Socioduality framework: arXiv paper maps how humans and AI mutually shape each other

Traditional human-AI research tends to rank individual capabilities, measure combined performance, or score final outputs—but these approaches discard the dynamics of how one party's response becomes the conditions for the other's next move. In a new arXiv paper (2608.11322), Mehmed Zahid Çögenli proposes socioduality, a relational process framework that treats human-AI interaction as a sequential, reciprocal, history-carrying exchange. The construct is designed specifically for human-AI dyads, breaking interactions into nested units: individual moves, confirmed sociodual episodes, linked pathways, and the broader interaction container. A minimal episode follows the pattern A1-B1-A2, where A is the human and B is the AI, and it must satisfy both response contingency and return contingency—meaning each contribution demonstrably reacts to the previous one.

Candidate episodes are then classified as confirmed, non-sociodual, or indeterminate, followed by secondary coding of response orientation and substantive contribution re-formation. To make the framework empirically tractable, the author defined a frozen operational protocol and ran two separate model-based evaluator series on three previously unseen natural human-AI records. The results were striking: move and candidate reconstruction converged exactly in two of the three cases, with the third differing by only a single local multimodal unitisation decision. Remaining disagreement clustered around return-contingency boundaries, highlighting where the framework is most sensitive.

The paper also puts forward three propositions about history-conditioned formation, pathway divergence, and robustness differences among endpoint-equivalent pathways. These suggest that two interactions can lead to the same final output yet follow very different relational trajectories—and that some pathways are more robust than others. Socioduality therefore gives researchers a bounded, reproducible way to analyze not just what humans and AI produce together, but how they co-create that output through interaction, preserving pathway information that endpoint-centred analysis cannot recover. For practitioners building agents, this could be a foundation for designing and evaluating AI systems that genuinely adapt to user contributions over time.