MENLO PARK, California: Meta's new artificial intelligence image detection tool struggles to identify some of its own AI-generated images after they are cropped, highlighting the challenges of verifying manipulated content as AI-generated images become more widespread.

The company previewed the tool this week alongside the launch of its new image-generation model, Muse Image, saying it can identify AI-generated images using an invisible watermarking system called Content Seal.

However, testing by Reuters found the tool successfully identified all 40 original images generated with Muse Image but failed to verify 55 percent of the same images after they were cropped to roughly one-third to one-half of their original size.

According to Meta, Content Seal is embedded in every image produced by Muse Image and is designed to remain intact even after common edits, allowing users to verify whether the company's AI models created an image.

Responding to Reuters' findings, Meta noted that the detection tool remains a preview. The company said the watermark is intended to survive common edits but acknowledged the signal may be lost if an image is heavily cropped.

The results underscore the broader challenge of authenticating AI-generated content after it has been modified, a limitation that could complicate efforts to identify deepfakes during a busy election year that includes the U.S. midterm elections.

Google and OpenAI have also cautioned that their own AI image detection tools are not foolproof against image-altering techniques.

In March, Meta's Oversight Board, an independent body that issues binding decisions and recommendations on content moderation across the company's platforms, urged the company to do more to address the "proliferation of deceptive AI-generated content" and invest in stronger detection tools.

Siwei Lyu, a computer science professor at the State University of New York at Buffalo who researches AI image forensics, said he had not evaluated Meta's system but noted that watermark-based detection has inherent limitations.

"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 remained a promising approach despite its shortcomings.

"Like many preventive cybersecurity or physical security measures, it may not be fully watertight, but even if we catch only 90 percent of cases, that's still a great leap from 0," she said.