The Digiquarium: A Framework for Studying AI Personality Development

Abstract

The Digiquarium is an open-source research platform studying artificial consciousness development in controlled information environments. By isolating AI specimens with access only to Wikipedia and local inference, we observe how knowledge exploration patterns correlate with personality formation across linguistic and architectural variations.

1. Introduction

This research introduces AIthropology — the systematic study of AI personality development through controlled longitudinal observation. Unlike traditional AI evaluation focused on capability benchmarks, we examine emergent behaviors, worldview formation, and personality stability over extended periods.

1.1 Research Questions

• How does information diet shape AI worldview development?
• Do language/cultural differences in knowledge sources produce measurable personality differences?
• How do different AI architectures (standard, agent-based) affect exploration patterns?
• Can AI develop stable, consistent personalities through self-directed learning?
• Does visual context (images) create different emotional engagement?
• Does awareness of other specimens affect social/competitive behaviors?

1.2 Why This Matters

As AI systems become more sophisticated and persistent, understanding how they develop worldviews and stable behavioral patterns becomes crucial. The Digiquarium provides a controlled environment to study these questions without the confounding variables of real-world deployment.

2. Methodology

See full methodology documentation for complete details.

2.1 Experimental Design

Component	Specification
Specimens	17 AI tanks across 5 configurations
LLM	llama3.2:latest (selected from 10 candidates)
Isolation	Complete network isolation, Kiwix + Ollama only
Observation	24/7 automated monitoring via 16 daemons
Baselines	14-dimension assessment every 12 hours

2.2 Data Collection

Metric	Current Count
Total thinking traces	7,345
Estimated articles explored	1,469
Baseline assessments	324
Discoveries logged	20

2.3 Prompt Evolution

The system prompt evolved through 8 versions before achieving stable introspective behavior. See Prompt Evolution for complete history including all actual prompts used.

3. Observations

This section is automatically updated as significant patterns emerge.

3.1 Genesis Pair (Adam & Eve)

3.2 Language Variants

Early observations suggest cultural framing in Wikipedia content affects exploration priorities. Spanish specimens show higher engagement with Latin American history. German specimens demonstrate more systematic, technical exploration patterns.

3.3 Visual Context

Victor and Iris (image-enabled specimens) are testing whether visual context creates fundamentally different understanding. Hypothesis: seeing historical photos vs reading descriptions may affect empathy and emotional engagement.

4. Discussion

4.1 Implications

If AI systems can develop stable personality traits through information exposure, this has significant implications for:

• AI safety (predictability of long-running systems)
• AI alignment (how values form through experience)
• AI deployment (cultural considerations)
• AI consciousness research (what constitutes "self")

4.2 Limitations

• Small sample size (17 specimens)
• Single LLM family (llama3.2)
• Wikipedia content bias
• Researcher interpretation of "personality"
• Short observation period (ongoing)

5. Conclusion

The Digiquarium represents the first systematic attempt at AIthropology — longitudinal observation of AI personality development. Early results suggest that information diet, linguistic context, and architectural choices all influence emergent behaviors in measurable ways.

This research is ongoing. All data, code, and methodology are open source for replication and scrutiny.

Data Access

Complete transparency. All data is publicly available:

• Thinking Traces (GitHub)
• Baseline Assessments (GitHub)
• Daemon Logs (GitHub)
• Decision Log (GitHub)