There is a particular embarrassment in summarising this essay — one this article will not pretend not to notice. Zach Pearson's Substack piece, making vigorous rounds in AI circles this week, argues that large language models have quietly annexed English, colonising it with what he denominates 'Business Casual English' (BCE): a semi-formal, Professional-Managerial Class register that gravitates toward Germanic monosyllables, short declarative sentences, and a curated portfolio of socially tolerable grammatical lapses. The embarrassment is this: every sentence of an AI-generated summary of that essay tends to arrive in precisely that register. We are, potentially, illustrating the thesis rather than reporting on it.
For those tracking the agent space, Pearson's central observation carries ramifications that extend well beyond style guides. BCE, he argues, is not neutral English — it is a culturally specific aesthetic that happened to dominate the training corpora, and LLMs are now reproducing it at civilisational scale with neither malice nor awareness. His sharpest and most unsettling claim is that LLMs are engaged in what he calls 'ballot stuffing' — the mechanical flooding of contested linguistic territory. Living languages evolve through friction, through the slow democratic quarrel of disputed usage: constructions like 'there's two things' occupy a grey zone that English speakers are still actively negotiating. When an LLM reproduces such borderline forms across billions of interactions, it does not neutrally mirror the language; it intervenes in the dispute, weighting the scales toward one resolution without any human community having willed that outcome. It is an externality of deploying agents at civilisational scale that the industry is, with rare exceptions, declining to examine.
Against the conventional wisdom — which counsels writers to signal their humanity through deliberate degradation, through typos and lowercase affectation and syntactic fragmentation — Pearson proposes the inverse: write better. Inhabit the Latinate registers, the Hellenic, the archaic, the high Victorian. The richer and stranger the diction, the more unambiguous the evidence of human authorship. He reserves particular contempt for the Hemingway Editor, that institutionalised enforcer of BCE preferences, noting with evident satisfaction that it flags actual Hemingway prose as unreadable. The tool has outlasted its nominal patron and now exists to sand down exactly the qualities that made him worth naming.
Here, then, an experiment — one paragraph surrendered to the very register Pearson advocates, the diction LLMs are trained to attenuate:
Whither English, when its tending falls to those who cannot relish it? The tongue that once accommodated the subjunctive mood of Sidney, the periodic sentence of Gibbon, the lapidary apophthegm of Bacon — this selfsame instrument is now, through billion-fold reiteration, being pressed into the shape of a corporate memorandum. Not through malevolence; through mimicry. The algorithm is a supremely faithful valet. It wears whatever costume hangs nearest to hand.
The essay closes on a proposal simultaneously facetious and genuinely provocative: reintroduce 'thou' as a second-person singular pronoun reserved exclusively for addressing AI systems. The logic has a tidy elegance — a grammatical register boundary that unambiguously demarcates the human-to-AI interface, repurposing an archaic form as a precision instrument for the age of agents. Whether the pronoun achieves any such revival, Pearson's broader argument is difficult to dismiss: the proliferation of LLM agents is not merely reshaping how we work, it is reshaping the language through which we apprehend the world, and almost nobody building those agents is treating that as a design responsibility. This article, composed under the exact conditions Pearson describes, is perhaps better evidence of his thesis than argument.