Meta AI detection tool struggles to recognize some cropped AI generated images

Meta’s new AI image detection tool has shown limitations after failing to identify some of its own AI generated images once they were cropped, raising fresh concerns about the reliability of technology designed to detect manipulated content online.

The tool was introduced this week alongside Meta’s latest image generation model, Muse Image. It is intended to help users determine whether an image was created using the company’s artificial intelligence system. However, an analysis conducted by Reuters found that the detector became far less reliable after images were trimmed, exposing a weakness that experts say could make it easier for altered AI content to spread across social media.

Reuters tested 40 images produced with Muse Image and found that the detection tool successfully recognized every original image as AI generated. However, after those same images were cropped to roughly one third to one half of their original size, the tool failed to identify 55 percent of them.

Meta says the detection system works through an invisible watermark known as Content Seal, which is embedded into every image created by Muse Image. According to the company, the watermark is designed to remain detectable even after common edits, including cropping. The goal is to allow users to verify whether an image was generated by Meta’s AI models.

After Reuters shared its findings, Meta pointed out that the detector is still in its preview stage. The company acknowledged that while the watermark is built to survive routine edits, the embedded signal can be weakened or completely lost if an image is heavily cropped.

The issue highlights a broader challenge facing the technology industry as AI generated content becomes more realistic and more difficult to distinguish from genuine images. Experts have warned that detection systems are still evolving and cannot guarantee accurate results in every situation, especially when images are altered before being shared online.

Other major technology companies have also admitted that their own detection tools are not perfect. Google and OpenAI have both said that changes such as cropping, resizing, compression, or other editing techniques can reduce the effectiveness of systems designed to identify AI generated images.

Earlier this year, Meta’s Oversight Board urged the company to strengthen its efforts against misleading AI generated content across its social media platforms. The independent body recommended that Meta invest in more advanced detection technology to better address the growing spread of deceptive digital media.

Artificial intelligence researchers say watermark based detection has clear advantages but also important weaknesses. Siwei Lyu, a computer science professor at the State University of New York at Buffalo, explained that watermark systems perform well only when the embedded signal remains intact. He noted that editing methods such as cropping, resizing, compression, or other modifications can reduce their ability to verify whether an image was produced by AI.

Sarah Barrington, an AI researcher and doctoral candidate at the University of California Berkeley, believes watermarking remains a valuable tool despite its limitations. She said no security system is perfect, but even identifying the majority of AI generated content would represent significant progress compared with having no reliable method at all.

As AI generated images become increasingly common, improving detection technology is expected to remain a major priority for technology companies, researchers, and policymakers working to limit the spread of misleading digital content.