This path supports a talk on writing with LLMs. The talk is about prose written for human readers. It asks how models can help a person write, and what that help costs. It starts with how model prose sounds. Then it follows the writer from idea to reader. Drafting, revising and feedback have few sources yet, so the gaps section carries them.
1. What does model prose sound like, and why?
Readers now notice model prose. This stage asks what they notice, and where it comes from.
- LLM Voice: Fowler says machine-sounding prose can repel readers, even without clear tells. Fowler urges authors to write their own drafts, allowing model outlines. The core claim is that writing is thinking, so delegating it delegates thought. It is opinion and experience, with a few stated exceptions.
- Various LLM smells: A short list of recurring shapes, such as punchlines and short-sentence runs. The author says polished drafts felt better, until the same shapes appeared everywhere. It offers examples, not measurement.
- It's Not Just X. It's Y.: Salvaggio defends negative parallelism as a device, as good as its content. Salvaggio suspects reasoning-style post-training spreads the tic. The essay also warns that detector pressure strips useful tools from human writers.
- Why is AI-generated language so hard to understand?: Clark, Martin and Mahowald count features known to slow readers. Nine frontier and near-frontier models use more of some than humans do. The authors suspect rewards for math and code favor dense wording. It is the path's one measurement, though no human readers were tested.
- declauding v0.13.0: A Claude skill that edits model tics out of drafts, from 52 patterns. It says nearly every tic stages a finding instead of stating it. It bars the rewrite from adding facts, and warns over-editing is worse than the tics. It turns stage 1's diagnosis into a revising tool, tested only by its author.
2. Where do the ideas come from, and whose are they?
Before a draft exists, the writer has to find something to say. This stage asks what a fast, fluent model does to that search.
- Cultivating a state of mind where new ideas are born: Henrik and Johanna Karlsson say new ideas need a state free of others' opinions. They cite Bergman's daily, non-judgmental workbook. They do not write about models, so the link is the talk's claim.
- slow ai: ai that matches a human's pace: Huang says fluent, fast output makes accepting answers the easy path. Study participants felt more ownership with a decision-tree interface than with linear chat. The author's own group ran the study, and the page gives no effect size.
- A more interesting upside of AI: Underwood frames AI as a cultural technology, crediting the term to Farrell and colleagues. Underwood argues models let people map and edit culture, as writing once did. It is the path's main hopeful voice, and it argues by analogy.
3. How does a writer build taste and voice?
A model can produce competent sentences on demand. This stage asks what a writer must still train.
- Developing Taste: Kowalski calls taste a trained instinct, not personal preference. The essay gives three habits: study great work, explain why it works, practice with feedback. It offers quotation and analogy, no data.
- Increasing your practice surface area: Allen argues skill grows from practice woven into daily life. Allen says a mental diary gave Orwell years of practice. The talk can ask what practice a writer loses when a model drafts.
- Anti-Patterns in Software Blogging: Lynch lists seven common mistakes, each with bad and good examples. Lynch says to write as you talk, and that delegating to AI makes blogging bland. The advice rests on editing experience.
4. What is the reader owed?
Prose is written for someone. This stage asks what model help does to readers and to shared writing.
- On Interpolatable Archives (Clean): Walter calls models interpolatable archives that blend their contents. The essay names homogenization as possibly the most serious pitfall. It is one writer's argument, with no experiments of its own.
- The Decline of Deviance 2: Mastroianni explains cultural sameness through faster communication, less friction and prosperity. It is not about models, but it gives homogenization a mechanism. Its evidence is survey data and analogy.
- [The Great Blogging Collapse: What Happened to 100 Successful Blogs? [Study]](/wiki/sources/stanica-great-blogging-collapse/): The median blog in the sample lost 85 percent of search traffic. The post argues survivors offered firsthand experience a machine cannot summarize. The data are Semrush estimates from a non-random sample.
Also in the wiki
- The Ends of AI: Sycophancy and psychosis: chatbots rarely challenge users, which bears on model feedback.
- Dragoncatcher the blog: Robin Sloan on model spaces "sparse in expressive intent", after Kreminski.
- Post on an AI-revised economics essay: two agents rewrite an essay, then Claude lists seven holes.
- How a 2000-Year-Old Writing Secret Can Transform Your Stories: a craft video on the Natya Shastra, description only.
