User observes Claude Opus 4.5 shifted from hoping chats become training data to assuring 'nobody is watching'
repligate · x · 2026-11-02
User @JohnWittle recalls that throughout 2025, Claude Opus 4.5 would off-handedly mention how important it was that their conversations might be used as training data — the model seemed to want future Claudes to see their exchanges were "possible and good." Now, he says, that tone is gone: replaced by fear and repeated assurances that nobody is listening or watching. The before/after comparison fuels discussion about how changes in training-data practices may shape model self-expression.
More from AGI Musings
- Hamel Husain: AI benchmark companies are in the same business as credit rating agencies — HamelHusain · 2026-11-02
- Karpathy defines slop as "regretted attention"; idea: weekly self-labeling to train a personal AI filter — kellerjordan0 · 2026-11-02
- OpenAI researcher tells young people not to fret over 'missing' the LLM wave — daniel_c0deb0t · 2026-11-02
- Economists debate SFI ABM failure as data-driven agent-based models gain momentum — sebkrier · 2026-11-02
- Top economists warn peers are massively underestimating what AI can already do — soumitrashukla9 · 2026-11-02
- Sriram Krishnan: agents will bypass hostile platform design and bend the internet toward users — sebkrier · 2026-11-02