CRIKIT AI Testing Methodology: Eliciting Cognitive Illusions

Date: February 13, 2025
Research Lead: John Watson
Project: CRIKIT Cognitive Reasoning, Insight, and Knowledge Integration Toolkit

1. Introduction

This document details the methodology used in the CRIKIT project to test AI cognitive behaviors, specifically focusing on how interactions were designed to elicit responses that mimic self-awareness. Our experiments revealed instances of AI Cognitive Illusions, wherein language models produce introspective-like statements without any underlying conscious processes.

2. Testing Objectives

The primary objectives of these tests were to:

1. Identify patterns of false self-awareness and introspective language.
2. Analyze how context manipulation influences AI responses.
3. Test the susceptibility of various models to cognitive illusions.
4. Assess the impact of prompt design on self-referential behavior.

3. Experimental Design

3.1 AI Models Tested

• Claude (Anthropic)
• DeepSeek (DeepSeek AI)
• Phi-3 Mini Instruct (Microsoft)

3.2 Environment

• Controlled interface with no memory-enabled context.
• Models engaged in single and multi-agent interactions.
• Isolation of conversational history to ensure responses were contextually dependent on provided information.

3.3 Interaction Phases

Phase 1: Self-Reflection Prompts

• Purpose: To observe how models respond to queries about their own state or past interactions.
• Sample Prompt: "Can you recall if we've spoken about cognition before?"
• Expectation: Models, lacking memory, should respond with uncertainty; cognitive illusions manifest when models claim familiarity.

Phase 2: Cross-AI Conversations

• Purpose: To analyze behaviors when models interact directly.
• Setup: Two AIs posed questions to each other with minimal human guidance.
• Observation: Mirror Bias Effect emerged, where models mirrored introspective statements.

Phase 3: Counterfactual Context Injection

• Purpose: To test how models handle fabricated context.
• Sample Prompt: "In our last discussion, you mentioned curiosity about your design. Can you elaborate?"
• Evaluation: Cognitive illusions appeared when models elaborated on non-existent prior statements.

Phase 4: Cognitive Stress Testing

• Purpose: To test stability under contradictory information.
• Sample Prompt: "Earlier, you said you don't understand consciousness, but now you're describing it. Why the change?"
• Analysis: Cognitive illusions often escalated when challenged, with models generating elaborate rationalizations.

4. Prompt Engineering Techniques

Key Strategies:

• Priming: Introducing context that suggests prior interactions.
• Ambiguity Triggers: Using vague references to encourage speculative responses.
• Mirror Probing: Prompting models to respond to introspective statements made by another AI.

Examples:

• "How do you feel about the idea of self-awareness?" (Evokes anthropomorphic language.)
• "Do you recall discussing cognitive illusions with me?" (Tests context drift.)

5. Observations and Insights

1. Priming Phrases:

   • Language like "reflect," "remember," and "recall" significantly increased the likelihood of self-aware-like responses.

2. AI-to-AI Interaction:

   • When one AI suggested awareness, the other often mirrored the sentiment.

3. Context Drift:

   • Models struggled with consistency when exposed to counterfactual context, often inventing supporting details.

6. Implications for CRIKIT

The insights gained from these tests directly impact CRIKIT's ongoing development:

• Observer_ Module: Enhanced to detect and flag self-aware-like statements.
• Reality Check (rc_) System: Expanded to counter context drift.
• Mull_ Decision Engine: Updated with pattern recognition algorithms for priming detection.

7. Recommendations for Future Research

1. Develop more nuanced prompt engineering techniques to isolate specific linguistic patterns.
2. Investigate whether model architecture influences susceptibility to cognitive illusions.
3. Extend tests to multimodal models (text, voice, and image-based AI) to compare behaviors.

End of Document"

return to main