With the rapid development of extended reality, virtual agents, social robots, and multimodal artificial intelligence, multimodal human-computer interaction (HCI) has become increasingly prevalent. A central unresolved question is why similar multimodal configurations produce divergent—sometimes opposite—outcomes across conditions: some improve objective task performance, others enhance subjective experience while leaving performance unchanged, and still others simultaneously increase workload and behavioral efficiency. Existing reviews have largely organized findings by technology category or engineering implementation without offering a unified psychological account of when and why synergy versus interference emerges.

The present paper argues that this inconsistency becomes theoretically tractable when multimodal processing is represented by two jointly estimated functional components rather than by a single global effect. The proposed framework does not claim that its component mechanisms are new or that the two paths are independent neural modules. Instead, it embeds established mechanisms in one causal architecture. Integration benefit refers to uncertainty reduction and search-space restriction produced by common-cause binding, reliability-weighted evidence combination, and valid attentional guidance. Processing cost refers to resource competition and the additional control demands generated by conflict detection and reprocessing. Binding and weighting are sequential operations, conflict monitoring and reprocessing form a trigger-execution sequence, and attention can facilitate either path depending on cue validity.

Critically, the two functional components are separable only in an experimental sense: benefitting enhancing antecedents should change integration mediators while task demand and conflict are held constant, whereas cost-enhancing antecedents should change effort or control mediators while evidence value is held constant. Both components may nevertheless co-occur and covary within one episode. This joint-estimation requirement is the framework's irreducible contribution over applying Bayesian integration and workload theories separately: it predicts when performance, workload, and subjective certainty will dissociate and when the same configuration will reverse direction as reliability changes.

Cross-modal configurations are analyzed at three necessary levels. The information level asks what task-relevant evidence is available; the content level asks whether channel-specific representations support a compatible interpretation; and the temporal level asks whether the signals can be bound as one event. The sequence therefore follows evidence availability, representational compatibility, and event bindability. The levels interact but are not interchangeable: a repeated warning can preserve the same task proposition while conflicting in source, location, or social meaning, and compatible evidence can still fail to integrate when it falls outside the temporal binding window.

The relative strength of the two pathways is moderated by three classes of boundary conditions: user characteristics (expertise, working memory capacity, prior experience), task-modality fit (degree of resource overlap between channels and primary task demands), and system reliability (whether the potential value of additional channels can be consistently realized or is diverted into monitoring and verification costs).

The paper defines an empirical identification strategy relative to a preregistered reference condition. Rather than assuming a universal net effect or tipping point, it requires prespecified primary outcomes, acceptable cost constraints, equivalence bounds, and mediator-based model comparisons.

Four conditional predictions follow. Complementary, congruent, synchronous evidence under intermediate capacity should improve performance while increasing effort; low-value redundancy should leave performance within a prespecified equivalence bound while increasing cost; valid orienting cues under a performance ceiling should improve certainty without changing objective performance; and declining reliability should weaken benefit and increase monitoring and verification demands, potentially reversing the direction of the observed outcome. The core framework is disconfirmed if validated benefit and cost manipulations cannot differentially change their proposed mediators, or if a single-factor model explains the joint indicators as well as or better than the two-component model.

The paper concludes by identifying three priorities: dynamic measurement of the two components, preregistered calibration of primary outcomes and equivalence bounds, and adaptive interfaces that switch modality configurations when monitoring demands exceed a prespecified cost constraint. In multimodal AI, trust is treated as a downstream decision outcome rather than as processing cost. The four processing predictions therefore remain applicable to proximal performance and effort, while anthropomorphic cues and prior beliefs must be modeled separately when predicting reliance. The framework is conceptual and scoped primarily to system-to-user information integration; differential mediator effects and model comparisons define the next empirical steps.