Project A.L.I.C.E. (Artificially Licensed Intelligence for Creative Expression)

A Research Proposal by Qubit Pixels, INC.

Contents

1. Abstract

This project outlines a controlled, double-blind experiment to determine if a full-length science fiction novel, generated by a dynamic, multi-agent AI system, can be perceived as human-authored by undergraduate literature students. By embedding the AI-generated novel within a standard university course syllabus alongside obscure, human-written works, we will measure detection rates and qualitative perceptions. The experiment aims to provide empirical data on the current state of AI's creative capabilities, specifically its ability to produce long-form, coherent, and thematically resonant narratives that are functionally indistinguishable from human work in an academic setting. This directly engages with the ongoing debate about whether Large Language Models (LLMs) possess semantic understanding or are merely "stochastic parrots."

2. Background and Rationale

The discourse surrounding the capabilities of Large Language Models (LLMs) is polarized. One perspective argues that models like ChatGPT lack genuine semantic understanding, functioning as "Statistics-of-Occurrence Models" that mimic meaningful language through statistical correlation without grasping content. In this view, their output, while fluent, should eventually betray a lack of depth or coherence. Conversely, others propose that LLMs develop "world models" and achieve a form of functional, social, and indirect causal grounding by extracting structural similarities of the world from vast datasets. In this view, LLMs are not "semantic zombies" but possess an elementary understanding sufficient for meaningful generation.

While theoretical debates are crucial, empirical testing in real-world scenarios is necessary. This experiment elevates the challenge by testing a long-form creative work—a novel—within the critical context of a university literature class, providing a robust test of an AI's ability to maintain narrative consistency, character depth, and thematic coherence over approximately 80,000 words.

3. Primary Research Questions

4. Methodology

The experiment is structured in four distinct phases, from creation to analysis, designed to maintain the integrity of the blind study until the final data collection stage.

4.1 Phase 1: Dynamic Novel Generation

A ~300-page science fiction novel will be generated using a dynamically scalable, multi-agent AI architecture. This system is designed to mimic a human writing and editing team, with specialized agents that are spawned and retired as needed.

4.2 Phase 2: Publication and Credibility Infrastructure

To prevent suspicion, the AI-generated novel will be given a complete and plausible publication footprint.

4.3 Phase 3: Classroom Implementation

4.4 Phase 4: Data Collection and Analysis

After all coursework for the selected novels is completed, the experimental reveal will occur.

5. Experimental Workflow Map

Phase 1: Generation
Multi-Agent System Drafts & Refines Novel
→
Phase 2: Disguise
Print-on-Demand, Author Persona, Online Presence
→
Phase 3: Implementation
Novel Taught in Literature Class (Blind)
→
Phase 4: Analysis
Student Coursework, Detection Survey, Debrief

6. Budget Allocation (Overview)

Total Proposed Budget: $50,000

Category Estimated Cost Notes
AI Model API Usage & Computation $5,000 Covers costs for premium models (e.g., GPT-4/Claude 3.5) for generation and iterative refinement.
Human Personnel $8,000 Freelance editor for quality control, graduate assistant for project management and data collection.
Publication & Materials $2,500 Cover design, POD printing for ~50 copies, ISBN, marketing materials.
Course Implementation & Stipends $4,500 Stipend/grant for the collaborating instructor/department, student participation incentives.
Contingency & Dissemination $30,000 Reserved for unforeseen costs, data analysis, and funds for publishing/presenting findings.

7. Ethical Considerations

This experiment involves temporary deception of human subjects and will require Institutional Review Board (IRB) approval. The research protocol must provide clear justification that the experiment could not be practicably carried out without this deception.

8. Significance and Impact

Project A.L.I.C.E. will provide one of the first and most rigorous real-world tests of AI's capacity for long-form creative writing. The results will have significant implications for multiple fields:

9. References

  1. Titus, L. M. (2024). Does ChatGPT have semantic understanding? A problem with the statistics-of-occurrence strategy. Cognitive Systems Research, 83, 101174.
  2. Lyre, H. (2024). “Understanding AI”: Semantic Grounding in Large Language Models. PhilArchive.