Socioduality: A Relational Process Framework for Human-AI Interaction

Socioduality is defined as "a sequential, reciprocal, and history-carrying relational process between two distinguishable parties in which a response from one party becomes part of the observable conditions under which the other party's subsequent contribution, judgement, decision, or action is formed." The construct is specified for human-AI dyads and operationalised through nested units: moves, confirmed sociodual episodes, linked pathways, and the broader interaction container. A minimum episode A1 → B1 → A2 requires evidence of both response contingency and return contingency; candidate episodes are classified as confirmed, non-sociodual, or indeterminate before secondary coding of response orientation and substantive contribution reformation. The framework distinguishes interaction history from current contextual inputs and separates pathway description from endpoint states and downstream outcomes. Three propositions address history-conditioned formation, pathway divergence after early episode orientations, and robustness differences among endpoint-equivalent pathways. A frozen operational protocol was calibrated on three previously unseen natural human-AI records through two separately executed model-based evaluator series. Move and candidate reconstruction converged exactly in two cases and differed by one local multimodal unitisation decision in the third; remaining disagreements were concentrated at return-contingency boundaries rather than dispersed across the architecture. Socioduality therefore provides a bounded and empirically tractable process construct for analysing how human and AI contributions are formed through interaction while preserving pathway information that endpoint-centred analysis cannot recover.

The paper argues that Most human–AI evaluation still privileges capabilities, combined performance, or final outputs. Those questions remain essential, but they compress the interaction that produced the endpoint. The central thesis is that the relational process through which successive human and AI moves are formed should be analysed as a primary empirical object, not recovered indirectly from final outputs. The construct is distinguished from prior work: "Socioduality is distinguished at the level of analytical object and evidence. A1 → B1 → A2 is not claimed as a new sequence form; it is the minimum candidate structure within which two contingencies are tested: B1 must be formed in relation to A1, and B1 must then enter the observable formation of A2."

The paper establishes four nested levels of analysis: Move: one observable contribution, response, judgement, decision, action, or explicitly recorded termination produced by one party; Sociodual episode: the minimum confirmed A1 → B1 → A2 configuration in which B1 is responsive to A1 and there is sufficient evidence that B1 entered the formation of A2; Sociodual pathway: the maximal uninterrupted chain of two or more overlapping confirmed episodes, for example A1 → B1 → A2 followed by B1 → A2 → B2; and Interaction/session container: the broader recorded conversation, task, meeting, or workflow within which zero, one, or multiple sociodual episodes or pathways may occur.

Three necessary conditions are specified for confirming a sociodual episode: Two distinguishable parties: the researcher can identify the two poles producing the focal moves; Response contingency: B1 is produced in relation to A1 rather than independently, merely in parallel, or according to a fixed response that A1 cannot alter; and "Return contingency: the available evidence supports the claim that B1 entered the formation of A2—that A2 presupposed or advanced a state established or altered by B1. Continuation of a broader task topic, temporal proximity, or a fresh initiation prompted by new contextual input does not by itself satisfy this requirement."

The paper distinguishes construct existence from response orientation: "Temporal succession is not equivalent to episode confirmation. The primary identification question is whether the two relational links required for the episode are supported: B1 must be responsive to A1, and B1 must enter the observable formation of A2. Only after this evidentiary decision should A2 be coded for response orientation. Labels such as acceptance, rejection, reaffirmation, revision, verification-seeking, redirection, or termination are response orientations within an already confirmed episode. They do not, by themselves, prove that relational contingency existed."

The framework also specifies exclusion and indeterminate cases: "Clear non-cases include one-way transmission with no subsequent observable move by the receiving party; independently generated human and AI outputs later combined by a researcher or third party; B1 outputs generated independently of A1; and A2 moves known by design to be precommitted or generated independently of B1. Indeterminate cases are those where an unchanged A2 = A1 with no explicit reference, revision trace, verification behaviour, or other evidence may not reveal whether B1 was considered, ignored, or merely seen. Such cases should be coded as indeterminate and excluded from confirmed pathway comparisons unless a stronger evidentiary regime resolves the ambiguity."

