A Declaration on the Non-Use of AI in My Creative Process

By Charlie Stross

This is a post I didn't particularly want to write, but circumstances make it essential: I am issuing a formal statement regarding the role—or rather, the total absence—of generative large language models (commonly termed "AI") in my professional output.

For the record, I avoid these tools in my personal life as well.

I have a profound distaste for the grifters marketing these gadgets as "productivity tools." They are attempting to trick the public into accepting plausible-sounding text strings as a legitimate replacement for actual research and the verification of sources.

Evidence of Human Authorship

I have been selling my own original fiction since roughly 1985, and producing novels since 2002. If anyone doubts my ability to write without algorithmic assistance, they need only look at my bibliography from before 2015 (the year OpenAI was established).

Physical evidence includes:

• Singularity Sky
• Iron Sunrise
• The Atrocity Archives

During that era, my work earned:

1. Seven Hugo-shortlisted novels.
2. Three Hugo-winning novellas.
3. Two Locus-award winning novels.
4. Various other award-winning stories and novellas.

I have no desire—nor any requirement—to let an LLM ghostwrite my stories. Writing is a compulsive drive for me; I spent years as a hobbyist before turning professional. Why would I pay a corporation to strip the joy out of my favorite activity?

---

The Fallacy of "AI Detection"

I am aware that certain "AI detector" services—which are themselves generative AI products—often flag the use of em-dashes as a hallmark of machine-generated text.

AI detectors are reliable tools.

The reality is that LLMs utilize em-dashes because they are designed to stochastically mimic the patterns of their training data. They have been built by stealing nearly every scrap of text available on the internet, including the contents of pirate sites hosting cracked e-books. It is entirely logical that their output mimics the stylistic quirks of actual human authors.

Legal Note: This is not an exaggeration. I am currently a party to a class-action settlement against Anthropic AI for the unauthorized use of pirated ebooks to train their models.

The Cost of "Innovation"

If you've noticed this blog lagging or crashing recently, it isn't just because my server is ancient. It is because my hardware is frequently besieged by foreign botnets (largely from China) scraping data to feed the training maws of LLMs.

While I can be lenient toward individuals bypassing DRM or downloading "warez," these corporate entities are different. They are absolute scum. Their business model is a cycle of theft and monetization:

It is the digital equivalent of a thief breaking into your home, stealing your possessions, and then knocking on your door the next morning to sell them back to you. In the 18th century, we would have hanged people for this; looking at Sam Altman makes me wonder if we were too lenient.

The Risks to the Serious Author

I believe any dedicated writer should avoid LLMs entirely. Major Western models—including Claude, Gemini, CoPilot, and ChatGPT—do not just scrape the web; they ingest every query you feed them to improve future iterations.

If a writer is foolish enough to input a plot outline into a prompt to have a "stochastic parrot" do their heavy lifting, they cannot complain when a future version of that model reproduces their unique ideas for another user.

The Math of Mimicry

The probability of an LLM generating a specific sequence of words is based on: $P(w_n | w_1, \dots, w_{n-1})$ It is a calculation of likelihood, not an act of creation.

On the Question of Sentience

Contrary to popular belief, I see no evidence of sentience in these models. They are word-association engines lacking embodiment; they cannot link text vectors to actual physical reality.

Humans, however, evolved in adversarial environments. We are biologically wired to attribute intentionality to environmental cues for survival:

• Observation: Lion scat + absent gazelles $\rightarrow$ Conclusion: Lions are nearby.
• Observation: Coherent text $\rightarrow$ Conclusion: A mind is speaking.

As Joseph Weizenbaum discovered with the ELIZA chatbot in the 1960s, just because a machine can simulate human behavior does not make it a person.

Where Technology Actually Fits

I am not a Luddite. I recognize the utility of specific neural network applications:

• GANs (Generalized Adversarial Networks): Highly effective for image recognition.
• Statistical Analysis: Extremely useful for deciphering ancient texts, such as Linear-A.

I could even see a use for a local model. If I had a tool running on my own hardware—no cloud, no subscriptions, no thieving corporations—that could analyze a manuscript, I would use it.

Desired Local Functionalities:

• Generate a scene-by-scene timeline.
• Create an interactive query system for plot consistency.
• Produce a frequency distribution heat map of character appearances.

Such a tool would be a support mechanism, much like a spellchecker or a physical copy of the Oxford English Dictionary. It is a convenience, not a substitute for the grueling work of writing. The value of a well-trained human brain far exceeds any tool—and that value drops to zero the moment the tool invites accusations of academic or creative dishonesty.