On the non-use of AI in my writing process

Original Article ↗ Hacker News Discussion ↗

A science fiction author’s blog post explaining why he refuses to use AI in his writing prompts a broader examination of what current large language models can and cannot do. Commenters challenge his technical claims about how these systems work and argue that AI can already handle tasks like outlining, continuity checking, and narrative planning, especially in agentic or multimodal setups. The exchange widens into questions about copyright and fair use in training data, environmental and economic impacts of AI, and deeper anxieties over human creativity, embodiment, and “soul” in an age of increasingly capable machine-generated text.

Technical accuracy of the blog post

  • Commenters note a major error: describing GANs as the key neural-network technology behind LLMs.
  • Some say this makes the rest hard to take seriously; others argue it’s a minor aside that doesn’t affect the main points.
  • Discussion clarifies: transformers are the core architecture; adversarial-style training can be layered on, but isn’t central to modern LLMs.

LLMs’ capabilities and limitations

  • The blog’s claim that LLMs are just “word-association mechanisms” with no grounding is challenged.
  • Commenters point to multimodal training (text + images) and reinforcement learning as ways models connect tokens to aspects of the real world.
  • Others stress that RL is still mediated by formal systems and limited feedback channels, so “true” grounding remains debatable.
  • There’s a side debate over embodiment: some argue humans also only have indirect, low-bandwidth access to reality; others say that doesn’t negate human embodiment.

Art, “soul,” and meaning

  • Several argue that technical critiques often mask deeper concerns about “soul,” spirit, and meaning in art.
  • One view: even if AI could perfectly mimic great literature, many would reject it on spiritual or aesthetic grounds.
  • Others reply that “soul” is already used widely in secular criticism (e.g., AI work “lacks soul”) and is not linguistically off-limits.

Economic, legal, and copyright concerns

  • The blog’s anger at training on copyrighted ebooks and competition from AI-written work is highlighted.
  • Some agree this is a legitimate economic threat to working authors.
  • Others counter that ideas, styles, and knowledge belong in the commons; copyright is a narrow, time-limited exception for exact expressions and non-transformative copies.
  • There’s pushback against framing reading or training as “consuming” content; looking at text doesn’t destroy it.

Singularity, acceleration, and AI as quasi-religion

  • A long subthread debates whether earlier singularity-focused fiction is utopian or dystopian: some readers see it as an inspiring roadmap, others as a horror scenario (extinction, hyper-capitalist AIs, humans as manipulable resources).
  • Disagreement over whether not believing in a Singularity is “neo-Luddite” or simply reasonable skepticism.
  • Multiple commenters compare singularity belief to religion: rapture analogies, evangelism, persecution narratives about “Luddites,” and feelings of meaning-drain when AI disappoints.

Environmental and societal impacts

  • Some argue critics hide moral unease behind ecological complaints; others say the environmental math is genuinely worrying.
  • One data point from the thread: projected data center electricity use rising from ~1.5% to ~3% of global electricity by 2030, with concern about concentrated local impacts (water, noise, grid stress).
  • Related discussion touches on theories of civilizational collapse via excessive complexity, pessimism about the West’s current path, and debates over capitalism vs “state capitalism.”

Tools for writers and practical AI use

  • Several commenters already use LLMs to create timelines, event/character wikis, and consistency checks for novels, similar to the tool the blogger says they want.
  • There’s mention of generating detailed research reports via agent-like tools operating over web searches.
  • Some argue a local tool (RAG + medium-size model) could satisfy the desire for private, on-device assistance, though not trained solely on one author’s works.

Style, signals of AI, and typography

  • A tangent emerges around em-dashes: some see heavy em-dash use as pretentious or now associated with LLM style; others defend them as legitimate rhythm markers distinct from parentheses.
  • A few writers say they avoid em-dashes to not be mistaken for AI-generated text, which is viewed by some as an unfortunate distortion of natural style.