AI Language Output Quality Research

Research on how input/output language combinations affect AI (Claude) reasoning quality, detail coverage, and thinking patterns.

Date: 2026-03-03 Model: Claude Opus 4.6 Test scale: 36+ subagents across 5 rounds

Summary

• Input language has negligible impact on reasoning quality
• Output language affects detail coverage (not correctness) — English output scores ~8 points higher on average
• All 36+ test agents reached identical core conclusions regardless of language
• "Think in English" CLAUDE.md instruction consistently improves Cantonese output by +3-5 points (validated across 5 diverse topics)
• The improvement is strongest for reasoning/judgment tasks (+4-5) and smallest for terminology-heavy tasks (+1-2)

Contents

• report.md — Full research report with tables and findings
• claude-md-config.md — Optimized CLAUDE.md language settings
• test-results/ — Raw test outputs from all rounds