Research Brief · Emergent Novelty
Can a human-AI pairing produce something neither one could have produced alone?
Micheal Charles Preble · Preprint, not peer reviewed
The Question
Can sustained, deliberate human-AI collaboration produce a measurable rate of output attributable to neither the human's prior knowledge nor the AI's training data alone, and does that rate change as the collaboration develops over time?
The Problem
The academic literature on human-AI creative collaboration relies almost exclusively on naive participants working with a single AI system in a single short session. The paper argues this paradigm has structural limitations: it cannot detect the effects of developed “AI fluency,” cannot capture co-evolution over long timescales, and does not measure the production of complex intellectual output such as named frameworks or books.
The Contribution
Introduces Emergent Novelty (EN): output of a human-AI dyadic system not attributable to either party alone, but to the specific interaction dynamics of that pairing. Proposes a three-question Novelty Attribution Test (could the human have produced this without AI? could the AI have produced this without this specific human? would a different pairing have produced the same output?) and the Emergent Novelty Rate (ENR) as a calculable metric. Proposes a five-phase developmental model for human-AI co-evolution: Discovery, Emergence, the Hinge, Formalization, and Centaur.
How the Argument Works
The paper anchors its framework to one piece of real, independently published evidence: Shin, Kim, van Opheusden & Griffiths (2023, PNAS) analyzed over 5.8 million professional Go move decisions across 71 years and found that human decision quality and novelty rose after superhuman AI appeared, in ways not reducible to imitation. The paper's own contribution beyond that anchor is a single retroactive case study: the Novelty Attribution Test was applied to 18 deliverables produced by one operator over a 14-day window, yielding a strict ENR of 83.3% (88.3% counting partial credit for augmentation).
Why It Matters
If validated independently, the framework would suggest a human-AI dyad can function as a distinct generative system rather than only an augmented human or a supervised AI. The paper explicitly frames this as a hypothesis-generating pilot, not an established finding, and invites replication.
Evidence Status
Evidence status: Preliminary single-case evidence. The Novelty Attribution Test in this case was applied retroactively, by the same AI system that participated in producing the outputs being classified — a disclosed conflict of interest. The author is also the sole case-study operator and holds IP interests in the frameworks described; no external funding is disclosed.
What This Does Not Claim
- Does not claim the 83–88% Emergent Novelty Rate is independently validated — the paper calls it “a first-pass estimate” that “may be lower when evaluated by external raters.”
- Does not claim the five-phase developmental model is universal; it is derived from one operator's trajectory and its generalizability is explicitly unknown.
- Does not claim the “neurodivergent advantage” hypothesis in Section 7 has been tested — the paper states plainly “no study has tested this prediction; it remains a hypothesis,” and part of that section's support rests on a co-researcher's LinkedIn profile rather than a peer-reviewed source.
- Does not claim independence from evaluator bias: the classification of the operator's own outputs was performed by an AI system that helped produce those outputs.
Open Questions
Whether the five-phase arc reproduces across different operators, whether independent human evaluators would classify the same 18 deliverables the same way, and whether neurodivergent-AI pairings show a higher ENR than neurotypical pairings are all explicitly open, untested questions in the paper's own words.
Read / Cite the Research
Read the canonical paper record → · Full paper PDF · View on SSRN
Preble, Micheal Charles. “Emergent Novelty in Human-AI Dyadic Systems: A Theoretical Framework and Preliminary Evidence” Available at SSRN 7234219, 2026.