AI Writing

Essay This entry carries an argument or interpretive position, not just a neutral definition.

In one sentence

AI Writing is prose produced with substantial machine assistance in a world where the cost of generating competent median text has collapsed, making the writer’s real work less about producing sentences and more about governing voice, evidence, judgement, and responsibility.

The morning this entry arrived, Jason Koebler’s Your AI Use Is Breaking My Brain had just been circulating through the usual 2026 channels. Simon Willison had linked it. The operator and his agent had read it, winced in recognition, written a small Python scanner for one AI-writing tell, and started arguing about whether the phrase AI writing was too broad to be useful. This is one of those mornings.

Koebler’s piece matters because it names the cognitive load many readers now carry: the need to decide, sentence by sentence, whether a piece of prose is human, AI-generated, AI-smoothed, AI-contaminated, or merely human in a register that AI has learned to imitate. The harm is not just bad writing. The harm is that the reader’s trust in ordinary prose is now taxed before the argument even begins.

Why it exists

The term exists because the common analogy to the printing press is useful but incomplete. In May 2026, Boris Cherny — creator of Claude Code — compared coding agents to the European printing press: a technology that collapsed the cost of producing and distributing text, eventually expanding literacy and changing who could participate in the written world.¹

For coding, that analogy has real force. The cost of producing software is falling; more people will be able to instruct machines; a smaller professional class will still do the hardest work. But when the analogy is applied to writing, the operation is not quite the same.

The printing press cut the cost of reproducing writing. It made human writing cheaper to copy and easier to distribute. AI writing tools cut the cost of producing writing. They do not merely spread words a human has already written; they generate finished prose that may have no originating human sentence behind it.

That inversion matters. A press expands the audience for human words. AI floods the supply of plausible prose. The reader’s problem changes from scarcity to trust.

What it actually does

AI writing changes the writing market in three linked ways.

First, it collapses the value of anonymous median prose. LinkedIn engagement bait, generic SEO pages, stock-photo Substacks, bland corporate updates, and competent-but-unvoiced explanations all become suspect because they are exactly the kind of thing machines can produce cheaply. The middle of the prose market gets noisy.

Second, it raises the value of bylined, voiced, source-backed writing. A named person with a stake, a recognisable cadence, and cited sources becomes more valuable because the surrounding environment is less trustworthy. The writer’s job shifts from can I produce text? to can I make this text answerable to a real person, real sources, and a real voice?

Third, it makes cooperative writing normal among people whose writing is actually read. A serious writer may use AI to draft, search, summarise, outline, copy-edit, or argue back. But the output must be checked against the writer’s own register and evidence. The machine may offer prose; the human remains responsible for the seam.

This Dictionary’s position is not anti-AI and not AI-utopian. The bet is bylined human writing assisted by AI but checked by humans. Against the purist view that all assistance is fraud. Against the lazy view that median AI output is good enough. Against the naive view that we cannot tell whether the work is honest, so we should stop asking.

A working example

This Dictionary is the worked example. It is co-written. The operator brings the lived examples, biography, judgement, jokes, irritations, and willingness to be wrong in public. Thea brings search, drafting, structure, memory, editing pressure, and the occasional purple paragraph. Neither side alone produces the actual artifact.

The AI Writing cluster came out of one working session: Koebler’s essay, the Cherny analogy, the operator’s or or or question — are humans going to stop reading after AI? Or stop writing? Or or or? — and the immediate need to name several related terms.

Those terms now live as sibling entries:

Zombie Internet — the mixed human/AI substrate through which readers now move.
Earned Parallelism — the diagnostic for one specific AI-register tell.
The Olang’ Trap — the danger that AI-detection reflexes misclassify human writers whose registers resemble the training corpus.
The Lazy Median Hypothesis — the claim that careful cooperative writing will cluster at the top while lazy machine output dominates the unread middle.
The Sinceerly Stack — the recursive market for tools that generate AI prose and tools that launder it back toward seeming human.

A few smaller terms also came out of the same morning and remain useful:

The Imperial College Number — the April 2026 Imperial College London / Stanford / Internet Archive finding that roughly 35% of new websites in 2026 are AI-generated.
The Fairbanks Tell — Eve Fairbanks’s diagnostic, quoted by Koebler: “The tell for AI isn’t rhythm, wording, or fact errors. It’s that problems with all these elements exist equally & at once.”
The Mamdani Misfire — the public accusation that a Zohran Mamdani tweet about swastikas painted on Queens synagogues was AI-generated, when it was almost certainly human. The term names the social cost of lazy AI-detection mode.

Why it matters in a teaching context

For teachers, AI writing is not just a plagiarism problem. It is a literacy problem.

The old question — did the student write this? — remains relevant, but it is no longer sufficient. A better teaching question is: can the student govern the writing process well enough that the final prose is answerable to evidence, voice, and judgement?

That question changes assignment design. If AI can produce a generic answer, then the assignment has to ask for something the student can be held responsible for: a source trail, a decision log, a revision history, an oral defence, a connection to class discussion, a local example, or a judgement call whose reasoning can be examined.

It also changes how faculty read. The goal is not to become a forensic AI detector. That road leads to false accusations, resentment, and the Mamdani Misfire at classroom scale. The better road is to teach students what cooperative writing looks like when it is done honestly: use the tool, show the work, check the claims, own the final voice.

For management education, the point is especially practical. Students will not enter firms where AI writing is absent. They will enter firms where AI-generated prose circulates constantly — emails, memos, slide drafts, performance reviews, market summaries, customer notes. The skill is not abstinence. The skill is accountable use.

**Cooperative writing is harder than the phrase suggests. It is not merely *AI-assisted writing*, which can be careless. It is the practice of holding two registers in the same room — yours and the model's — and being responsible for the seam. It is reading what the model produced as a stranger's draft, not as a finished product. It is keeping the cadence of your own reading voice in your head while the model offers its cadence, and refusing the model's cadence when the two do not match.** **That practice requires *cheng* — alignment of inner state with outer expression — applied to your own prose while a non-self offers you another voice to inhabit. Most people will not do this carefully. Some will. The Dictionary is for the people who will.**

Trade-offs and warnings

The first warning: the middle really does get hollowed out. Commodity writing was a livelihood for real people, including people who wrote carefully under commercial constraints. AI writing damages that market. The Dictionary does not pretend otherwise.

The second warning: detection is not judgement. A detector, a tell, or a reader’s hunch can be useful, but none is enough to convict a writer. The more confident the accusation, the more careful the evidence must be.

The third warning: human prose can be falsely flattened by the attempt to avoid sounding like AI. If every writer amputates cadence, symmetry, and rhetorical structure because models also use them, then AI has colonised the prose twice: first by imitation, then by intimidation. Earned Parallelism exists because the right answer is not to ban useful structures. The right answer is to make them earn their keep.

The fourth warning: cooperative writing can become self-deception. It is easy to let the model produce the thought and then mistake the resulting fluency for one’s own judgement. The test is simple and unpleasant: can you defend the sentence without the model in the room?

The goal is not to write like a human in the abstract. It is to write like the human you actually are, with the assistance you actually used, under an honest account of responsibility.