Spymarks, not Watermarks
Invisible “spymarks” embedded in digital text and images are blurring the line between benign watermarks and covert tracking tools. Commenters weigh potential benefits like proving AI-generated content and deterring forgeries against serious risks to privacy, whistleblowers, and user control, noting how closed, proprietary systems (e.g., SynthID-style schemes, device-level detectors) could turn everyday media into a pervasive surveillance and DRM layer. Many argue that open standards, visible or verifiable marks, and local control are critical if provenance is to be improved without enabling ubiquitous tracking.
Terminology and definitions
- Many argue “spymark” is useful to distinguish tracking marks from traditional watermarks that signal authenticity or ownership.
- Others prefer terms like “tracking watermark” or see spymarks as just an application of steganography.
- Several note that classic banknote watermarks are visible and not meant to track users, so they shouldn’t be conflated with spymarks.
Privacy, tracking, and whistleblowers
- Strong concern that invisible, user‑specific marks can secretly track who created, viewed, or forwarded content, similar to email tracking pixels.
- Examples raised: identified whistleblowers via printer dots and company‑specific watermarked screeners or internal screenshots.
- Spymarks are seen as dangerous for whistleblowers, dissidents, and anyone targeted for views or affiliations.
Watermarks vs DRM and device control
- Some see these marks primarily as DRM tools, enabling enforcement and identification of “pirates.”
- A major fear is closing the “analog hole”: devices refusing to display or capture content if watermarks indicate restrictions.
- Others respond that the deeper problem is lack of user control over devices; spymarks are just one datapoint in that power imbalance.
Technical aspects: robustness and removal
- For text, marks can be encoded via word choice; copying text doesn’t necessarily remove them.
- For images, some suggest blurs, noise injection, recompression, or analog copying; others note modern schemes (e.g., SynthID) are surprisingly robust and hard to erase without heavy degradation.
- It’s acknowledged you can’t generally prove a watermark’s absence; only detect known marks.
SynthID and open alternatives
- Mixed views: some see SynthID‑style marking of AI output as a net positive for authenticity; others worry about closed, unverifiable, and potentially surveillance‑oriented designs.
- One commenter describes building an open, tamper‑resistant watermark for provenance, but is unsure whether releasing it would be perceived as protective or as enabling tracking.
Legal and policy questions
- Debate on whether user‑identifying spymarks fall under GDPR as personal data; some say yes if they uniquely identify users, others point to broad exemptions or argue they track content, not people.
- App‑store rejection of tools removing C2PA metadata is cited as evidence that removing such marks may be discouraged or blocked.