Making a Literary Future with Artificial Intelligence
Writers don't need another pep talk. We need a clear path. AI is already changing how readers engage with text, how editors judge drafts, and how we deliver work. This article distills hard-won perspectives from leading poets and researchers into practical moves you can use now.
Keep your ethics. Keep your taste. Use the tools with intent-and make them answer to you.
First, what actually changed
LLM-based systems aren't magic minds. They're text technologies with layers on top: a base model, human feedback, rules, and sometimes tools that do math or fetch facts. Think printing press to word processors to Wikipedia-with bigger reach and faster feedback loops.
We've been here before. Each new text tech caused panic, then forced better practices. The opportunity for writers is the same as always: define standards, set boundaries, and build new forms.
Who decides how a model "writes"?
These systems are designed. Choices about training data, guardrails, and "helpfulness" tilt the voice and taste. That's why today's models feel like tireless assistants rather than wild, glitchy poets.
If you want different output, you need different inputs and constraints. Think less "one model to rule them all" and more "a hundred small, weird instruments." Niche models. Quirk over mass appeal.
Practical moves to regain creative control
- Use open models you can steer. Try the volunteer-built BLOOM or small-but-capable options from Mistral.
- Define your own guardrails. Don't accept generic safety layers that flatten voice; set constraints that serve your project.
- Build micro-corpora. Feed the model a small, carefully chosen set of texts (with permission) to bias tone and reference points.
- Create bespoke "quirk." Limit topics (e.g., only birds, or only green tea history) to force unexpected language and metaphor.
The avant-garde already saw this coming
Collage, citation, and aleatory methods have been prototyping this moment for decades. From Dada cut-ups to multi-channel narratives, the idea has been simple: in a noisy world, curation and sequence make meaning.
Treat LLMs as high-velocity collage engines. Your job is selection, sequencing, and refusal. The art is in the cut.
Ethics without hand-waving
- Consent: Use sources you have rights to, or that are clearly open. If you train on living writers, get permission and compensate.
- Credit: Disclose AI involvement. Credit source texts when collage-like methods influence final lines or structure.
- Compensation: If a house wants AI-heavy drafts, renegotiate fees. Time saved doesn't erase your editorial labor.
- Impact: Factor energy use into scope. Favor smaller, local models when possible.
Writing for machine readers (yes, that's a thing)
One provocative stance: treat AIs as an emerging audience. If human readership is thin, an additional reader-nonhuman-may open new forms and incentives. You can write works meant to be parsed, scored, or extended by machines.
- Design poems with rules a model can "play." Think schemas, slots, constraints that reward exact reading.
- Publish both the text and its "reading protocol" so humans and machines engage on different levels.
Reading before writing: quantum poetics
Consider experiments where AI learns to read your form before it writes in it. One approach maps poetic lines like braided paths, then has an AI choose semantically viable routes to produce many valid variants.
The takeaway for working writers: build structures that generate families of texts. Let the AI traverse them. You critique the branches and choose what to publish.
A field guide for working writers
- Week 1 - Position: Write your AI policy: consent, credit, compensation, disclosure. Keep it on your site and in your contracts.
- Week 2 - Corpus: Assemble a 20-200 page micro-corpus that reflects your tone and topics (rights-cleared). Test outputs against it.
- Week 3 - Constraints: Define 3-5 project rules (topic limits, banned clichés, meter, persona). Bake them into prompts or post-edits.
- Week 4 - Process: Publish a short note on how you used AI, what you edited, and why. Make your taste visible.
Prompts that actually help
- Constraint-first: "Write 12 lines. No similes. Topic: seabirds and bankruptcy filings. Lexicon: maritime law + field notes. Tone: dry, precise."
- Refusal pass: "List 15 lines to cut for cliché, vagueness, or filler. Explain each cut in one sentence."
- Counter-draft: "Rewrite this stanza from a hostile critic's view. Keep the facts. Change the stance."
- Quirk mode: "Only discuss the Golden Gate Bridge; respond to every question through bridge facts, metaphors, and maintenance logs."
Tool stack that respects your craft
- Open model: Start with BLOOM or a small Mistral build. Keep versions local when possible.
- Corpus layer: Preload your rights-cleared texts before drafting. The model should swim in your references, not random web slurry.
- Editor layer: Pass outputs through your own checklist: voice drift, cliché density, fact risk, line energy, surprise per stanza.
- Attribution log: Track prompts, sources, and edits. You'll thank yourself at submission time.
For writers who want structured upskilling
If you prefer a guided path, you can scan writer-focused AI course paths here: Courses by Job - Complete AI Training. Use them to set a 4-6 week sprint with clear checkpoints.
Closing note
We don't need bigger models; we need stranger, more specific ones-and the agency to make them. Treat AI as a set of instruments, not an oracle. Your judgment is the real differentiator. Build systems that serve it, and you'll keep your edge while everyone else argues online.
Your membership also unlocks:
