The development has renewed concerns about the difficulty of verifying AI-generated images, particularly as digital manipulation tools become more advanced and online platforms face increasing challenges in detecting deepfakes.
The issue is especially significant during major election periods, when misleading AI-generated content can be used to spread false information and influence public opinion.
A Reuters analysis found that Meta’s detection tool successfully verified 40 images generated using the company’s new image-generation model, Muse Image, in their original form. However, when the same images were cropped to between one-third and one-half of their original size, the tool failed to verify 55% of them.
Meta had previously stated that its preview detection system was capable of identifying AI-generated images created by Muse Image even after cropping. The system relies on an invisible watermarking technology known as Content Seal, which is embedded into every image produced by Meta’s AI models to help determine whether content was generated using the company’s technology.
Following the Reuters findings, Meta said the tool was still in preview and explained that while the watermark was designed to survive common image edits, more extensive modifications could affect its effectiveness.
The company said the watermark signal may disappear when an image undergoes significant cropping or other major alterations.
The challenge is not unique to Meta. Other major technology companies, including Google and OpenAI, have also acknowledged that AI detection systems cannot guarantee perfect identification of manipulated images, particularly after editing techniques are applied.
Earlier this year, Meta’s Oversight Board urged the company to take stronger action against the growing spread of deceptive AI-generated content on its platforms and invest in more effective detection measures.
Experts warn of watermarking limitations
Researchers in digital forensics have pointed out that watermark-based identification systems can be useful but remain vulnerable to certain forms of modification.
Siwei Lyu, a computer science professor at the State University of New York at Buffalo who studies AI image forensics, said watermarking technology can work effectively when the embedded signal remains intact but may become weaker after image manipulation.
"Watermark-based methods can be highly effective when the watermark remains intact, but any modification that removes or weakens the embedded signal — such as cropping, resizing, heavy compression, or editing — may reduce their effectiveness, depending on how the watermark is designed," Lyu said.
Sarah Barrington, an AI researcher and Ph.D. candidate at the UC Berkeley School of Information, said watermarking remains a promising approach for identifying AI-generated content, even though it may not completely eliminate the problem.
"Like many preventive cybersecurity or physical security measures, it may not be fully watertight, but even if we catch only 90% of cases, that’s still a great leap from 0," she said.
As generative AI tools continue to improve, technology companies are under growing pressure to develop reliable methods for distinguishing authentic images from synthetic ones. The Meta case highlights the ongoing arms race between AI content creation and the systems designed to verify it.