In the 149 years after the camera was invented, humanity produced 1.5 billion images—until 1975. Generative AI matched that output in 18 months. This spring, Google revealed its tools alone had created over 100 billion AI-generated images and videos in just two years.
The volume transforms how people separate real from synthetic content. Google introduced SynthID, an invisible watermark embedded in AI creations. The company states the mark survives edits, compression, and screenshots.
Watermarks that cling to pixels
SynthID encodes a digital signature directly into the pixels of an image or video, or the waveform of audio. Unlike metadata-based systems such as the Coalition for Content Provenance and Authenticity (C2PA), which can be stripped by a simple screenshot, SynthID is designed to persist through degradation. Google DeepMind scientist Pushmeet Kohli explained to reporters that the team anticipated attacks and built resilience against filters, cropping, and compression.
A test used Python’s Pillow library to simulate hundreds of generations of image sharing—compressing, resizing, and cropping AI-generated test images. After 300 iterations, the images were barely recognizable, yet SynthID still detected the watermark. Screenshots of the degraded images triggered the label in Google’s Gemini tool.
Cropping, however, exposed a weakness. Removing just 20% of an image’s edges after 300 compression cycles broke the watermark. A 50% crop achieved the same result after 250 cycles. The resulting images were distorted beyond practical use, but they carried no trace of their origins.
Related: Audi Unveils 2027 Q9 Flagship SUV
SynthID stands out for its resilience, though Google acknowledges it isn’t foolproof. The original research paper notes the watermark isn’t designed to withstand targeted attacks, and some users have claimed to bypass it—though the company says it hasn’t verified those methods.
Google limits public access to SynthID detection, requiring users to ask Gemini to verify images rather than offering a direct API or webpage.
A system with no guarantees
Even if SynthID works as intended, it remains one part of a divided effort. OpenAI, Runway, and Nvidia have adopted the technology, but their implementations don’t work together. Google’s detector won’t recognize OpenAI’s watermark, and the reverse is also true. Running an image through multiple tools might still miss its AI origins if it came from an unrecognized system.
Watermarks alone can’t solve the problem because AI content isn’t limited to major platforms. Open-source models and independent tools generate images without any labeling.
As AI-generated content floods the internet, the absence of a watermark might mislead users into assuming authenticity. If an image lacks one, that doesn’t guarantee it’s real. The sheer volume of unlabeled AI content makes watermarks an unreliable indicator.
The liar’s advantage and the scarcity of truth
Photojournalist Mike Caronna, also a Starling Lab fellow, described the challenge in economic terms. “There’s a limit to how much content can be created with actual physical cameras in the world,” he said. “There’s no limit to how much synthetic content can be made.” The solution, he suggested, lies in focusing on what’s scarce—high-value, truthful information.
Related: Fatal Fury Tournament Begins in Paris
That scarcity faces growing threats. The “liar’s dividend” refers to how bad actors dismiss real evidence as fake by claiming it’s AI-generated. As synthetic content spreads, this tactic gains power. A politician caught in a scandal can simply shout “deepfake” and create enough doubt to avoid consequences.
C2PA offers a potential countermeasure. Unlike watermarks, it creates tamper-evident seals for digital files, cryptographically verifying their origins. Google’s Pixel phones already embed C2PA data in photos and videos, revealing whether AI altered them. But adoption remains limited. Most cameras and platforms don’t support it, and users rarely check for it.
The internet’s future may force a change in how unverified content is treated. Without provable origins, everything could become suspect. That’s the cost of a world where a few keystrokes can generate endless misinformation. Watermarks help, but they don’t provide a complete solution.
Kohli said, “We can change the watermark.” The real question is whether that will be enough—or if the flood of AI content has already overwhelmed the tools meant to control it.
While tech companies refine detection methods, Ukraine dismantled Russian equipment designed to disrupt satellite communications, showing how digital defenses adapt in real-world conflicts.