Anthropic replaces the 2023 constitution with a new one
Anthropic published a new constitution for Claude, released under a CC0 license, and the 2023 post carries an update note dated 2026-01-21 (the new-constitution page itself is dated 2026-01-22).
- Date
- 21 January 2026
- Who
- Anthropic
- Confidence
- High (Anthropic's own posts; I have not read the full document line by line)
- Deep dive
- RLHF and instruction tuning (how base models became assistants)
Tier: Supporting · Significance: 3/5 · Org(s): Anthropic · Confidence: High (Anthropic's own posts; I have not read the full document line by line) Anthropic published a new constitution for Claude, released under a CC0 license, and the 2023 post carries an update note dated 2026-01-21 (the new-constitution page itself is dated 2026-01-22). Where the 2023 list gave standalone principles (B05-32c), the new text explains why behaviors matter, on the stated theory that models generalize better when they understand reasons than when they follow rules. It ranks four properties in order (broadly safe, broadly ethical, compliant with Anthropic's guidelines, and "genuinely helpful"), with hard constraints for behaviors Claude should never perform. Anthropic also says the constitution is itself used to generate synthetic training data, including conversations, value-aligned responses and rankings of possible responses, building on Constitutional AI (Anthropic, new constitution; Anthropic, 2023 post with update note). It describes itself as a living work in progress and acknowledges possible gaps between intent and actual model behavior. It belongs in the CAI lineage (B05-21) because the human input to alignment moved from a short principle list to a long explanatory document that doubles as training data, 162 weeks after the paper. Depth is in B06. Sources: Anthropic, new constitution · Anthropic, 2023 constitution