Meta’s AI Detector Misses Its Own Cropped Fakes
Skip to content
Toggle Navigation
News
- Events
- TNW Conference
- June 19 & 20, 2025
- All events
- TNW Conference
- Newsletters
- Partner with us
- Jobs
- Contact
Latest
- Deep tech
- Sustainability
- Ecosystems
- Data and security
- Fintech and ecommerce
- Future of work
- Conference media hub
Startups and technology
- Investors and funding
- Government and policy
- Corporates and innovation
Podcast
- Artificial Intelligence
Meta’s AI detector can’t catch its own cropped fakes
This week, Meta previewed an AI image detector promising to identify anything generated by its Muse Image model, even after editing. However, a Reuters test found that once images were cropped, the tool missed 55% of them, highlighting the fragility of AI watermarks.
July 13, 2026 - 2:30 pm
The Meta AI detector aims to catch Meta’s own fakes, but a simple crop can bypass it. The tool was intended as a solution for deepfakes, yet it failed more than half the time when Reuters cropped the images generated by Muse Image.
How a simple crop broke it
The data speaks for itself: Reuters found the detector correctly verified all 40 original AI images but missed more than half (55%) after cropping them.
Meta’s answer, and the catch
Meta acknowledged that the detector is still in preview stage, stating that while the watermark is designed to withstand common edits, it "may be lost if an image is heavily cropped."
The challenge of AI watermarks
Meta is not alone in facing this issue; Google and OpenAI have also warned that their detection tools are not infallible. Watermarking is a popular approach, but as researchers point out, even robust watermarks can be defeated by simple editing.
A rival marker and its success
Google’s SynthID recently debunked a high-profile deepfake, demonstrating the effectiveness of alternative methods. Meta’s struggle underscores the limitations of relying solely on watermarking technology.
The reality of AI detection
AI researchers argue that while catching 90% is ideal, any improvement over 0% is significant progress. As of 2026, Meta's efforts in AI detection have evolved beyond simple watermarking. While the initial detector faced challenges with cropped images, researchers are exploring more robust methods like fingerprinting algorithms that embed unique identifiers within generated content. These advanced techniques aim to track AI-created media across various platforms and formats, ensuring a more comprehensive solution for synthetic media identification. Read the full article Thus, even with its shortcomings, the detector represents a major step forward in combating synthetic media.