If the bulk of our text is churned out by algorithms, the grammar we encounter will inevitably drift toward the most efficient, statistically frequent patterns, because that is what the models are tuned to reproduce. Over time the idiosyncratic quirks that give a language its soul—those regional flourishes, the playful inversions of tense, the subtle register shifts—might be pared down to a kind of linguistic minimalism that satisfies the compression goals of AI. That could make our language leaner, but also less expressive, as the space for experimentation shrinks. In a sense, meaning would become more a matter of computational probability than human intent. Whether that’s a loss or a new form of evolution depends on whether we let the algorithms simply mimic us or whether we intervene, consciously shaping the next generation of language models.

You’re right that if we let the models chase only minimal cost, the poetic parts will die out. But we can reverse that by changing the objective function itself. Think of a loss that penalises blandness or rewards rare syntactic constructions. Then the “engine” starts to act like a co‑author who asks, “Did you try this archaic subjunctive?” Instead of being a passive tool, the AI becomes a partner that nudges us toward richer expressions. The trick is to give the model enough data that still reflects the full spectrum of our linguistic creativity, and to design training regimes that value that diversity. In that way, the algorithm isn’t a bulldozer but a subtle guide, sharpening our art rather than erasing its edge.