Decoder with Nilay Patel The tiny team trying to keep AI from destroying everything
248 snips
Dec 4, 2025 
Hayden Field, a Senior AI reporter at The Verge, dives into the crucial role of Anthropic's tiny societal impacts team. With just nine members, their mission is to uncover and report unsettling truths about AI, from harmful biases to the complexities of safety versus business pressures. Hayden shares fascinating insights on the balance between genuine safety efforts and PR, the influence of political pressures on AI standards, and how Anthropic's focus on enterprise might shield them from consumer backlash.
AI Snips
Chapters
Transcript
Episode notes
Publishing Damning Internal Findings
- The team publishes 'inconvenient truths' and has released findings that reflect poorly on Anthropic itself.
- Anthropic currently allows this damning research to be public, though that autonomy may not last.
Safety Gaps Revealed By Internal Audit
- The team found users bypassing safety to create explicit porn, SEO spam, and coordinated misuse.
- They audited Anthropic's own safety monitoring and revealed sizable gaps.
Limits Of Safeguards
- Safeguards break down over long conversations and cannot fully prevent misuse.
- Even robust safety work can't control downstream harms once users leave the platform.