Microsoft proposed optimal standards suitable for implementation by major platforms.
The work became especially relevant in connection with California’s AI Transparency Act, which will take effect in the summer. The company took part in discussions of the requirements, but has not yet announced full implementation of the standards in its services, including Copilot, Azure and LinkedIn.
Microsoft emphasizes that the technology’s task is not to determine whether information is true, but to record the fact of manipulation and the origin of content. If the industry can agree on common standards, the digital environment may become more predictable and resilient. Not perfectly honest, but more understandable. And that is already a serious step forward.
