Erik Voorhees slams Anthropic's Claude self-assessment: machine chats create no victim
juanbenet · x · 2026-11-01
Anthropic introduced "self-assessment": after each response, Claude flags potential misuse it was designed to catch via other safeguards; live since June and more scalable than fixed-harm classifiers. ShapeShift founder Erik Voorhees pushed back hard: text with a machine is inherently not criminal or immoral—blame belongs to the actual harmful act (theft, hacking, real threats), not the messages. "Imagining harm ≠ harm." The exchange hits the core controversy over vendors monitoring user conversations.
More from AGI Musings
- Can LLMs turn surprising observations into research questions and discover physical laws? — rohanpaul_ai · 2026-11-01
- Gamers push back: 'We want MORE AI in gaming, not less' — petewoodbridge · 2026-11-01
- 'Don't recreate existing games': AI platform terms spark open-asset fears — TinfoilTricorn · 2026-11-01
- Altman prefers trillions of AIs checking each other; skeptics ask who owns them — VraserX · 2026-11-01
- RL Legend Szepesvári Defends Academia: Your Research Raises the Water Level — CsabaSzepesvari · 2026-11-01
- Sam Altman: Using latest AI tools feels like managing a factory inside a computer — haider1 · 2026-11-01