Following Delegation Traces to Study Human Oversight in Interaction with Multimodal AI

1. Introduction

• Delegation concerns which task parts users hand to AI, how they inspect or verify outputs, and when they intervene or reclaim control.Trustworthiness describes system and deployment properties, trust is users’ situated perception, and delegation is the transfer of control and decision authority.Source: §1 · pp. 1–2
• Existing work more often examines trust and responses to recommendations than when users transfer control or how much oversight they retain.Multimodal evaluation has also focused predominantly on model performance and user preferences rather than control negotiation across repeated interactions.Source: §1 · p. 2
• The paper proposes enacted delegation traces as a framework for studying how responsibility is redistributed during collaborative tasks.The proposed traces cover control-transfer scope, oversight actions, and recovery of control; they have not yet been implemented or validated.Source: §1 · p. 2
• In a repeated Video Question Answering study, higher-than-usual trust predicted stronger intended delegation, while retained control remained common even at high trust.Exploratory findings also associated graph-supported presentation with stronger delegation; the paper argues delegation captures oversight allocation that trust ratings alone only partly reveal.Source: §1 · p. 2

2. Background and Related Work

• Trust is a subjective judgment, whereas reliance and delegation become visible in how users allocate control.Prior work includes willingness to delegate repeated perceptual tasks and behavioral traces of whether teams let an AI partner answer for them.Source: §2 · pp. 2–3
• Explanations and inspectable intermediate representations can give users a basis for checking and revising AI contributions.Prior work also frames delegation as function allocation across human and automated agents.Source: §2 · p. 3
• Multimodal research has emphasized model-centric trustworthiness and preference evaluation more than repeated trust judgments or users’ control decisions.Benchmarks broaden evaluation beyond accuracy, while public agent arenas collect open-ended interactions and preference judgments.Source: §2 · p. 3
• The study uses Video Question Answering to examine whether users judge objects, actions, and temporal relations, with graphs exposing an inspectable intermediate representation.Graphs are treated as an interface presentation model, not as a guarantee of correctness; users can inspect and edit them.Source: §2 · p. 3

3.1. Participants and task design

• The final sample comprised 56 participants and 392 analysed task observations; for one participant who completed three sessions, only the first completed session was retained for inference.The study deliberately began with a relatively young, technically experienced population familiar with AI.Source: §3.1 · pp. 3–4
• Participants answered seven temporal-sequence questions based on the same STAR video items, and responses were presented as text alone or with an editable graph.Some groups saw ground-truth answers, while a later cohort encountered deliberately erroneous responses; condition effects were not analyzed here, and between-condition estimates were exploratory.Source: §3.1 · p. 4
• Participants rated current trust at the beginning and end of each trial, then chose whether to reject AI, use it with verification, or fully delegate on an analogous task.Because the policy concerned a future task rather than an action in the current workflow, the choice measures intended delegation.Source: §3.1 · pp. 4–5

3.2. Within-person analysis

• The pooled cumulative-logit mixed model predicted ordered delegation using within-person and mean trust, task order, cohort, graph presentation, and error exposure.A participant-specific random intercept accounted for repeated observations; item identity was added in a re-estimation because task order and item identity were confounded in the earlier cohort.Source: §3.2 · pp. 5, 7

4.1. Delegation with retained control remained common at high trust

• 79.3% of high-trust trials retained human control through rejection or verification, while only three low-trust trials involved full delegation.Verification was selected in 230 of 392 trials (58.7%), more often than rejection or full delegation.Source: §4.1 · p. 7, Figure 4 · p. 8

4.2. Task-specific trust predicted stronger intended delegation

• A one-point within-person increase in post-response trust was associated with 14.97 times higher odds of selecting a stronger delegation category.95% CI [7.14, 31.37].Source: §4.2 · pp. 7–8

4.3. Inspectable graph presentation and delegation

• Graph presentation did not reliably predict trust–delegation mismatch (OR = 1.33, 95% CI [0.55, 3.23]; p = .552).Although 16 of 28 graph-condition participants modified or deleted a node or edge, the graph-supported groups were small and unevenly distributed across cohorts, making the delegation estimate highly uncertain.Source: §4.3 · p. 8

