Durmus et al. — Towards Measuring the Representation of Subjective Global Opinions in Language Models (2023)

2 min read

Towards Measuring the Representation of Subjective Global Opinions in Language Models

Authors: Esin Durmus, Karina Nguyen, Thomas I. Liao, Nicholas Schiefer et al. Affiliation: Anthropic Canonical URL: https://arxiv.org/abs/2306.16388

Summary

A quantitative framework for measuring whose opinions an LLM’s responses are more similar to. The authors build GlobalOpinionQA from cross-national surveys (Pew, World Values Survey) and measure similarity between LLM responses and human responses per country.

Headline finding: by default, LLM responses are most similar to the opinions of certain populations — primarily the USA, parts of Europe, and parts of South America. This is a direct quantification of mirror imaging at the cultural / values level.

Key Findings

  1. Default LLM opinions are WEIRD-biased. Responses align most with Western, Educated, Industrialized, Rich, Democratic populations — primarily USA opinions.
  2. Country-conditioning helps but distorts. When prompted to consider a specific country’s perspective, the model shifts — but introduces stereotypes and may caricature the target population.
  3. Translation does not fix it. Translating into the language of the target country does not meaningfully shift opinion alignment toward that country.
  4. Implications for agentic systems modeling foreign actors. An LLM agent asked to model an adversary’s reasoning will, by default, model a USA-flavored adversary.

Relevance to This Wiki

  • Direct empirical foundation for Mirror Imaging as an LLM phenomenon. Quantifies what was previously a qualitative claim.
  • Critical for H4 (Red Team produces adversarially robust plans). If a red-teaming LLM has WEIRD priors and can’t reliably shift them via country-prompting, its adversarial modeling will be biased toward Western threat archetypes. This is a specific failure mode for H4 testing.
  • Caution for Red Team Analysis in LLM contexts. Persona prompting works but unevenly; check that adversary models aren’t sanitized versions of Western actors.
  • Caution for Outside-In Thinking. “Considering external context” is undermined if the model’s default external context is also WEIRD.

See Also

  • Mirror Imaging
  • Red Team Analysis
  • Atari et al. (2023) Which Humans? — the WEIRD critique
  • Cao et al. (2023) Assessing cross-cultural alignment between ChatGPT and human societies

Graph View

Backlinks