AI Text Now Has a Secret Tracking Code
Summary
Taylor Lorenz and guest Neil Chilson discuss Anthropic’s global implementation of AI text watermarking to comply with the EU AI Act. They analyze how this technique alters word choices to embed hidden markers, raising concerns about reduced output quality, privacy risks, and potential government surveillance. The episode also explores the implications for free speech and the growing importance of open-weight models as an alternative to regulated systems.
Topics discussed
Sponsor reads: LinkedIn, Perfect Bistro, Visible, United, Starbucks
Intro: EU AI Act, Anthropic watermarking, and guest Neil Chilson
Why AI policy matters for civil liberties and marginalized groups
How AI text watermarking works and the EU AI Act Transparency Code
Technical details: Probabilistic word choice bias in LLMs
Sponsor reads: LinkedIn, Perfect Bistro, Visible
Issues with watermarking: Reduced quality and false positives
Host support plea: Patreon and Substack newsletter
Risks in legal, scientific, and code contexts due to precision loss
Security vulnerabilities and privacy concerns with secret keys
Potential for undetectable ideological bias in AI outputs
Critique of EU tech laws and comparison to GDPR cookie banners
Anthropic's worldwide implementation and First Amendment concerns
Diplomatic pressure and US response to EU regulations
Open weight models as a liberty safeguard and episode conclusion
Listen ad-free on Castria