Her 2025 experiment, now known as , found that when asked to generate counterfactual histories (e.g., "What if the printing press had been invented in 100 AD?"), models trained primarily on English produced 40% less creative divergence than models fine-tuned on Romance languages.

In an era obsessed with alignment, safety, and scaling, Rednik is the strange, Slavic-inflected whisper reminding us that before we align AI with human values, we should probably make sure we aren't confusing "human values" with "English syntax."

Her conclusion was stark: By training our AIs on a global, flattened English corpus, we are not just standardizing language. We are standardizing imagination. Naturally, the tech world has pushed back. OpenAI’s chief ethicist called her work "linguistic determinism dressed up as data science." A prominent Google DeepMind researcher accused her of "romanticizing non-English syntax."

The Unspoken Pattern (Rednik, 2023) | "The Rednik Threshold" (arXiv:2503.08821) What do you think? Is grammar destiny for AI? Or is Rednik overthinking the subjunctive? Drop your take in the comments. Author Bio: Jordan M. is a recovering digital strategist and M.A. candidate in Language & Technology at Columbia.

But the more pointed critique came from literary circles. Critics like Harold Voss (The New Criterion) argued that Rednik reduces literature to a mere wiring diagram. "She treats Proust's subjunctives as engineering schematics," Voss wrote. "The soul is missing."

She demonstrated that languages with a strong subjunctive mood (Romance languages, German, Greek) encode uncertainty and counterfactual thinking within the structure of a sentence . English, by contrast, relies on auxiliary verbs ("would," "could," "might"), which are statistically rarer in LLM training corpuses.