Sensity says watermarks won’t stop AI deepfakes for police
Sensity AI emphasizes that digital watermarking and content provenance metadata, like those from OpenAI and Google's SynthID, are insufficient on their own for law enforcement in identifying AI-generated content. The article argues that these 'opt-in' solutions can be easily bypassed by malicious actors and calls for a multi-layered approach incorporating specialist forensic tools for robust media verification.
Key Takeaways
- OpenAI says C2PA metadata can be stripped, lost through uploads, or broken by file format changes, resizing, and screenshots.
- Google expanded SynthID detection from Gemini to Lens, AI Mode, Circle to Search, and Gemini in Chrome.
- Sensity says Stable Diffusion with 10-15% denoising strength can remove a watermark, and a GitHub repo already exists to reverse SynthID.
- The article says metadata is often degraded when social platforms copy audio and video into a new container during upload.
- Sensity argues that law enforcement needs complementary forensic tools and an auditable chain of custody for court authentication.
Why It Matters
The immediate takeaway is that watermarking and provenance metadata are useful triage signals, but not enough to authenticate AI-generated media in court. Sensity frames the competitive and ecosystem issue as a trust problem: C2PA and SynthID depend on voluntary participation, so they mainly help conforming creators and mainstream platforms, not malicious actors who strip labels or use non-labeling generators. The key signal to watch is whether law-enforcement workflows start pairing watermark checks with specialist forensic tools at evidence ingestion, as Sensity recommends.
Read full article at sensity.ai