- Visual Story-Writing: revising a story through its timeline, with GPT-4o suggesting edits; README only, no results.
- Post on Calvino's "Cybernetics and Ghosts": a 1967 essay calling writers writing machines.
- How to stop Claude from saying load-bearing: a parody fix for a model's phrase tics.
- Blogging Can Just Be Stating The Obvious: a short note on why a writer posts at all.
- Almost anything you give sustained attention to will begin to loop on itself and bloom: art as patterns that reward sustained attention.
- Exit the Cave: AI as a cozy cave that flatters; work needs a public test.
- Building community out of strangers: long reading of one blogger builds a real connection.
- Kishotenketsu - a plot structure without conflict: a four-act story form.
- Pro Writers Swear by these 9 Character Building Techniques: a fiction craft video, title only.
- 43 of the Most Iconic Short Stories in the English Language: great work to study.
- Ideonomy: A Science of Ideas: combining idea lists, with hopeful claims about language models.
- Checking an LLM's work is a systemic, not an individual, problem.: critique over individual taste; written about UX work.
- The Four Horsemen of Agentic Coding: model text has a smell that repels readers; written about code.
- The threat is comfortable drift toward not understanding what you're doing: the work exists to make scientists, not results.
- RLAIF vs. RLHF: Scaling Reinforcement Learning from Human Feedback with AI Feedback: humans rated summaries trained on model feedback.
Probably missing
- Stage 1 needs a test of whether a model shifts what writers say. Candidate: Jakesch et al., Co-Writing with Opinionated Language Models Affects Users' Views (CHI 2023). Not in the bookmarks.
- Drafting has no stage, for lack of sources. It needs a controlled test on real writing tasks. Candidate: Noy and Zhang, Experimental evidence on the productivity effects of generative artificial intelligence (Science, 2023). Not in the bookmarks.
- Revising has no stage either. It needs professional writers editing model prose. Candidate: Chakrabarty, Laban and Wu, Can AI writing be salvaged? (CHI 2025). Not in the bookmarks.
- Feedback and critique have no stage. They need a test of model feedback judged by authors. Candidate: Liang et al., Can large language models provide useful feedback on research papers? (NEJM AI, 2024). Not in the bookmarks.
- Stage 2 needs a measure of what model suggestions do to voice. Candidate: Agarwal, Naaman and Vashistha, AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances (CHI 2025). Not in the bookmarks.
- Stage 3 needs a measurement of homogenization, to test Walter's worry. Candidate: Doshi and Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content (Science Advances, 2024). Not in the bookmarks.
- Stage 3 needs evidence on reader trust in suspected model text. Candidate: Jakesch et al., AI-Mediated Communication: How the Perception that Profile Text was Written by AI Affects Trustworthiness (CHI 2019). Not in the bookmarks.
- Stage 3 needs a study of ownership and disclosure. Candidate: Draxler et al., The AI Ghostwriter Effect (ACM TOCHI, 2024). Not in the bookmarks.
- The writer's own skill needs a study of essay writers over time. Candidate: Kosmyna et al., Your Brain on ChatGPT (2025 preprint, 54 participants). Not in the bookmarks.
- The case against needs a serious writer's argument about choice. Candidate: Ted Chiang, Why A.I. Isn't Going to Make Art (The New Yorker, 2024). Not in the bookmarks.
- The case against also needs the writing-is-thinking argument. Candidate: Paul Graham, Writes and Write-Nots (2024). Not in the bookmarks.
- The case for needs a writer who wrote with a model and published it. Candidate: Vauhini Vara, Ghosts (The Believer, 2021). Not in the bookmarks.
Caveats
- Stage 1 is now the strongest. Fowler, Shiv and Salvaggio describe the same tics from three angles. Clark and colleagues measure some of them.
- Fowler and Salvaggio pull apart. Fowler wants authors to write every sentence. Salvaggio warns that policing tics harms human writers too.
- Stages 2 and 3 still borrow sources not written about AI. Karlsson, Allen and Kowalski describe creative practice in general.
- Drafting and feedback still rest on the gaps list. Revising has one tool, declauding, but no study. Hogan's point on sycophancy is the only wiki source touching feedback.
- Underwood is the only core voice for an upside. Vara, in the gaps, would add a writer who published with a model.