■ CRITICAL RISK ■ Technology
For first-draft documentation of well-understood features, yes — large language models do it in seconds and adequately. The writers who survive stop being typists of explanations and become owners of information architecture, accuracy, and the questions engineers forgot to answer.
“AI writes documentation that's equally boring but infinitely faster.”
Our AI replacement risk score — how we score jobs
Technical writers turn engineering reality into usable words: API references, user guides, release notes, runbooks, compliance documents. The unglamorous truth of the job is that much of it was always transformation work — interviewing an engineer, reading a spec, and restructuring that material into clear prose against a style guide. Docs-as-code workflows made writers half-engineers already: Markdown, Git, static site generators, CI checks. Then language models arrived, and transformation work is precisely what they do.
Today an LLM drafts an API reference from a code annotation, converts a changelog into release notes, rewrites for reading level, and answers style-guide questions instantly. Auto-generated reference documentation from OpenAPI specs was normal before AI; now the connective prose around it is generated too. Companies under headcount pressure notice that one senior writer with AI tooling produces what a team of four once did — and the junior-writer apprenticeship pipeline, where people learned by writing routine docs, is the first casualty. Support content, tooltips, and knowledge-base articles are increasingly machine-first with human review, when they get review at all.
The resistant core is everything before and after the prose. Someone must discover what actually needs documenting — which means interviewing engineers, using the half-broken feature, and finding the gap between what the code does and what the PM believes it does. Someone must own information architecture across a thousand pages, decide what to delete, and catch the confident hallucination in generated text before a customer follows it into a production outage. Regulated industries — medical devices, aviation, pharma — still require accountable authors. Our 82 lands on the occupation as staffed today: the verification-and-architecture roles remain, but they need far fewer hands than the drafting work that filled most job descriptions.
Automatability: our editorial assessment of current and near-term AI capability
The pressure is current, not projected: LLMs draft competent documentation today, tech-sector layoffs have repeatedly hit documentation teams, and junior writing roles are quietly not being backfilled. Through the late 2020s expect the standard model to become one senior writer-editor orchestrating AI drafting for what a team once produced. Regulated-industry documentation and docs-engineering roles erode slowest.
Not dead, but violently reshaped. AI now drafts routine documentation well enough that teams need fewer writers, and entry-level roles are evaporating fastest. What survives is the senior layer: information architecture, docs tooling, engineer interviews, and accuracy verification. If your value is producing clean prose from clear inputs, you are exposed; if it is finding out what is true and structuring it, you are not — yet.
Common patterns today: generating API references from code annotations, drafting release notes from changelogs, producing knowledge-base articles from support tickets, and keeping docs synchronized with product changes. Humans review for accuracy — with varying diligence. The practical effect is fewer writer-hours per page, which translates to smaller teams rather than better-rested ones.
Three stacks: docs engineering (Git, static site generators, CI, and AI-generation pipelines), discovery (interviewing engineers, testing products, reading code well enough to spot gaps), and editorial verification of machine output. Domain depth in a regulated field is a strong fourth. The unifying theme: move your value upstream or downstream of the drafting that AI now owns.
For routine material, often yes — and 'good enough, instantly, for free' wins a lot of business arguments. The failure mode is confident inaccuracy: generated docs describing parameters that do not exist or steps that silently changed. Companies that skip human verification ship those errors to customers. That risk is the strongest remaining case for professional writers, and smart ones are building careers on it.