Sci-fi authors warn AI threatens writing’s provenance

Two prominent science-fiction authors, Charles Stross and John Scalzi, have condemned the effect of generative AI on writing, arguing that large language models create copyright risks, undermine authorship, and make it harder for publishers to establish whether a work is genuinely human-written.

Stross: LLM training is copyright theft

Stross, author of the Laundry Files novels, wrote in a Saturday post that he has no interest in using an LLM to produce fiction because writing is itself the point.

“I do not want or need a large language model to write my fiction for me. I write fiction compulsively – before I was published I wrote for many years as a hobbyist – so why on earth would I pay someone else to take my fun away?”

He also said serious authors should avoid LLMs and pointed to his involvement in the class-action lawsuit against Anthropic AI, which alleged that the company pirated eBooks to train its models. Stross said his objections extend beyond the use of copyrighted material: his website has allegedly been overwhelmed by bots scraping data for LLM training, leaving an old server sluggish or unavailable.

Stross called AI companies “absolute scum”, arguing that they take copyrighted work, train systems intended to compete with its creators, and then plan to monetize access to those systems through token sales and subscriptions. He compared the model to a burglar stealing someone’s possessions and then trying to sell them back.

Recommended reading

MIPT gives rescue drones a visual memory

Scalzi warns of provenance problems

Scalzi, best known for the Old Man’s War series and the Hugo Award-winning novel Redshirts, made similar points while commenting on a report about a highly anticipated debut novel that was withdrawn from publishing and film or television consideration amid concerns that AI had been used in its creation.

He argued that the central issue is not simply whether AI-generated material can receive copyright protection. The risk, he wrote, is that models are trained on copyrighted material, allowing rights holders to challenge prose generated from that training.

Scalzi expects publishers to look for evidence of AI use, even though existing detection tools are notoriously unreliable. That could put new writers at a disadvantage because they may not have a large body of pre-ChatGPT work to demonstrate their authorship.

“No, Microsoft, I don’t fucking want Copilot to 'help me write' in Word, and no, Google, I don’t fucking want Gemini to 'help me write' in Google Docs.”

Scalzi said Microsoft and Google were embedding AI tools into the writing software authors depend on, creating provenance concerns even when writers do not use those features. He advised authors to avoid AI editing, artwork, translation, and other publishing services because using one component could cast doubt on the rest of a book.

“When I say that I have it in my contracts that I require every aspect of my book production to be done by humans, it’s not just because I want to honor human work and input. It’s also because I’m protecting my own reputation by doing so.”

The limited uses Stross accepts

Stross acknowledged that a locally running model could be useful for analyzing his own manuscripts, provided it operates on his hardware without a cloud service. He suggested tools that could:

• Derive a scene-by-scene timeline from a manuscript.
• Map where and when characters appear.
• Produce a frequency-distribution heat map of character names.

He compared those functions to a spelling checker: useful for decision support, but not a replacement for the author’s work. He also warned that submitting an outline to ChatGPT or Claude could create a risk that a future model trained by the same company might reproduce part or all of it.

Stross does see a legitimate artistic use for LLMs. In an excerpt from a space opera he is editing, an LLM appears as a fictional translation system that converts names, institutions, measurements, and culturally specific language while warning readers that it may have introduced hallucinations.

Neither author discussed George R. R. Martin, who writes on a PC without an internet connection using DOS and the old WordStar word processor. The source suggests that drafts and revision histories from such systems could provide another way for authors to demonstrate that their work is their own.