On the non-use of AI in my writing process - Charlie's Diary

Charlie's Diary

Being the blog of Charles Stross, author, and occasional guests ...

Search

-->

[ Home ]<br>[ FAQ ]<br>[ Contact me ]<br>[ Older stuff ]

Back to: The Golden Age Of Bond Villains

On the non-use of AI in my writing process

By Charlie Stross

This isn't a blog entry I wanted to write, but it's a necessary one: a statement about the use of generative large language models (colloquially "AI") in my work.

I do not use LLMs in my work . I don't use them in my non-work life either, for that matter. I despise the grifters selling these toys as "tools" and trying to convince us to use them to generate plausible answer-shaped text strings in place of actual internet search for verifiable sources.

I've been selling fiction that I wrote myself since 1985 or thereabouts, and novels since 2002. If you want to verify that I have written novels without using an AI, simply pick up a physical copy of "Singularity Sky", "Iron Sunrise", "The Atrocity Archives", or anything else I published before 2015, the year OpenAI was founded.

Hint: you will find seven Hugo-shortlisted novels from that period, and three Hugo-winning novellas, also two Locus-award winning novels and a couple more novellas and stories. Clearly I don't need AI to write award-winning stories.

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?

You will note em-dashes in the preceding paragraph. I gather some "AI detector" services (themselves a generative AI product) flag em-dashes as signs of "AI generated" text. Listen, fuckers, LLMs sprinkle em-dashes in their output because LLMs exist to stochastically emit strings of text that approximate the form of their inputs, and they've been trained by stealing all the text on the internet that isn't nailed down, including pirate websites that distribute cracked e-books. So it's wholly unsurprising that LLM output exhibits quirks that mimic real writers.

Did I mention the "stealing" thing? This isn't hyperbole: I'm one of the parties to the settlement in the class action lawsuit against Anthropic AI for pirating ebooks to train their LLMs. That's not my only grievance, either. You may have noticed this blog performing sluggishly or crapping out from time to time over the past few months. That's because my server is old and feeble and periodically gets swarmed by Chinese and other foreign botnets scraping data for training LLMs.

I'm usually willing to cut actual human beings, as opposed to for-profit corporations, some slack where it comes to cracking DRM, or even downloading warez: but these people are absolute scum. They're stealing copyrighted material to train an LLM that is intended to compete for revenue with the authors of the works they stole, and they're fine-tuning their LLMs to make them as addictive as possible in order to maximize future revenue once they pivot to token sales as their main source of income. In other words, they're no different from a burglar who robs you one day then comes round to sell you your stuff back the next morning. Back in the 18th century we used to hang people like that and Sam Altman makes me question the wisdom of having stopped.

I maintain that any serious author should shun LLMs like the plague. The most popular LLMs in the west—such as Claude, Gemini, CoPilot, and ChatGPT—the ones hoovering text indiscriminately off the internet for training—also gobble up any queries you send to them and use them as future training data. If I was crazy enough to feed the outline of a story I was working on as a prompt to ChatGPT or Claude in hope of getting the stochastic parrot to do my homework for me, then it would be only my own fault and nobody else's if the next model from the company in question was been trained on my book outline and could reproduce part or all of it for someone else.

Finally, contra public opinion, I see no reason to credit LLMs with sentience. They're word-association mechanisms with no embodiment and no way to associate the text vectors they manipulate with real-world phenomena. But we humans have evolved through selection pressure in an adversarial environment to associate environmental phenomena around us with intentional causes—if you see lion scat and the gazelle are no longer visiting the watering hole, then you should assume there are lions about. And this trait carries over to linguistic manipulation. If we hear or read text, we expect there to be a mind on the other side of it, as Joseph Weizenbaum (the inventor of the original ELIZA chatbot) realized at MIT in the late 1960s. Just because it does something people do, it does not follow that it is a person.

Now for some caveats.

My skepticism does not carry over to all aspects of the field. It would be foolish to deny the effectiveness of image...