The Scale of Synthetic Media
The rapid expansion of artificial intelligence generated media poses significant challenges for digital information systems. According to an analysis by Starling Lab reported by Ars Technica, generative AI tools produced approximately 1.5 billion images in an 18-month period. For comparison, Starling Lab noted that traditional photography required 149 years—from its inception in the early 19th century until 1975—to reach the same milestone.
This volume has continued to grow. At its spring 2026 I/O event, Google reported that over 100 billion images had been generated using its tools. To address the resulting challenges around synthetic content and misinformation, tech companies are developing labeling systems. Google has also backed European Union rules regarding AI content transparency and watermarking frameworks.
Testing SynthID and Watermarking Robustness
One prominent technical approach is Google's SynthID, which embeds an invisible watermark directly into AI-generated media. While designed to resist modification, practical testing indicates limits to its durability.
In stress tests utilizing Python's Pillow library, the SynthID watermark demonstrated resilience through multiple cycles of compression and saving. However, researchers identified critical vulnerabilities when compression was combined with cropping:
Post-Compression Cropping: SynthID proved vulnerable to cropping after undergoing compression, making it harder to detect.
Crop Vulnerability: When compression cycles were paired with cropping, the watermark detection rate fell significantly or failed.
This limitation is not unique to Google. For instance, Reuters reported that Meta's Content Seal faces similar cropping issues, showing that minor image alterations can disrupt detection across different platforms.
Access and Interoperability Limits
Even when watermarks remain intact, verifying them presents practical hurdles. Currently, Google's detection tools face access constraints, such as a daily limit of 10 checks per user on Gemini.
Furthermore, there is a lack of interoperability between major tech platforms. Currently, Google and OpenAI detectors do not share a common standard, meaning Google's detector cannot read OpenAI's watermarks, and vice versa. This lack of cooperative infrastructure limits the effectiveness of automated platform-wide detection.
The Challenge of Open Models and Verification
A deeper structural limitation is the availability of open-source models. Industry experts note that centralized watermarking cannot address content generated by local, offline open-source models that do not enforce watermarking frameworks.
As Adam Rose observed, relying solely on centralized creators to apply watermarks leaves a significant gap, as open models can generate unlabeled content freely. Mike Caronna similarly noted that as truth becomes scarcer, verifying authentic content becomes essential to combat claims that genuine footage is fabricated.
To establish a verifiable record of real-world captures, some hardware manufacturers are integrating authentication standards directly into devices. For example, Google's Pixel 10 integrates C2PA standards, allowing cryptographic verification of an image's origin. However, until these systems are widely adopted and standardized, the digital ecosystem remains highly fragmented.