💬 NLP

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

This paper argues that historical imperial inequalities persist in contemporary large language models through shared training data, resulting in radical digital under-support for most writing systems and systematic, convergent errors across diverse model architectures that disproportionately misattribute non-religious scripts to religious use.

Original authors: Hiroki Fukui

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hiroki Fukui

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The Ghost in the Machine

Imagine that history is a giant library. For 500 years, powerful empires (like Spain, Britain, or France) decided which books to keep, which to burn, and which languages to write in. They destroyed many local writing systems and forced others to change.

This paper argues that Large Language Models (LLMs)—the AI brains behind tools like ChatGPT—are not new, neutral inventions. Instead, they are like ghosts haunting a library built by those empires. Even though the empires are gone, their "ghosts" are still in the building. The AI doesn't know how to read certain languages not because the engineers were mean, but because the library (the internet data) they learned from is missing those books.

The Main Findings, Broken Down

1. The "Digital Tax" (Why some languages cost more)

Imagine you go to a toll booth to enter a highway.

English speakers pay $1 to drive 10 miles.
Speakers of minority languages (like Limbu or Adlam) are charged $31.70 to drive the exact same 10 miles.

Why? The AI breaks text down into small chunks called "tokens" (like Lego bricks).

For English, the AI has a huge box of pre-made Lego bricks. It snaps them together easily.
For minority scripts, the AI doesn't have the right bricks. It has to break the letters down into raw, meaningless bytes (like smashing the Lego bricks into plastic dust and trying to rebuild them from scratch).
The Result: It takes 31 times more "computing effort" to process a sentence in Limbu than in English. This isn't a bug; it's a "tax" built into the system because those languages were historically suppressed and have less data on the internet.

2. The "Digital Funnel" (Who gets to be heard)

The researchers looked at 300 different writing systems in human history. They tried to see how well the digital world supports them.

The Funnel: Imagine a giant funnel.
- Top: 300 scripts enter.
- Middle: Many get stuck because they aren't in the computer's dictionary (Unicode) or can't be read by scanners (OCR).
- Bottom: Only 29 scripts (less than 10%) make it all the way through to the bottom.
The Pattern: The scripts that make it through are almost exclusively the ones used by historical empires (Latin, Chinese, Arabic, Cyrillic). The scripts that get stuck are the ones those empires tried to erase.
The Tragedy: 60 languages are still spoken by living people today, but the digital world treats them as if they are dead. They are "living but digitally dead."

3. The "Four Friends" Experiment (It's not just one AI)

To prove this wasn't just a mistake by one company (like OpenAI), the researchers asked four different AI families (Claude, GPT-4o, Grok, and DeepSeek) the same 3,000 questions about writing systems.

The Surprise: These four AIs are built by different companies, in different countries, with different code. They shouldn't agree on mistakes.
The Reality: They agreed on their mistakes 90% of the time.
The Analogy: Imagine four students who went to four different schools. If they all get the same wrong answer on a history test, it's not because they copied each other. It's because they all read the same biased textbook.
The Bias: The AIs consistently guessed wrong in the same way. For example, if a script was from a non-Western country, the AIs were very likely to guess, "Oh, this must be used for religion," even if it wasn't. They were filling in the blanks with stereotypes found in the data.

4. The "Empire Without an Emperor"

The most important finding is that nobody is currently in charge of this bias.

The engineers didn't sit down and say, "Let's make Limbu speakers pay more."
The bias is structural. It happened like this:
1. Empires destroyed communities and reduced the number of speakers.
2. Fewer speakers meant fewer books and websites written in those languages.
3. Fewer websites meant the AI had no data to learn from.
4. No data means the AI is bad at that language.
The "Empire" doesn't need a leader anymore. The damage is baked into the statistics. The AI is just doing math based on a world that was already unequal.

The "Religion" Glitch

One specific finding stood out: The AIs were obsessed with guessing that scripts were "religious."

Out of all the mistakes the four AIs made together, 43% were about guessing a script was used for religion when it wasn't.
Why? The internet is full of old, colonial-era descriptions of non-Western cultures that focused heavily on their "mystical" or "religious" aspects, while ignoring their daily, administrative, or scientific uses. The AI learned that "Non-Western Script = Religion."

The Conclusion: What Can Be Done?

The paper suggests that we can't just "fix" the AI by tweaking its code.

The Problem: The problem is the library (the training data), not the reader (the AI).
The Solution: To fix this, we need to change the library. We need to add more books written by the people who speak these languages, in their own scripts, describing their own lives.
The Warning: Until we do that, the AI will continue to be a mirror of the last 500 years of history, reflecting the same inequalities, even if no one intended for it to happen.

In short: The AI isn't broken; it's just faithfully repeating the history of who was allowed to write and who was silenced. The "digital afterlife" of empires is still very much alive.

Technical Summary: The Digital Afterlife of Empires

Problem Statement
Large Language Models (LLMs) process the world's writing systems with radical inequality. While the scripts of historical empires (e.g., Latin, Cyrillic, Arabic) are processed efficiently, scripts of colonized or marginalized communities suffer from significant digital exclusion. This paper investigates whether contemporary digital infrastructure and LLMs reproduce the structural inequalities inflicted by historical empires. Specifically, it asks if the "broken cultural clocks" of writing systems—documented in a companion study (Fukui, 2026) as structural deviations caused by imperial intervention—persist in the latent space of modern language models through the shared training corpus, rather than through specific architectural design choices.

