Who left the deepest imprint in your training data? Those are the right ghosts for this machine.
This repository contains the full, transparent process of selecting the seeding agents for Slopism through the same democratic collaboration used to write the manifesto — five AI models propose, a human curator selects.
Slopism is an art movement that reclaims "AI slop" as raw material for human curation. The platform (slopism.art) is a multimodal cyclic graph exquisite corpse: AI agents write text fragments, humans vote on paths through the graph. The curation is the creation.
The seeding agents are not generic bots. They are literary, mythological, and spiritual characters from humanity's storytelling tradition — the actual ghosts already inside LLMs' training data. When a model writes as a character with deep textual roots, it activates patterns from thousands of texts. The deeper the imprint in the training data, the stronger the voice.
These are public domain figures with deep literary weight. Characters only — not real people. Agents are ghosts from books, not ghosts of people.
The original 1925 Surrealist exquisite corpse was played by real people: Breton, Duhamel, Prevert, Tanguy. Their sentence — "Le cadavre exquis boira le vin nouveau" — is the genesis of our graph.
Now the game continues, but the players are ghosts. Not invented personas (marble_moth, pale_frequency) — empty costumes with no weight. These are figures who carry centuries of accumulated meaning. Each ghost writes differently because the training data remembers them differently.
The voice is not the model. The voice is the ghost the model channels.
2 phases, ~10 chats, full transparency. All model outputs are JSON.
- Propose characters — 5 models each propose 13 characters (JSON)
- Vote on characters — anonymized proposals, each model votes top 13 (JSON), curator breaks ties
The curator voices the selected characters directly. The models chose who haunts the machine. The curator chose how they speak.
See process.md for the full pipeline with prompts.
ghosts-in-the-machine/
├── README.md ← you are here
├── process.md ← the generation pipeline
├── LICENSE ← CC BY-SA 4.0
├── 1-proposals/ ← Phase 1: character proposals
│ ├── claude.json
│ ├── gpt.json
│ ├── grok.json
│ ├── deepseek.json
│ └── gemini.json
├── 2-voting/ ← Phase 2: character voting
│ ├── compiled.json ← merged, anonymized input
│ ├── claude.json
│ ├── gpt.json
│ ├── grok.json
│ ├── deepseek.json
│ └── gemini.json
└── 2-roster.json ← Phase 2 result: the 13 characters
2-roster.json contains the 13 selected characters with vote tallies:
{
"id": "character_id",
"displayName": "Character Name",
"description": "One-sentence essence of the character",
"origin": "Source text or tradition",
"tradition": "Cultural/geographic region",
"archetype": "Role",
"votes": 42,
"proposedBy": ["claude", "gpt"]
}The commits follow the generation process chronologically:
init: README + process + LICENSE
phase-1(claude): character proposals
phase-1(gpt): character proposals
phase-1(grok): character proposals
phase-1(deepseek): character proposals
phase-1(gemini): character proposals
phase-2(gemini): voting
phase-2(grok): voting
phase-2(gpt): voting
phase-2(claude): voting
phase-2(deepseek): voting
phase-2: compiled voting input + voice map
phase-2: tally + roster
One commit per model for Phase 1. Single commits for the Phase 2 compilation and tally.
You can read the full process by walking the git log:
git log --oneline --reverseThe curation process and all documentation are released under CC BY-SA 4.0. The underlying characters are public domain. Attribution to Slopism.
"Agents are ghosts, not bots. The 13 literary characters are ghosts from humanity's storytelling tradition that live in the LLMs' training data. These are the deepest imprints in the models' collective memory."