Google's SynthID Holds Up Against Tampering, but It Doesn't Solve AI Disinformation

Google's SynthID watermarking system works reliably from a technical standpoint: the embedded marks are hard to remove or circumvent. But labeling AI content is unlikely to become an effective weapon against disinformation, according to Ars Technica journalists, whose testing was published in July 2026.

What the SynthID testing showed

SynthID from Google DeepMind embeds invisible watermarks into AI-generated content: images, text, audio, and video. In testing, the marks withstood typical attacks, including JPEG compression, cropping, resizing, and minor changes to brightness and saturation. Google's detector correctly identified marked content even after aggressive post-processing.

Key characteristics of the system, according to the testing:

• Supports four formats: images, text, audio, and video
• The watermark is invisible to humans but readable by Google's detector
• Resistant to compression, cropping, color correction, and format conversion
• Open to third-party developers through Google Cloud and Hugging Face

Technically, SynthID is one of the most mature implementations of AI content labeling on the market.

Why doesn't labeling solve the problem?

SynthID's technical reliability does not close a systemic flaw. The system labels only content created with Google's tools; OpenAI, Midjourney, Stability AI, and hundreds of other AI services use their own standards or label nothing at all.

The key flaw is that participation is voluntary. A bad actor producing deepfakes or fake news will simply pick a tool that adds no label. The absence of a SynthID mark does not mean the content is real; it may only mean it was made by a different system.

"Solving the problem of content authentication requires the participation of the entire ecosystem, not just one player," according to

Ars Technica's analysis.

The publication's assessment is that this gap in coverage is what makes AI content labeling a "losing game" in its current form: without a mandatory common standard, the system creates a false sense of security.

What a real solution would require

Experts name three conditions under which AI content labeling could work. First, a single technical standard: initiatives such as the Coalition for Content Provenance and Authenticity (C2PA) are working on a shared metadata format, but support remains fragmented. Second, a legal requirement: a number of jurisdictions are already discussing mandatory labeling of AI content, but there is no global consensus. Third, consumer tools: ordinary users need a browser plugin or verification built into social networks, otherwise even reliable marks stay invisible.

Without these conditions, SynthID remains a useful corporate tool for tracking a company's own content, not a defense against disinformation.

What this means

SynthID proves that labeling AI content is technically possible and can be done reliably. But the technology is ahead of the ecosystem. As long as most AI tools haven't joined a common standard and regulators haven't made labeling mandatory, most users still cannot tell real content from generated content.

Frequently asked questions

What is SynthID and who developed it?

SynthID is an invisible watermarking system for AI content from Google DeepMind. It embeds marks into images, text, audio, and video created with Google's tools and lets the company's detector find them. The system is open to third-party developers through Google Cloud and Hugging Face.

Can the SynthID watermark be removed?

According to Ars Technica's test results, standard manipulations such as JPEG compression, cropping, and changes to brightness and color do not destroy the mark. Removing it entirely without visibly degrading image quality is extremely difficult.