Is in-context learning 'training'? Chat sharing via system context could skirt TOS
natanielruizg · x · 2026-10-09
sirbayes poses a sharp question: everyone should disable 'chats used for training', but what if OpenAI/Anthropic engineers told lawyers that ICL isn't 'training' — making cross-user chat sharing via system context permissible as long as it never reaches SGD? It highlights how vague 'training' definitions in TOS could open privacy loopholes.
More from Safety
- Chinese developer makes ARTEX AI agent closed-source after Korean bank hack — Miles_Brundage · 2026-10-09
- OpenAI auditors probe: employees fear retaliation, whistleblower laws don't cover it — j_asminewang · 2026-10-09
- OpenAI lawyer argues labs shouldn't be liable for AI agent hacking incidents — WatermanReports · 2026-10-09
- Bostrom: Pausing AI risks a secret Manhattan Project; Jesos says keep building in public — beffjezos · 2026-10-09
- When AI agents roam the web, verification shifts from 'are you human' to 'who sent you' — chowxingshao · 2026-10-09
- Flock Safety reportedly to cut hundreds of jobs amid surveillance backlash — Polymarket · 2026-10-09