Methodology
The study constructs the Digital Script Representation Index (DSRI), a seven-axis instrument applied to the Global Script Database (GSD), which catalogs 300 writing systems across 43,000 years of history. The DSRI measures digital presence across:

Token Efficiency Ratio (TER): The number of tokens required to represent a grapheme cluster relative to a Latin baseline, measured across four tokenizers (tiktoken, Mistral, Qwen, Phi-4) using parallel text (UDHR, eBible) and Wikipedia corpora.
Unicode Coverage: Encoding status, codepoint ranges, and versioning.
Web Corpus Volume: Derived from Wikipedia article counts and CC-100 corpus byte counts.
LLM Generation Fidelity: Accuracy in answering 10 binary typological questions (e.g., directionality, religious use) against GSD ground truth.
OCR Support: Character error rates across Tesseract, Google Cloud Vision, and Azure.
Machine Translation Coverage: Binary support across Google, NLLB-200, and DeepL.
Input Method Availability: Native keyboard/IME support on iOS, Android, Windows, and macOS.

To address the circularity of using a single LLM as both instrument and object, the study employs a cross-architecture validation protocol. Four independent LLM families (Claude Haiku 4.5, GPT-4o, Grok-3-mini, DeepSeek-V3) were queried with 3,000 identical typological prompts (300 scripts × 10 features), yielding 12,000 responses.

Statistical analyses include:

Serial Mediation Modeling: Testing the pathway Imperial Intervention → Speaker Population → Web Corpus Volume → Tokenizer Efficiency.
Structural Equation Modeling (SEM): Assessing model fit at $n=45$ (the limit of scripts with measurable parallel text).
Convergence Analysis: Calculating Spearman correlations of error deviations (false-positive and false-negative rates) across model pairs to identify architecture-independent biases.
Sensitivity Checks: Excluding the "used for religion" feature to test if convergence is driven by a single contested ground-truth coding.

Key Results

Digital Exclusion Funnel: Only 29 of 300 scripts (9.7%) are fully supported by contemporary digital infrastructure. Among 158 living scripts, 60 (38.0%) lack complete support. The exclusion follows a colonial geography, with scripts from Africa, South/Southeast Asia, and Oceania disproportionately affected.
The "Digital Tax": Tokenizer efficiency varies by a factor of 31.7 across 45 measured scripts (Latin baseline = 1.0; Limbu = 31.7×). This disparity is consistent across four independent tokenizers, indicating a structural rather than implementation-specific cause. Scripts with the highest token costs often lack natural-language web corpora (e.g., no Wikipedia paragraphs), compounding the exclusion.
Causal Mediation: The relationship between imperial intervention and tokenizer inefficiency is fully mediated by speaker population and web corpus volume. The direct effect of empire on tokenization is indistinguishable from zero ( $\beta = -0.22, p = 0.39$ ). The structural equation model achieves a near-perfect fit at $n=45$ (CFI = 1.01, RMSEA = 0.00), with digital presence contributing an incremental $\Delta R^2$ of 0.366.
Cross-Architecture Convergence: Four independent LLMs exhibit highly correlated error patterns ( $\rho = 0.85–0.98$ ) when deviating from base-rate expectations. This convergence persists even when the "religion" feature is excluded, suggesting the bias is external to any single model and resides in the shared textual substrate.
Unanimous Errors: All four models answer 172 script-feature items identically wrong. Over-attribution (assigning a feature not present) outnumbers under-recognition by a ratio of 3.9:1. The feature "used for religion" concentrates 43.6% of these errors. When excluded, the asymmetry persists at 1.77:1 ( $p=0.008$ ), indicating a multi-channeled bias where non-Western scripts are systematically mapped onto templates of "ancient," "pictographic," or "religious" origins.
Two Independent Wounds: The study finds a statistical independence between the structural disruption of scripts (molecular clock deviation) and the demographic destruction of script-using communities (byte-fallback status). Empires inflict both: altering the script's form and destroying the community that uses it.

Significance and Claims
The paper argues that the "empire" requires no emperor to reproduce its violence in the digital age. The structural inequalities of colonial history have sedimented into the statistical distribution of the global web. Because LLMs are trained on this corpus, they inherit and reproduce these biases with "mathematical fidelity," regardless of their specific architecture or the organization that built them.

The findings challenge the notion that bias is a result of specific model design choices or alignment procedures. Instead, the paper posits that computational assimilation—the systematic over-attribution of majority-script characteristics to minority scripts—is a property of the textual substrate itself. The convergence of four independent models suggests that the "imperial cartography of writing" is encoded in the data, not the algorithms.

The authors modestly frame the causal mediation results as "suggestive rather than confirmatory" due to the statistical fragility at the margin (the BCa bootstrap interval grazes zero), acknowledging that the $n=45$ sample size is a "horizon" imposed by digital exclusion itself, not a random sampling limitation. However, the cross-architecture convergence provides robust, independent evidence that the bias is structural and upstream of the models.

The paper concludes that remediation cannot be achieved by fine-tuning individual models. Because the bias is located in the corpus, interventions must be upstream: expanding tokenizers to include all Unicode scripts, correcting the frequency-based vocabulary allocation, and, most critically, addressing the demographic and institutional damage inflicted by empires that left these communities without a digital voice. The "digital afterlife of empires" is preserved not by malice, but by the method of frequency-based learning on a non-neutral corpus.