Socioduality: a new framework that treats the sequence of human–AI exchanges — not just final outputs — as the primary object of analysis, offering explicit episode/pathway identification rules and an operational protocol that the author calibrates on three real interactions, though broader empirical validation remains outstanding.

Main Finding

The paper introduces "socioduality" as a bounded, empirically tractable process construct for human–AI interaction: a sequential, reciprocal, history-carrying relational process in which one party’s response becomes part of the observable conditions under which the other party’s next contribution is formed (minimum evidentiary unit A1 → B1 → A2). The construct is operationalised with nested units (moves, candidate/confirmed sociodual episodes, linked sociodual pathways, interaction container), an identification logic (evidence of response contingency and return contingency), and a frozen coding protocol that was calibrated on natural human–AI records. The central claim is that interaction-history/pathway information cannot be recovered from endpoints alone and should be treated as a primary empirical object.

Key Points

• Definition: Socioduality = sequential, reciprocal, history-carrying relational process between two distinguishable parties (human–AI dyads focused here).
• Minimum evidentiary episode: A1 → B1 → A2, requiring observable evidence that (1) B1 was formed in relation to A1 (response contingency) and (2) B1 entered the observable formation of A2 (return contingency).
• Evidence states for candidate episodes: confirmed, non-sociodual, or indeterminate. After confirmation, episodes receive secondary codes for response orientation (e.g., acceptance, rejection, verification) and substantive contribution re-formation (degree/direction of change).
• Nested empirical units:
  • Move: the atomic contribution (turn, multimodal unit).
  • Candidate episode: potential A1 → B1 → A2 structure.
  • Confirmed sociodual episode: passes both contingency tests.
  • Sociodual pathway: maximal uninterrupted chain of overlapping confirmed episodes that preserves relational history.
  • Interaction container: the full recorded interaction (may terminate without observable sociodual termination).
• Pathway-irreducibility principle: endpoint state cannot reconstruct the relational sequence that produced it; different pathways may produce identical endpoints but have different process properties and downstream implications.
• Three testable propositions (summarised):
• History-conditioned formation: prior episodes condition the formation of subsequent moves (history affects what is observed next).
• Early-orientation divergence: early episode orientations (e.g., acceptance vs. rejection) lead to divergent pathways.
• Robustness heterogeneity: pathways that converge to the same endpoint can differ in robustness to perturbation, continuation, or downstream effects.
• Calibration: a frozen operational protocol was applied to three previously unseen natural human–AI records using two separate model-based evaluator series. Move and candidate reconstruction converged exactly in two records; one disagreement involved a local multimodal unitisation decision; remaining disagreements clustered at return-contingency boundaries.
• The framework distinguishes interaction history from concurrent contextual inputs, and separates pathway description from endpoint measures and downstream outcomes.

Data & Methods

• Nature of the contribution: primarily conceptual/theoretical with an operational identification architecture and an initial empirical calibration.
• Operationalisation:
  • Identification rules require explicit observable markers for response contingency (evidence B1 depends on A1) and return contingency (evidence that B1 entered A2’s formation).
  • Candidate episodes are first reconstructed (unitisation of moves and multimodal signals), then classified (confirmed/non-sociodual/indeterminate), then coded for orientation and substantive change.
• Calibration exercise:
  • Dataset: three natural human–AI interaction records (previously unseen by the coder protocol designers).
  • Evaluators: two separately executed series of model-based evaluators (i.e., independent coders/annotation runs using the frozen protocol).
  • Outcomes: high convergence in move and episode reconstruction (exact in two cases; single local unitisation difference in third). Disagreements were concentrated around return-contingency boundaries, suggesting those boundaries are empirically the most challenging.
• Relation to existing empirical methods:
  • Compatible with and complementary to advice-taking measures (e.g., Weight of Advice), IRF/IRE-style analyses, grounding/repair frameworks, reciprocal learning and mutual-adaptation models, and multi-turn LLM evaluation protocols (e.g., MINT).
  • Distinct emphasis on episode-level reciprocal contingencies and maximal pathway reconstruction, rather than only endpoint shifts or aggregate feedback effects.
• Limitations noted by the author:
  • Dependence on observable records; indeterminate status arises when evidence is insufficient to confirm contingencies.
  • Calibration sample small (three records) — further validation and inter-coder reliability testing on larger, diverse datasets needed.
  • Multimodal unitisation choices can affect episode reconstructions.

