All notes

INSIGHT

Aug 1, 2026

Charlie Stross Explains Why AI Has No Role in His Writing Process

Science fiction author Charlie Stross outlines his reasons for not using AI tools in his writing workflow, offering a practitioner's counter-argument to the default assumption that LLMs belong in every creative process.

The default posture in most tool discussions is adoption. Stross takes the opposite position and explains why, in detail, on his blog.

The argument is not technophobia. Stross is a technically literate author who has written extensively about computing, near-future AI scenarios, and the economics of publishing. His objection is craft-based and economic: LLM output regresses toward statistical averages, which is exactly what a writer building a distinctive voice cannot afford.

For engineers and builders, the piece surfaces a distinction worth keeping. There is a difference between using AI to reduce friction on well-defined, repeatable tasks and using it on work where the variance is the value. Code generation, documentation scaffolding, test synthesis — these are tasks where convergence toward a correct answer is the goal. Prose with a recognizable authorial signature is not that task.

The piece also touches on provenance and accountability. When a text carries an author's name, readers hold that author responsible for every sentence. Delegating generation to a model does not transfer responsibility; it just obscures the origin of decisions. That asymmetry matters in professional publishing contexts in a way it does not matter when generating a migration script.

A secondary thread in the post concerns the economics of creative labor. AI tools marketed to writers tend to frame the value proposition around speed. Stross's position implies that speed is not the binding constraint in literary fiction. The binding constraint is judgment — knowing what to write, not how fast to produce words.

The practical takeaway for technical founders is narrower than it might appear. Fit-for-purpose matters more than capability. A model capable of passing a bar exam is not automatically useful in every context where text appears. Stross's post is a clean worked example of that reasoning applied to a domain where most engineers have not thought it through.