Existing studies of human-AI creative collaboration use naive participants in single short sessions and cannot detect effects of developed fluency or long-term co-evolution.
Research Poster · Emergent Novelty
Emergent Novelty in Human-AI Dyadic Systems: A Theoretical Framework and Preliminary Evidence
Preprint, not peer reviewed · SSRN 7234219
Problem · Background · Research Question
Anchored to Shin et al. (2023, PNAS): human novelty in expert Go play rose after superhuman AI appeared, not reducible to imitation.
Can sustained human-AI collaboration produce a measurable rate of output attributable to neither party alone?
Framework: the Novelty Attribution Test
- Q1: Could the human have produced this without AI? If yes → not emergent.
- Q2: Could the AI have produced this without this specific human’s input? If yes → retrieval or augmentation, not emergence.
- Q3: Would a different human-AI pairing have produced the same output? If yes → general AI capability, not dyad-specific emergence.
- Only when all three answers are no is the output classified as Emergent Novelty.
Central diagram: the five-phase Temporal Vector
- 01Discovery
Transactional AI use, treated as an encyclopedia. Emergent Novelty Rate near zero.
- 02Emergence
Exploratory use; output quality improves but completion rates stay low.
- 03The Hinge
A phase transition: sudden increase in output completion and session complexity.
- 04Formalization
Named operations and deliberate methodology develop. ENR begins climbing.
- 05Centaur
The dyad operates as one system; the method is used to describe itself.
Propositions / Observable Implications
- Compounding emergence: emergent outputs can become inputs to further emergent outputs, at increasing levels of abstraction.
- Dyadic Amplification (ENA): the degree to which combined output exceeds the best available human-alone or AI-alone output; tracking ENA over time would test the compounding hypothesis.
- Neurodivergent advantage (untested hypothesis): if true, neurodivergent-AI dyads would show a higher ENR than neurotypical-AI dyads. No study has tested this.
Evidence Required
Retroactive classification of 18 deliverables from a 14-day window (March 26–April 9, 2026), single operator: 15/18 fully emergent, 3/18 augmentation, 0/18 retrieval. Strict ENR 83.3%; inclusive ENR 88.3%. 8 of 15 emergent deliverables (44%) showed further downstream emergent outputs (“strong emergence”).
Limitations
Single operator, single cognitive profile; generalizability unknown. Classification performed by the AI system that participated in production — a disclosed conflict of interest. Author holds IP interests in the frameworks described. The 83–88% figure is a first-pass, non-independently-validated estimate.
Falsifiers / Open Questions
Independent human evaluation of the same 18 deliverables could produce a materially different ENR. Whether the five-phase model reproduces across different operators, and whether the neurodivergent-advantage hypothesis holds, are both untested.
Citation / QR
Suggested citation
Preble, Micheal Charles. “Emergent Novelty in Human-AI Dyadic Systems: A Theoretical Framework and Preliminary Evidence” Available at SSRN 7234219, 2026.
Preble, Micheal Charles. “Emergent Novelty in Human-AI Dyadic Systems: A Theoretical Framework and Preliminary Evidence” Available at SSRN 7234219, 2026.