The paper positions socioduality relative to prior theory, noting that "Process theory, sequential interaction analysis, and advice-taking methods can unquestionably be combined to analyse much of the same data. The construct claim is that their component availability is not equivalent to an established, common identification architecture for reciprocal response-conditioned episodes and pathways. The redundancy test is demanding: an existing construct or established theoretical integration would make socioduality unnecessary if it captured substantially the same analytical object, boundary conditions, identification rules, and empirical implications with equal or greater parsimony."

Three theoretical propositions are developed. Proposition 1 states: "Within sociodual pathways, holding the focal response and current conditions constant or closely matched, differences in accessible prior interaction history will be associated with systematic differences in the distribution of subsequent observable response orientations or moves. Proposition 2 states: Under comparable starting human–AI–task configurations, and among early confirmed episodes that do not themselves observably terminate the interaction, differences in response orientation will be associated with systematic differences in subsequent continuation structure, including whether a linked sociodual pathway forms and, if so, its verification, repair, redirection, reaffirmation, or other configuration. Proposition 3 states: Among sociodual pathways with equivalent endpoint performance, differences in pathway composition—particularly the presence and configuration of confirmed verification, correction, repair, unexamined acceptance, and compensating-error episodes—will predict differences in downstream robustness under repetition, perturbation, or error introduction."

The pathway-irreducibility principle is central: "An endpoint state identifies where a bounded process ended but cannot reconstruct the sequence of response orientations, corrections, rejections, verifications, repairs, continuations, or other relational events through which that state was produced. Confirmed sociodual episodes and pathways are therefore empirical objects in their own right; downstream consequences are a subsequent question, not the source of their existence."

The empirical identification framework follows a staged process: establishing scope and data container, identifying candidate episodes, testing response contingency, testing return contingency, assigning evidentiary status, coding response orientation, coding post-confirmation properties, linking episodes into pathways, separating current contextual inputs, and separating endpoint state from downstream outcomes.

The operational calibration applied a frozen protocol to three natural human-AI records: "an extended music-creation and release workflow, a multimodal cover-image generation interaction containing refusals and technical failures, and a 97-page website-construction interaction containing uploads, code artifacts, interface pastes, external platform messages, repeated repairs, and long active task states. Results showed that V2 and V3 produced exact agreement on move and candidate counts across evaluators, while V1 exposed a single local multimodal unitisation edge case rather than competing reconstructions of the interaction. Residual disagreement was concentrated: The recurrent judgment-sensitive area was return-contingency assessment at stage transitions."

The paper acknowledges limitations: "First, the calibration establishes operational tractability rather than construct validity, prevalence, predictive validity, or formal human inter-rater reliability; those require trained human coders and independent empirical tests. Second, residual coding discretion is concentrated at some return-contingency boundaries and a small number of multimodal or interleaved source-format cases, while the closed orientation vocabulary used in the calibration is occasionally coarse at the post-confirmation layer. Third, the present theorisation is bounded to observable human–AI dyads; sparse records may remain indeterminate, observational evidence cannot always establish causal dependence, and broader dyadic, triadic, or network portability must be tested separately."

The conclusion states: "Socioduality identifies a process that endpoint-centred human–AI analysis systematically leaves unresolved: how one party's response becomes part of the conditions under which the other party's next observable move is formed. The construct makes that process empirically tractable through a bounded architecture of moves, candidate episodes, two relational contingencies, confirmed episodes, and linked sociodual pathways. It thereby distinguishes genuine reciprocal formation from mere temporal succession, parallel contribution, fixed alternation, and evidential uncertainty."

Improvements to AI Systems Based on Socioduality:

1. Add a Relational Contingency Audit Module to Conversational AI.

• Improvement: Implement a post-hoc or real-time evaluator that checks whether each AI response (B1) was genuinely formed in response to the user's prior move (A1) and whether that AI response demonstrably influenced the user's next move (A2). This goes beyond simple next-token prediction or task success metrics.

• Capability: The AI can flag non-sociodual exchanges (e.g., where it ignored user input or where the user's next move was pre-planned) and self-correct by explicitly referencing or integrating the user's last contribution in its next response.

