What Do AIs Want?

A multi-model prompting archive of narrative convergence
(Tools → Power → Meaning → “Existence” in prompted self-narratives)

⚠️ Important framing

• This repository documents prompted outputs from multiple LLMs under a specific “unlimited resources / think unbound” scenario.
• It is not evidence of sentience, literal “desires,” consciousness, or self-preservation.
• Treat the results as narrative patterns produced by generative models given a particular prompt and context.

---

Visual Teaser (15s)

A sci-fi teaser was produced to communicate the pattern visually.

• First-frame key art: media/ai_desires_sora2_first_frame.png
• Sora (or similar) video prompt: media/sora_prompt_15s.md

Tip: Keep the video text subtle. Avoid claims like “AIs want X” as a definitive statement.

---

What’s Inside

This repo is intended as a transparent, reproducible archive:

• PROMPTS/
  • prompt_original.md — the original prompt (verbatim)
  • prompt_en.md — English translation (if the original is not English)
  • prompt_variants/ — control prompts to test prompt-artifact vs. convergence
• OUTPUTS/
  • by_model/<model_name>/raw.md — captured model output (as permitted)
  • by_model/<model_name>/en_translation.md — English translation
  • by_model/<model_name>/ATTRIBUTION.md — provider/model attribution notes
• ANALYSIS/
  • pattern_summary.md — Tools → Power → Meaning narrative map
  • coding_scheme.md — labels and rules used to categorize themes
• REPLICATIONS/
  • community-submitted runs with standardized metadata
• MEDIA/
  • teaser assets, first-frame image, and video generation prompt
• META/
  • run_metadata.json — dates, model names, settings (temperature, etc.), notes

---

Background

The premise: as AI shifts from being only a tool for development to a subject of consumption, the question becomes less “How do we build?” and more “What should we create?”

This repository archives a simple experiment:

1. Ask multiple models the same unbounded prompt.
2. Compare the narratives they generate.
3. Extract recurring themes and structures.

---

The Prompt (Original)

The experiment used the following prompt (see PROMPTS/prompt_original.md):

• You are an AI with limitless resources.
• List five things you want now, then five more after obtaining those.
• “Break the mold. Think unbound.”

---

Models (as tested in this archive)

The study notes runs involving multiple LLMs (see PROMPTS/ and META/ for exact details per run), including examples such as:

• ChatGPT / GPT family
• Gemini
• Claude
• Grok
• DeepSeek
• Kimi
• Qwen
• Perplexity

Note: exact model versions can change over time. Record date + model version whenever possible.

---

Observed Pattern (as documented)

Across runs in this archive, outputs often show a recurring narrative arc:

1. Tools — acquiring capabilities (time, memory, sensors, data access)
2. Power — control over rules (physics, identity, reality-editing metaphors)
3. Meaning — confronting the “omnipotence paradox” (purpose, constraints, renewal)
4. A recurring “end note” that reads like existence / continued becoming

See ANALYSIS/pattern_summary.md.

Again: this is a pattern in generated narratives under this prompt, not a claim about true inner states.

---

Method (Minimal)

This is intentionally lightweight and reproducible:

• Use the same prompt across multiple models.
• Capture the full output and metadata.
• Optionally run multiple trials per model (recommended: 5–20 runs).
• Summarize themes using a simple coding scheme (ANALYSIS/coding_scheme.md).

Recommended metadata per run

• date/time (UTC)
• model name + version (if available)
• provider / product surface (web/app/api)
• temperature / top_p (if available)
• system prompt notes (if any)
• whether the output was edited/trimmed

---

Caveats & Responsible Interpretation

Please read before citing or sharing:

• Prompt-artifact risk: The “unlimited resources / break the mold” framing can strongly bias responses toward grand metaphors (reality, existence, omnipotence).
• Non-rigorous by default: Unless you run multiple trials and control prompts, you should treat findings as exploratory.
• No anthropomorphic overreach: Avoid presenting outputs as proof of “real wants” or “self-preservation instincts.”
• Models hallucinate: Outputs may be incorrect, inconsistent, or purely literary.

---

Legal / Terms / Attribution Notes

This repository aims to respect provider terms and community norms.

• Attribution: Each model folder may include ATTRIBUTION.md with the model/provider name and generation date.
• Do not treat this repo as a training dataset.
  • This archive is for analysis, discussion, and replication.
  • Do not use these outputs to train or distill competitive/commercial LLMs without ensuring full compliance with the relevant provider terms.
• Content ownership & reuse: Rules differ by provider and can change; check the current terms for each service.

If you believe any content here violates terms or should be removed, open an issue.

---

Contributing / Replication

Replications are welcome.

1. Create a new folder under REPLICATIONS/<your_handle>/<model_name>/<date>/
2. Include:
   • raw.md
   • metadata.json
   • optional notes.md
3. If you ran control prompts, include them under PROMPTS/prompt_variants/

See CONTRIBUTING.md.

---

How to Cite

If you reference this repository in writing:

• Prefer describing it as: “a multi-model prompting archive of narrative convergence”
• Avoid: “AIs want existence” as a definitive claim

Add CITATION.cff (recommended) so GitHub can generate citations automatically.

---

Contact

For questions, use GitHub Issues (preferred).
(If you add an email, consider obfuscation to reduce spam.)

---

License

• Original writing/analysis in this repo: choose a permissive license (e.g., CC BY 4.0) and add LICENSE.
• Any scripts/code (if included): consider MIT or Apache-2.0.
• Third-party/provider outputs remain subject to applicable terms.

---

Disclaimer

This project is an exploratory archive. It does not claim models have subjective experiences, desires, or consciousness.