Charlie Stross says his fiction stays 100% LLM-free

Science fiction author Charlie Stross has published a blog post stating that he does not use large language models, colloquially called AI, anywhere in his fiction writing process, and does not use them in his personal life either. He says he has been selling fiction he wrote himself since 1985 and novels since 2002, and offers proof: anyone can pick up a physical copy of Singularity Sky, Iron Sunrise, The Atrocity Archives, or anything else he published before 2015, the year OpenAI was founded. From that period alone he counts seven Hugo-shortlisted novels, three Hugo-winning novellas, and two Locus-award winning novels, which he says shows he does not need AI to write award-winning stories.

Stross's objection centers on how the models are built. He says he is one of the parties to the class-action lawsuit settlement against Anthropic over pirating ebooks to train its large language models, though he gives no settlement amount or terms. He also says his blog's server, which he describes as old and feeble, periodically slows or crashes because it gets swarmed by Chinese and other foreign botnets scraping content to train LLMs. He argues that the most popular Western models, naming Claude, Gemini, CoPilot, and ChatGPT, both scrape text indiscriminately for training and retain the queries users send them as future training data, meaning an author who fed a story outline into one of these tools as a prompt could see a later model reproduce parts of it for someone else.

He also rejects claims that LLMs are sentient, arguing they are word-association mechanisms with no embodiment and no way to tie the text they generate to real-world phenomena. He connects the illusion of a mind behind generated text to an observation he attributes to Joseph Weizenbaum, inventor of the original ELIZA chatbot, who realized at MIT in the late 1960s that people expect a mind on the other side of any text they read.

Stross adds caveats. He does not dismiss the underlying technology altogether: he calls it foolish to deny the effectiveness of image recognizers built on generative adversarial networks, which he describes as the key neural network technology underlying LLMs, and says LLMs are genuinely useful for large-scale statistical text analysis, citing Linear A as an example. He says he could see value in a tool, run entirely on his own hardware with no cloud service and no subscription fees, that digests a manuscript into a scene-by-scene timeline he could query while editing, similar in role to a spelling checker: a decision-support aid, not a substitute for doing the writing himself.

He closes by stating that as of August 2026 his fiction remains 100% LLM-free, and says he will update the declaration if that changes. As a closing aside, he shares a snippet from a space opera he is currently editing: a fictional 'Translator's Note' claiming the account was translated by a non-sapient large language model, which he calls the only use he has found so far for framing LLMs within his fiction, used to sidestep the problem of inventing alien terminology from scratch.

Key facts

Charlie Stross says he has never used LLMs to write fiction, pointing to physical copies of novels published before 2015, the year OpenAI was founded, including Singularity Sky, Iron Sunrise, and The Atrocity Archives.
From before 2015 he counts seven Hugo-shortlisted novels, three Hugo-winning novellas, and two Locus-award winning novels as evidence he does not need AI to write award-winning fiction.
He says he is a party to the class-action settlement against Anthropic over pirating ebooks to train its LLMs, though no settlement terms are given.
He says his blog's server periodically slows or crashes because it gets swarmed by Chinese and other foreign botnets scraping data to train LLMs.
He argues that the most popular Western LLMs, Claude, Gemini, CoPilot, and ChatGPT, both scrape text indiscriminately for training and retain submitted queries as future training data.

Why it matters

A commercially and critically successful genre author using his own platform to publicly refuse LLMs adds a concrete voice to the copyright fight over AI training data. Stross ties the refusal to a live legal stake: he says he is a party to the class-action settlement against Anthropic over pirated ebooks used for training, which moves the objection from general principle to a case he is personally involved in.

Who it affects

Authors and publishers weighing whether to use generative tools in creative work, and readers trying to judge whether a given book was AI-assisted. The piece also names the companies behind the models he singles out: Anthropic, whose training practices are the subject of the lawsuit he is party to, and the makers of Gemini, CoPilot, and ChatGPT, which he groups together as models that retain user queries for future training.

How to use it

Stross draws a line between rejecting LLMs for the writing itself and accepting narrow, local tooling. He says a tool that runs entirely on his own hardware, with no cloud service and no subscription fees, and that turns a manuscript into a queryable scene-by-scene timeline showing where each character appears, would be useful to him in the same way a spelling checker is: a decision-support aid, not a replacement for doing the writing by hand.

How solid is it

The claim rests on Stross's own testimony and on a bibliography readers can verify directly: he names specific award-winning titles published before 2015 and gives exact counts of Hugo and Locus honors from that period. The lawsuit claim is also checkable in principle, since he names Anthropic as the defendant, though the post gives no settlement amount, terms, or court outcome, and no total count of his published novels or novellas beyond the pre-2015 award tally.

Risks and caveats

Stross does not treat all AI technology as suspect: he separately credits generative adversarial network based image recognition as effective and calls LLMs genuinely useful for large-scale statistical text analysis, citing the study of the undeciphered Linear A script. He also gives no specific figures for the botnet traffic he says has slowed his server, so the scale of that particular claim is not independently quantifiable from the post.

“I maintain that any serious author should shun LLMs like the plague.”

— Charlie Stross