The Social Contract of Writing in the Age of Large Language Models

Large language models have slipped into almost every commercial niche; from drafting emails to composing code, the promise of automation is everywhere. Yet the most profound incursion is into the act of writing itself. The convenience of generating paragraphs at scale tempts authors who face relentless pressure to publish, but the consequences run deeper than a simple productivity boost.

A Forgotten Agreement

When a reader opens a piece of text, there is an unspoken expectation that the author has invested more mental labor than the reader will. This assumption underpins the trust that the writer understands the subject and has taken the time to shape the argument. The Oxide RFD 576 document, a public request for discussion from Oxide Computers, spells this out in section 2.4: LLM‑generated prose undermines a social contract of sorts; absent LLMs, it is presumed that of the reader and the writer, it is the writer that has undertaken the greater intellectual exertion.

If the writer outsources that exertion to a model, the contract is broken. The reader can no longer infer that the text reflects the author’s own reasoning, even when the facts are correct and the grammar immaculate. The breach is not merely aesthetic; it is epistemic, because the effort behind the words signals reliability.

Homogenization and the Rise of “Model‑Style” Phrases

Repeated exposure to model‑generated output creates a feedback loop. Certain constructions—it is not X, it is why; the overuse of em‑dashes; the formulaic you’re not imagining it, the problem is real—have become linguistic shortcuts that many writers now mimic unconsciously. A study from the Max‑Planck Institute for Human Development reported that ChatGPT’s preferred vocabulary has nudged human speakers to use words such as meticulous and adept more frequently. The phenomenon shows that even readers who never type a prompt are being reshaped by the corpus of AI‑produced text.

The Taint of Post‑November‑2022 Content

Anything published after November 30 2022 bears at least a trace of model influence, whether through direct generation or subtle stylistic borrowing. The only way to avoid this imprint is to confine oneself to pre‑LLM archives, a practical impossibility for anyone seeking contemporary discourse.

Why Originality Will Gain Currency

As the volume of AI‑written material swells, the market for genuine, effortful expression will tighten. Companies that train future models need high‑quality human text as training data, which makes authentic writing a scarce resource. In a sea of algorithmic prose, a piece that reveals the author’s unique voice—typos, idiosyncratic metaphors, even imperfect grammar—will stand out precisely because it deviates from the model norm.

A Thought Experiment

Imagine a student who can either spend weeks researching a topic and produce a competent essay, or submit a ChatGPT‑generated draft and receive a top grade. The latter option subverts the educational contract that grades should reflect personal learning. Over time, institutions may have to redesign assessment to value process over product, perhaps by requiring drafts, revision histories, or oral defenses.

Implications for Readers and Platforms

1. Critical Literacy – Readers will need tools to detect model‑generated text, similar to how we now evaluate deep‑fake media.
2. Platform Policies – Websites may adopt disclosure requirements, asking authors to label AI‑assisted content, thereby restoring some transparency.
3. Economic Incentives – Writers who market themselves as “LLM‑free” could command premium rates, much like hand‑crafted goods in a mass‑produced market.

Counter‑Perspectives

Some argue that the democratizing effect of LLMs outweighs the loss of the old contract. They point out that models can help non‑native speakers express ideas they could not otherwise articulate, expanding participation in public discourse. Others note that the stylistic convergence may be temporary; as models are fine‑tuned on more diverse corpora, the range of generated voices could broaden.

A Personal Commitment

The author of this piece chose to write without assistance, spending an afternoon drafting while watching classic cinema. The effort was not about perfect grammar but about honoring the reciprocal obligation that makes reading a meaningful act. By acknowledging the contract and choosing to keep it intact, the writer hopes to signal that human‑crafted text still matters.

---

If you found this analysis useful, you can follow the author on Bluesky at @jola.dev or support the work via GitHub Sponsors.