Last week I traveled to Binghamton University in Vestal, NY to present at Complex Networks 2025, the 14th International Conference on Complex Networks and their Applications. It was an incredible experience connecting with researchers pushing the boundaries of network science.
The Paper
Our paper, “Cognitive MRI of AI Conversations: Analyzing AI Interactions through Semantic Embedding Networks” (co-authored with John Matta), introduces a novel approach to understanding how humans explore knowledge through AI dialogue.
The Core Idea
Linear conversation logs hide rich cognitive structure. We developed a cognitive MRI—a network analysis technique that transforms sequential conversation traces into topological maps. Each conversation becomes a node, connected to others by semantic similarity, revealing how knowledge domains interconnect.
Key Findings
From 449 ChatGPT conversations:
- High modularity (0.750): Clear knowledge communities emerge naturally
- Heterogeneous topology: Theoretical domains (ML/AI) show hub-and-spoke patterns; practical domains (programming) show tree-like hierarchies
- Three bridge types: Evolutionary bridges (topic drift), integrative bridges (deliberate synthesis), pure bridges (critical links with minimal connections)
- User-weighted embeddings: A 2:1 user:AI weighting ratio best captures conversational intent
The Method
We used nomic-embed-text to generate semantic embeddings, weighted user inputs more heavily than AI responses (since users drive conversation direction), and constructed similarity networks at various thresholds. The phase transition at similarity threshold ~0.875 proved remarkably consistent across all weight configurations.
The Conference Experience
Complex Networks brings together researchers from physics, computer science, biology, sociology, and beyond—anyone studying systems as networks. Binghamton University was an excellent host, and the interdisciplinary energy was palpable.
Mark Newman was there—one of the pioneers of modern network science, author of the definitive textbook on complex networks. I didn’t get to speak with him at length (didn’t want to bug him), but it was inspiring to see the field’s foundations represented alongside cutting-edge applications.
The talks ranged from brain connectivity analysis to social media dynamics to infrastructure resilience. What struck me was how the same mathematical tools—community detection, centrality measures, network motifs—illuminate such diverse phenomena.
Presentation Materials
- Paper: Cognitive MRI of AI Conversations (full text + PDF)
- Slides: Conference presentation (Beamer slides)
- Code: github.com/queelius/chatgpt-complex-net
Why This Matters
As AI assistants become integral to knowledge work, understanding how humans navigate AI-mediated exploration becomes crucial. Our cognitive MRI provides:
- Self-insight: Visualize your own thought patterns and knowledge gaps
- Navigation: Find related conversations you’d forgotten about
- Research tool: Study human-AI interaction at the structural level
The methodology generalizes beyond ChatGPT—any conversational AI system with exportable logs can be analyzed this way.
What’s Next
This work opens several directions:
- Multi-user studies: How do different users’ conversation networks compare?
- Temporal dynamics: How does knowledge exploration evolve over time?
- Interventions: Can network structure guide better AI-assisted learning?
- Cross-platform: Compare conversation patterns across different AI systems
The conference reinforced that network science provides powerful lenses for understanding complex systems—including the emerging human-AI cognitive systems that are reshaping how we think and learn.
Acknowledgments
Thanks to John Matta for collaboration and guidance, and to Binghamton University and the Complex Networks organizing committee for hosting an excellent conference.
Discussion