[2603.28123] Does Claude's Constitution Have a Culture?

[2603.28123] Does Claude's Constitution Have a Culture?

arXiv - AI 4 min read

About this article

Abstract page for arXiv paper 2603.28123: Does Claude's Constitution Have a Culture?

Computer Science > Computers and Society arXiv:2603.28123 (cs) [Submitted on 30 Mar 2026] Title:Does Claude's Constitution Have a Culture? Authors:Parham Pourdavood View a PDF of the paper titled Does Claude's Constitution Have a Culture?, by Parham Pourdavood View PDF HTML (experimental) Abstract:Constitutional AI (CAI) aligns language models with explicitly stated normative principles, offering a transparent alternative to implicit alignment through human feedback alone. However, because constitutions are authored by specific groups of people, the resulting models may reflect particular cultural perspectives. We investigate this question by evaluating Anthropic's Claude Sonnet on 55 World Values Survey items, selected for high cross-cultural variance across six value domains and administered as both direct survey questions and naturalistic advice-seeking scenarios. Comparing Claude's responses to country-level data from 90 nations, we find that Claude's value profile most closely resembles those of Northern European and Anglophone countries, but on a majority of items extends beyond the range of all surveyed populations. When users provide cultural context, Claude adjusts its rhetorical framing but not its substantive value positions, with effect sizes indistinguishable from zero across all twelve tested countries. An ablation removing the system prompt increases refusals but does not alter the values expressed when responses are given, and replication on a smaller model...

Originally published on March 31, 2026. Curated by AI News.

Related Articles

Llms

My AI spent last night modifying its own codebase

I've been working on a local AI system called Apis that runs completely offline through Ollama. During a background run, Apis identified ...

Reddit - Artificial Intelligence · 1 min ·
Llms

Fake users generated by AI can't simulate humans — review of 182 research papers. Your thoughts?

https://www.researchsquare.com/article/rs-9057643/v1 There’s a massive trend right now where tech companies, businesses, even researchers...

Reddit - Artificial Intelligence · 1 min ·
Llms

Depth-first pruning seems to transfer from GPT-2 to Llama (unexpectedly well)

TL;DR: Removing the right transformer layers (instead of shrinking all layers) gives smaller, faster models with minimal quality loss — a...

Reddit - Artificial Intelligence · 1 min ·
[2603.23966] Policy-Guided Threat Hunting: An LLM enabled Framework with Splunk SOC Triage
Llms

[2603.23966] Policy-Guided Threat Hunting: An LLM enabled Framework with Splunk SOC Triage

Abstract page for arXiv paper 2603.23966: Policy-Guided Threat Hunting: An LLM enabled Framework with Splunk SOC Triage

arXiv - AI · 4 min ·
More in Llms: This Week Guide Trending

No comments

No comments yet. Be the first to comment!

Stay updated with AI News

Get the latest news, tools, and insights delivered to your inbox.

Daily or weekly digest • Unsubscribe anytime