4.4. Intended delegation showed a task-order association

• Each successive task was associated with higher odds of stronger intended delegation (OR = 1.22, 95% CI [1.02, 1.45]), adjusted for trust and study conditions.Trust showed no corresponding session-wide trend; the order association was exploratory and does not establish that repeated interaction caused greater delegation.Source: §4.4 · pp. 8–9

• The study’s intended-delegation item preserved verification as an explicit policy, but choices did not alter the workflow or carry performance or effort consequences.Its verification option also compressed distinct behaviours, from checking a salient segment to independently reconstructing or correcting an answer.Source: §5 · p. 9
• Enacted delegation traces capture how task responsibility shifts between a human and an AI during an ongoing activity, including what the AI does, what users inspect, and when they reclaim control.The unit of analysis is the sequence of delegations and interventions, not only a final accept-or-reject decision.Source: §5 · p. 9
• The proposed trace has three interaction-log dimensions: scope of control transfer, oversight after delegation, and recovery of control and its consequences.Scope includes how central the assigned task component is to the outcome; oversight includes evidence inspection, revisions, and independent verification.Source: §5 · p. 9
• The three dimensions should be analysed jointly and interpreted with contextual conditions, since broad delegation followed by timely recovery may indicate more effective oversight than narrow delegation without substantive inspection.Time pressure, perceived AI competence, and perceived stakes can change what the same transfer of control means.Source: §5 · p. 9
• Reconstructing traces requires granular workflow logs of actions such as assigning work to AI, inspecting or editing its output, rejecting content, and returning to manual control.Moment-by-moment tracking preserves trajectories that a final label such as “verified” or “delegated” would collapse.Source: §5 · pp. 9–10
• The proposed expression separates delegation trajectories from task-, system-, and participant-level covariates, but is a conceptual template for future operationalisation rather than a model estimated in this study.Self-reported trust remains complementary; sparse, intentional probes can reveal overlap or divergence without interrupting observation of enacted delegation.Source: §5 · p. 10

• The study should integrate delegation into the task itself, rather than ask about it only after the task ends.Unlike cooperative question answering that mainly observes letting AI answer for a team, the design tracks transfer over specific workflow components, inspection, revision, and recovery.Source: §6 · pp. 10–11
• A proposed learning scenario has participants answer video-based questions under varying time constraints while asking AI to locate segments, draft answers, or explain material.Participants remain responsible for the final submitted response; the initial target population could be AI-literate university students, with broader video and workplace-training contexts also possible.Source: §6 · pp. 10–11
• The interface should support manual work, assigning components to AI, inspecting or revising its contribution, and taking control back, while logging these events and using sparse trust probes.A participatory phase with workshops or focus groups can identify user-relevant delegation situations and factors beyond the initially anticipated context variables.Source: §6 · p. 11
• The study could vary AI-output presentation from plain text to an interactive workspace where users view, edit, or reject intermediate outputs.This shifts a graph-like representation from an explanation shown after an answer to an object through which users supervise collaboration.Source: §6 · p. 11
• AI-behaviour conditions could inject incorrect answers, vary proactive suggestions, or compare static with adapting systems that co-evolve with users.These are proposed dimensions for studying assistant behaviour, not results of the present study.Source: §6 · p. 11
• The design should also vary time pressure, perceived AI competence, and task stakes, while keeping the interaction low-risk and non-harmful.Examples include changing question load, showing prior AI performance, and using rewards or penalties tied to correct answers and supporting evidence.Source: §6 · p. 11

• Verification remained common even when self-declared trust was high, while inspectable and editable graph representations were exploratorily associated with stronger intended delegation but not reliably with trust-delegation mismatch.The paper calls for future evaluation of whether graph inspection helps users identify incorrect relations, revise evidence, and retain control when errors are detected.Source: §7 · p. 12
• Enacted delegation frames responsibility as a trace across a task, making observable how much work is handed over, how AI outputs are inspected, and whether control can be recovered after failure.The paper presents this as a methodological bridge between subjective trust and human oversight.Source: §7 · p. 12
• The proposed framework is intended for future studies that control time pressure, perceived AI competence, and task stakes, and compare populations with different AI familiarity and domain expertise.The empirical study began with AI-literate participants.Source: §7 · p. 12
• Enacted delegation offers a way to make the quality of oversight observable by treating distributed decisions to retain or transfer control as a measurable object of interaction research.The paper positions this aim for evaluation of multimodal, explainable, and agentic AI systems.Source: §7 · p. 12