Implications for AI Economics

Practical and research implications for economists studying AI adoption, productivity, market design, regulation, and welfare:

1. Measurement and evaluation

   • Move beyond endpoint-only metrics: incorporate episode- and pathway-level features (confirmed sociodual episodes, pathway types, orientation codes) when estimating AI’s economic impact.
   • Complement existing measures (e.g., Weight of Advice, final accuracy/productivity) with process-sensitive indicators (frequency of confirmed uptake, types of uptake: acceptance/verification/rejection, pathway length, branching/divergence).
   • Recognise that identical endpoint outcomes can mask distinct process histories with different propensity for future change, robustness, or downstream externalities.

2. Modeling dynamic complementarities and strategic interactions

   • Socio-dual pathways suggest modeling human–AI interaction as dynamic games or sequential-response processes where past moves enter state variables. This affects identification of complementarities, complementarities' persistence, and dynamic complementarities in production functions.
   • Pathway heterogeneity implies heterogenous treatment effects: the same AI intervention can have different impacts conditional on the pathway (e.g., early acceptance → reinforcing productivity; early rejection → little impact or erosion of trust).

3. Market-level feedbacks, adoption, and coevolution

   • Platform and network effects: feedback loops at the dyadic level can aggregate into market-wide coevolution (e.g., user behaviors that systematically accept or reject AI outputs can shape product design, training data, and market offerings).
   • Policy/regulatory consequences: endpoint-focused audits (e.g., accuracy checks) may miss pathway-driven harms (e.g., incremental internalisation of biased advice through repeated uptake). Regulations and audits should consider multi-turn interaction histories where feasible.

4. Experimental and empirical design recommendations

   • Design lab and field experiments to capture multi-turn sequences and code for response and return contingency; randomise early episode orientations to study causal pathway divergence.
   • Use panel/longitudinal dyadic data or structured microdata capturing turns to estimate pathway effects and persistence, and to instrument for exogenous variation in B1 or A1 when identifying causal effects on A2 and downstream outcomes.
   • Combine WOA-like continuous measures of directional adjustment with episode-level categorical descriptors (accept/reject/verify/redirect) to capture both magnitude and mechanism.
   • Anticipate measurement costs: multimodal unitisation and return-contingency judgment require annotation protocols and possibly automated classifiers trained on coded examples.

5. Welfare, robustness, and policy evaluation

   • When assessing welfare effects of human–AI systems, account for pathway robustness: two deployments delivering the same endpoint performance may have different sensitivity to shocks, manipulations, or long-run drift depending on pathway structure.
   • Consider incentives and design interventions aimed at pathway shapes (e.g., interfaces that promote verification vs. automatic acceptance) as policy levers that alter long-run outcomes beyond immediate performance gains.

6. Research agendas for AI economics

   • Estimate the distribution of sociodual pathway types across tasks and populations, and quantify how pathway types correlate with productivity, bias amplification, trust, and error propagation.
   • Build structural or reduced-form models that include episode-history state variables to evaluate dynamic complementarities and long-run coevolution between users and AI systems.
   • Develop scalable annotation/automated detection methods for confirmed episodes and return-contingency signals to enable large-scale empirical work.

Summary recommendation for economists: incorporate sociodual process constructs into empirical and theoretical analyses of human–AI systems — collect turn-level data, code or infer response/return contingencies, and treat pathways as mediating state variables that can change both short-run outcomes and long-run dynamics. This will improve inference on complementarities, robustness, and welfare beyond what endpoint-only analyses can deliver.