C3AI: Crafting and Evaluating Constitutions for Constitutional AI Conference: Proceedings of the ACM Web Conference 2025; April 28-May 2, 2025; Sydney, NSW, Australia Proceedings of the ACM Web Conference 2025 (WWW ’25), April 28-May 2, 2025, Sydney, NSW, Australia DOI: 10.1145/3696410.3714705 ISBN:
Yara Kyrychenko, Ke Zhou, Edyta Bogucka, Daniele Quercia · 2025
Read C3AI: Crafting and Evaluating Constitutions for Constitutional AI Conference: Proceedings of the ACM Web Conference 2025; April 28-May 2, 2025; Sydney, NSW, Australia Proceedings of the ACM Web Conference 2025 (WWW ’25), April 28-May 2, 2025, Sydney, NSW, Australia DOI: 10.1145/3696410.3714705 ISBN: by Yara Kyrychenko, Ke Zhou, Edyta Bogucka, Daniele Quercia (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-15861.txt.
Description
Yara Kyrychenko, Ke Zhou, Edyta Bogucka, and Daniele Quercia published this paper in 2025. It studies constitutional AI, where a language model is steered by a written constitution of principles. Which principles work best has been guesswork; this framework, C3AI, makes the choice systematic.
The team first crafts constitutions, selecting principles and framing them well before training. Principles phrased positively and tied to concrete behavior matched human preferences best. They then evaluate whether trained models truly follow the constitution afterward. In a safety test case, a graph-based selection method improved an existing constitution's safety results while keeping reasoning strong. A surprise emerged: trained models obeyed negatively framed rules better than positive ones, the reverse of the human pattern. The authors flag this gap as the field's key problem.
Catalog record
| CREATOR | Yara Kyrychenko, Ke Zhou, Edyta Bogucka, Daniele Quercia |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 84 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-15861