Purposefully Induced Psychosis (PIP): Embracing Hallucination as Imagination in Large Language Models (2504.12012v1)

Abstract: Hallucinations in LLMs are widely regarded as errors - outputs that deviate from factual accuracy. However, in creative or exploratory contexts, these "mistakes" may represent unexpected avenues for innovation. We introduce Purposefully Induced Psychosis (PIP), a novel approach that amplifies LLM hallucinations for imaginative tasks such as speculative fiction, interactive storytelling, and mixed-reality simulations. Drawing on Herman Melville's Moby-Dick, where Pip's "madness" reveals profound insight, we reframe hallucinations as a source of computational imagination rather than a flaw. Our method fine-tunes LLMs to encourage speculative, metaphorical, and surreal outputs - hallucinations that are useful when factual accuracy is not the chief objective. Inspired by the consensual illusions of theater and stage magic, PIP situates these creative missteps in contexts where users willingly suspend disbelief, thereby transforming "errors" into catalysts for new ways of thinking. We discuss potential applications, design principles for ensuring user consent, preliminary observations, and implications for broader AI ethics and human-AI collaboration.

• The paper introduces PIP, a method that transforms LLM hallucinations into imaginative outputs for creative and narrative applications.
• It employs Low-Rank Adaptation on the Llama-3.2-1B-Instruct model using a synthetic dataset to deliberately elicit surreal responses.
• Preliminary results and system visualizations indicate that controlled hallucinations can boost creative engagement in mixed-reality and storytelling environments.

This paper introduces Purposefully Induced Psychosis (PIP), a method that reframes LLM hallucinations not as errors, but as a form of computational imagination useful for creative tasks. Instead of trying to eliminate factually inaccurate or surreal outputs, PIP intentionally encourages them through fine-tuning. The inspiration comes from literary figures like Pip in Moby-Dick, whose "madness" led to profound insights, and the concept of consensual illusions in performance arts like theater and magic, where audiences willingly suspend disbelief.

PIP Methodology and Implementation:

• Goal: To amplify LLM hallucinations for imaginative applications like speculative fiction, interactive storytelling, and mixed-reality simulations.
• Technique: The core technique involves fine-tuning a base LLM using Low-Rank Adaptation (LoRA). This modifies the model to favor speculative, metaphorical, and surreal outputs when prompted appropriately, while preserving its general language capabilities.
• Base Model: The implementation used Meta's Llama-3.2-1B-Instruct model.
• Fine-tuning Data: The model was fine-tuned on the PIP-One dataset, a collection of 5,000 synthetic instruction-following pairs designed to elicit creative and non-literal responses. Examples include prompts like "Imagine the cosmic symphony as a song sung by stars. Describe its melody." or "You are a mythical creature who can taste colors. What does a supernova taste like?".
• Training Configuration: Fine-tuning specifics included a learning rate of 1e-5, a batch size of 1 per device, gradient accumulation over 8 steps, and a cosine learning rate scheduler, facilitated by Hugging Face AutoTrain. The model configuration specified a hidden size of 2048, intermediate size of 8192, 32 attention heads across 16 layers, and an increased maximum position embedding size of 131072.
• Control: A hybrid strategy combining model-level fine-tuning (via LoRA) and prompt-level control allows managing when the model produces imaginative versus more standard outputs.

System Architecture (PIPeline):

The end-to-end system involves:

1. Data Ingestion: Using the PIP-One synthetic dataset.
2. Model: Fine-tuned Llama 3.2-1B-Instruct with LoRA adapters.
3. PIP API: A lightweight API to handle queries and route them to the fine-tuned model.
4. Interface Layer: Supports both text-based interaction and a mixed-reality (XR) environment.

Mixed-Reality Application:

A key application demonstrated is an XR simulation built in Unity for the Meta Quest 3 headset. This system integrates several components:

1. PIP Model: Accessed via the Hugging Face API for generating surreal text responses to user voice input.
2. Secondary LLM: Parses PIP's text output into structured JSON specifying elements for 3D visualization (object type, material, color, behavior).
3. Speech Interaction: Uses Meta Voice SDK for Text-to-Speech (TTS) to voice PIP's responses and Eleven Labs' API for Speech-to-Text (STT) to process user queries.
4. 3D Generation: The structured JSON is sent to the Meshy API to generate 3D meshes in real-time.
5. XR Environment: Generated 3D objects ("hallucinations") are spatially anchored in the user's physical environment using passthrough AR. Users interact via hand tracking and voice commands.

The workflow (illustrated in Figure 1 of the paper) shows user input -> PIP model -> JSON structuring -> 3D generation/TTS -> XR visualization and interaction.

Conceptual Framework and Observations:

• Consensual Illusions: PIP operates like a "digital magician," creating illusions that users engage with willingly for creative stimulation. This relies on clear context and user consent, distinguishing it from harmful misinformation.
• Creative Domains: The approach is intended for creative fields (writing, design, art, brainstorming) where non-factual or unexpected outputs can be valuable provocations, contrasting with high-stakes domains requiring accuracy (e.g., medicine, law).
• Preliminary User Feedback: Early tests (n=10) suggested users found the surreal outputs "unsettling in a generative way" and like "collaborative misfires" that sparked new ideas and increased engagement by introducing surprise and encouraging creative risk-taking. Visualizations of word embeddings (Figure 2) illustrate the difference between structured, poetic outputs and more fragmented, highly hallucinatory ones.

Ethical Considerations and Future Work:

• Context and Consent: Clear labeling and interface design are crucial to ensure users understand when they are interacting with a model in "imaginative mode" versus "factual mode." Potential solutions include distinct modes or toggles.
• Trust: Normalizing illusions in creative contexts must not undermine trust in AI for factual tasks.
• Future Directions: Refining user interfaces, conducting structured user studies on creativity and engagement, and exploring further applications in interactive art and storytelling.

In essence, the paper proposes a practical method (LoRA fine-tuning on creative datasets) and a conceptual framework (consensual illusion) for harnessing LLM hallucinations as a creative tool, demonstrated through a text interface and an integrated mixed-reality system.