How Machines Absorb Cultural Heritage (And What Gets Lost in Translation)

Why It Matters

Mis‑representations in AI outputs can distort moderation, legal interpretation, education, and public discourse, amplifying bias and misinformation across the region.

Key Takeaways

• English-dominant training data yields fluent but shallow CEE language models
• Cultural riddles benchmark: ~48% factual accuracy, 16:1 hallucination ratio
• Models mirror nationalist framing of prompt language, not objective truth
• Local initiatives (Poland PLLuM, Cohere Aya) prioritize native data
• Treat AI training data as cultural infrastructure, not merely technical

Pulse Analysis

Large language models today are built on massive corpora that are overwhelmingly English. While this scale delivers impressive fluency, it leaves languages such as Hungarian, Croatian, and Serbian under‑represented. Recent internal benchmarks that probed cultural riddles native to these regions showed near‑perfect grammar but only about 48 % factual correctness, and a 16‑to‑1 ratio of confident hallucinations to admitted uncertainty. The models therefore act as linguistic mirrors, reproducing the syntax of a language without the embedded historical memory, folklore, or nuanced political context that give that language its identity.

The gap between surface fluency and deep cultural understanding is not academic; it surfaces in real‑world deployments. Content‑moderation tools trained on English slur lists may miss region‑specific hate symbols, while legal‑tech applications could misinterpret statutes that hinge on historically charged terminology. Educational chatbots risk passing distorted versions of national myths to students, and mental‑health assistants might overlook culturally specific expressions of distress. In contested historical narratives, the models often fabricate conciliatory “facts,” preferring a coherent story over a transparent admission of ignorance, which can erode public trust.

Recognizing AI as part of cultural infrastructure is prompting a new wave of localized initiatives. Poland’s PLLuM model, trained on 200 billion tokens of native Polish text, demonstrates that scale can be achieved without English translation pipelines. Cohere’s Aya project and Humane Intelligence’s bias research enlist native speakers to curate multilingual instruction sets, while Latvia’s TildeOpen offers open‑source models for 34 European languages with a focus on linguistic fidelity. Policymakers and industry leaders must fund similar data‑collection ecosystems across Central and Eastern Europe, ensuring that future AI systems preserve, rather than flatten, the region’s rich heritage.