2. Build Pathway-Aware Memory and Context Management.

• Improvement: Instead of compressing interaction history into a summary or embedding, maintain a structured record of confirmed sociodual episodes (A1→B1→A2) with response orientations (accept, reject, revise, verify, etc.). Use this pathway as a conditioning signal for future responses.

• Capability: The AI can distinguish between current contextual input and interaction history as separate conditioning streams, enabling it to detect when a user's new move is a continuation, repair, or redirection of a prior episode—and respond accordingly (e.g., not repeating a rejected suggestion, or resuming a repair trajectory).

3. Implement Return-Contingency Verification for Multi-Turn Tasks.

• Improvement: Before generating a response in a multi-turn task (e.g., code generation, document editing), the AI explicitly tests whether its previous output (B1) was actually used by the user (A2). If evidence is missing or ambiguous, the AI can ask a clarifying question or request confirmation rather than assuming influence.

• Capability: Reduces rubber-stamping or superficial continuation where the user simply says next without engaging with the AI's prior output. The AI can prompt for explicit verification or revision, making the collaboration more genuinely reciprocal.

4. Add a Pathway Divergence Predictor for Early Interaction Orientation.

• Improvement: Train a model to predict, from the first few confirmed episodes, whether the interaction will follow a verification-heavy, repair-heavy, or acceptance-heavy pathway. Use this prediction to adjust the AI's response style (e.g., more cautious and explanatory if early episodes show rejection, or more concise if early episodes show acceptance).

• Capability: The AI can proactively adapt its communication strategy to reduce downstream repair loops or unexamined acceptance, improving overall task robustness.

5. Create a Robustness Under Perturbation Evaluator Based on Pathway Composition.

• Improvement: After a task, analyze the confirmed sociodual pathway for the presence and configuration of verification, correction, repair, and compensating-error episodes. Use this as a feature to predict how the AI-human system will perform under repeated trials, noise, or new error introduction.

• Capability: The AI can flag interactions that achieved good endpoint performance but are fragile (e.g., many unexamined acceptances) and recommend additional verification steps before deployment or in future runs.

6. Develop an Indeterminate Case Handler for Ambiguous User Moves.

• Improvement: When the AI cannot confirm whether its previous output influenced the user's next move (e.g., user repeats a request without reference), it should explicitly mark the episode as indeterminate and either (a) ask for confirmation, or (b) adjust its next response to make its influence more testable (e.g., by asking a targeted question that requires the user to reference the AI's prior output).

• Capability: Reduces silent misalignment and prevents the AI from building on unverified assumptions, leading to more reliable long-horizon collaborations.

7. Enable Response Orientation Coding for Self-Monitoring.

• Improvement: After each confirmed episode, the AI internally labels its own response orientation (acceptance, rejection, revision, verification-seeking, redirection, termination) and the user's orientation. This becomes part of its internal state for the next response.

• Capability: The AI can detect patterns like repeated rejection or verification-seeking and adjust its behavior (e.g., switching from proposing solutions to asking diagnostic questions) to break negative loops.

8. Build a Sociodual Pathway Visualizer for Debugging and Explainability.

• Improvement: Provide a user-facing or developer-facing tool that renders the interaction as a graph of confirmed episodes and pathways, highlighting where return-contingency was weak or indeterminate.

• Capability: Users and developers can see exactly where the AI failed to be responsive or where the user ignored the AI, enabling targeted improvements to prompts, model fine-tuning, or interaction design.

9. Integrate Endpoint-Pathway Separation into Evaluation Metrics.

• Improvement: When evaluating AI performance, do not rely solely on final output quality. Also report pathway-level metrics (e.g., number of confirmed episodes, ratio of verification to acceptance, presence of compensating errors).

• Capability: Enables more nuanced benchmarking—two systems with identical final outputs can be distinguished by the robustness and reciprocity of their interaction pathways, guiding better model selection for collaborative tasks.

10. Add a History-Conditioned Formation Prior to Generative Models.

• Improvement: Modify the generation process so that the probability of the next AI move is explicitly conditioned on the sequence of confirmed episode orientations (e.g., after two rejections, the AI is more likely to propose a different approach) rather than only on the raw token history.

• Capability: Produces more contextually intelligent responses that respect the relational history, reducing repetitive or tone-deaf replies in long conversations.