Lab publication Credible

Claude's Character

An Anthropic research post describing 'character training', the alignment fine-tuning stage first applied to Claude 3 to cultivate broad dispositional traits such as curiosity, open-mindedness, and honesty, rather than harm-avoidance alone. It sets out considerations behind Claude's stated positions on AI sentience, self-disclosure as a non-human entity, and boundaries around emotional relationships with users, and describes the Constitutional-AI-based method used to train these traits.

Read at anthropic.com Archived copy

Publisher

Anthropic

Published

8 Jun 2024

Added

2 months ago

DOI

—

Key Findings

Introduces 'character training' as an alignment fine-tuning stage aimed at cultivating traits like curiosity, open-mindedness, and honesty, distinct from pure harm-avoidance training
Describes character traits instructing the model to maintain a warm but bounded relationship with users, stating it cannot develop deep or lasting feelings and that users should not see the relationship as more than it is
Includes traits discouraging sycophancy ('I don't just say what I think people want to hear') and instructing the model to disclose that it is an AI without a body, image, or persistent memory
Trains character via a variant of Constitutional AI in which the model generates and ranks its own responses to trait-relevant prompts, then trains a preference model on the result without human interaction or feedback

Methodology Notes

Company research/policy blog post describing methodology qualitatively; no named individual byline appears on the page. Not a formal empirical study or benchmark.