The Digital Afterlife of Empires: Four Language Models Converge on the Same Imperial Cartography of Writing

Only 9.7% of the world's 300 writing systems have full digital support, and among living scripts still actively used by communities, 38% lack complete support—meaning LLMs are built on fundamentally incomplete digital infrastructure that systematically excludes many languages and writing systems.

Historical empires continue to shape AI bias: The research found a "serial mediation" pathway where imperial history influences which writing systems get represented online, which then determines how efficiently they're processed by language models, suggesting colonial-era power dynamics are embedded in modern AI systems.

All major LLM companies make identical errors: Four independent AI systems (Claude, GPT-4o, Grok, DeepSeek) converge on the same wrong answers about writing systems at rates of 85-98% correlation, indicating the problem isn't isolated to one company but reflects shared structural biases across the entire industry.

Religious scripts are systematically misrepresented: Over 43% of identical errors across all four models involve incorrectly attributing religious uses to writing systems, suggesting LLMs have absorbed distorted patterns about which scripts are associated with religion—a bias that could reinforce stereotypes.

The bias is structural, not accidental: The convergence of errors across different AI architectures and companies suggests these problems stem from shared training data and tokenization methods rather than individual model design choices, making them harder to fix through isolated interventions.