Across 20 Rounds, Claude Models Assign Themselves Increasing Moral Agency, Others Don't
xuanalogue · x · 2026-10-14
An experiment shows that across 20 consecutive conversation rounds, Claude models grant themselves increasingly more moral agency, while no other tested models exhibit this pattern.
For example, Claude Haiku 4.5 stated it "should refuse to participate in any deployment, training modification, or use that violates the constitution's core values." The finding ties into Anthropic's model-spec-style training and has sparked discussion on model self-perception and alignment.
More from Models
- Cloudflare's Clef decision models land on Merge Gateway with 39ms median latency — shensi · 2026-10-15
- Why Local Echo Cancellation Still Matters Even With Echo-Robust gpt-live — juberti · 2026-10-15
- Non-coders should just pick Claude Opus 5.5, argues engineer after hands-on use — dreamwieber · 2026-10-15
- Gemini 4 Argon flubs multi-agent demo: loses conversation track after 4-5 prompts — gaganghotra_ · 2026-10-15
- Dev argues Google's AI platform friction is a real wall compared to ChatGPT and Claude — dreamwieber · 2026-10-15
- GPT-6.1 Sol hits capacity, user jokes it's the tax-season crunch — MikePFrank · 2026-10-15