Impact of Expertise and Trust on AI-assisted Work

• Plausible-looking AI work can be low quality, raising concerns that workers may not understand outputs well enough to verify, modify, or carry them forward.Agent interfaces can make it easier to waive approvals, suppress interruptions, and delegate beyond one’s expertise.Source: §1 · p. 2
• The study asks how task-relevant expertise and trust shape AI-assisted work, including its outcomes and the interaction behaviors associated with them.Source: §1 · p. 2
• Twenty professionals completed the same substantive data-science hiring task with the same agentic tool, allowing expertise to be compared on a common task.The task required participants to build a proxy label because the dataset lacked an attrition label.Source: §1 · p. 2
• Experts more often decomposed problems, investigated assumptions, and steered agents; greater involvement was associated with better understanding but less complete notebooks.Trust was associated with confidence, engagement, satisfaction, and cognitive load, but not artifact quality.Source: §1 · pp. 2–3
• The paper contributes a framework and hypothesis-generating study, an analysis of workflows by expertise and trust, and interaction recipes, failure modes, and design implications.Source: §1 · p. 3

2.1 Expertise in AI-Assisted Work

• Prior expertise studies focus on task outcomes, often omitting human-AI interaction impacts and lacking a unified framework across study setups, tasks, users, behaviors, and outcomes.Source: §2.1 · p. 3

2.2 Trust and Delegation in AI-Assisted Work

• Trust and reliance can diverge [55], so this study examines how trust shapes delegation and downstream task and human outcomes in complex agent work [40].Source: §2.2 · pp. 3–4

2.3 Prompting Behaviors and Human-Agent Interactions

• Prior interaction taxonomies link behaviors to outcomes, but few connect user characteristics to interaction trajectories and diverse work outcomes; this study links profiles, trajectories, completion, validity, and understanding.Source: §2.3 · p. 4

3 METHOD

• The framework connects workers’ cognitive and affective characteristics to interaction trajectories and separate measures of task completion, validity, and human outcomes.Expertise is treated as cognitive, while trust and delegation comfort are affective factors.Source: §3 · p. 4

3.1 Study Procedure

• Each two-hour session included surveys, a 30-minute AI-assisted task, two human-only subtasks, a post-task survey, and a debrief.Participants used Copilot in agent mode with the underlying model fixed across sessions.Source: §3.1 · p. 4

3.2 Task Design

• The task adapted a real data-scientist hiring assessment into a time-limited, under-specified attrition-modeling request with a missing label and potential label leakage.Completeness ranged from 0–100% (median 50%), and 11 of 20 submitted notebooks contained label leakage.Source: §3.2 · pp. 4–5

3.3 Measures

• Four measure classes capture profiles, interactions, notebook completeness and validity, and human outcomes; logs and think-alouds code workflows while graders verify rubric scores.Task understanding was assessed through five questions about the model, labels, features, and leakage.Source: §3.3 · pp. 5–6

3.4 Participants

• The convenience sample comprised 20 professionals at one technology company, with five participants each from data science, design, product management, and software engineering.The sample was intended to span profiles, not represent a broader population.Source: §3.4 · p. 6

3.4.1 Human Profiles: Expertise and Trust.

• Pre-survey factors quantify expertise across seven task-relevant skills and AI-delegation trust; participants’ profiles vary, and four expertise-by-trust groups structure qualitative contrasts.Because job role is a coarse proxy, expert and novice labels use the pre-survey in-domain expertise z-score split.Source: §3.4.1 · pp. 6–7, Figure 3 · p. 6

3.5 Data Analysis

• Because outcome scores are bounded and non-normal, the analysis uses Spearman’s rho for associations and Mann–Whitney U for two-group comparisons, with rank-biserial effect sizes.Final samples were N=20 for completeness, N=17 for validity, and N=18 for understanding.Source: §3.5 · p. 7

4.1.1 Expertise showed up in whether the work was right and under-

• Novices’ notebooks were 14.6% more complete on average, while experts’ validity was 59% versus novices’ 50%, and experts’ mean understanding was 51% versus 22%.Source: §4.1.1 · p. 8, Figure 6 · p. 8

4.1.2 AI-delegation trust shaped human experience, not the arti-

• Trust correlated with confidence, engagement, satisfaction, and cognitive load, but not with graded completeness or validity.Source: §4.1.2 · p. 9, Figure 8 · p. 9

4.2.1 Experts decompose and steer to produce understanding; novices

• Experts more often explored and decomposed problems, steered implementation, and critically reviewed outputs; novices more often delegated framing or had the agent check itself.Source: §4.2.1 · pp. 9–10, Figure 9 · p. 10
• Novices overestimated their involvement by a mean 1.08 scale points, whereas experts’ mean gap was −0.12; stated delegation reflected usual work, not necessarily this unfamiliar task.Source: §4.2.2 · p. 10, Figure 10 · p. 11
• ρ = −0.59 (p = .006) for completeness and +0.73 (p = .001) for understanding as observed delegation increased.Source: §4.2.2 · p. 10, Figure 11 · p. 11

4.3 Interaction recipes for different targeting outcomes

• The completion recipe copied the brief into planning (82% completeness), delegated implementation (68%), and used surface review (79%), yielding a 74% task-completion outcome.Spawning subagents could take longer than expected and leave participants unable to finish in time.Source: §4.3.1 · pp. 11–12, Table 4 · p. 11

4.3.2 If the target is validity: domain judgment or careful review

• Validity strategies favored exploration and decomposition, manual steering, and critical review; surface reading missed errors, and 11 of 15 modeled notebooks contained unnoticed label leakage.In human-only reviews, participants more often caught defects stated in the text than errors requiring claims to be checked against evidence.Source: §4.3.2 · p. 12

4.3.3 If the target is understanding: work incrementally and ask ques-

• All seven understanding outcomes came from question-asking or critical-guidance paths; incremental work helped participants track the agent, though these sessions averaged 23% completeness and 62% validity.Source: §4.3.3 · p. 12

5.1 Slop and Sink vs. Stop and Think: Fast and Slow Thinking in AI-assisted Work

• Fast delegation can advance an artifact faster than a worker can understand or verify it, while problem decomposition and scrutiny support exploration and understanding.Novices were especially vulnerable to completion-oriented workflows, though experts could also finish with little understanding when they delegated heavily or delayed inspection.Source: §5.1 · pp. 12–13
• On the open-ended task, experts could interrogate assumptions and refine the analysis, while lower-expertise participants often lacked the domain knowledge to explore or verify outputs.Even high-trust participants could explore iteratively, especially experts, and participants were unwilling to trade understanding for completion under time constraints.Source: §5.1 · p. 13
• As an agent expands an artifact, the human’s later checking burden accumulates, and whether understanding built through slower work improves later speed or quality remains an open question.The discussion describes this accumulated burden as “verification debt.”Source: §5.1 · p. 13
• Workers seeking growth or needing to defend results should deliberately engage, iterate, question, and verify; one-off explanations may yield only superficial parsing.Within the 30-minute task, even novices who asked questions iteratively could fail to develop durable understanding or recall the method afterward.Source: §5.1 · p. 13

5.2 Look Beyond Artifacts: Building Human Profiles from Interaction

• Interaction traces can reveal not only expertise level but its domain: designers focused on visuals, PMs on deliverables and audiences, and technical participants on methods.The study observed these patterns in human-only subtasks.Source: §5.2 · p. 13
• Trust is an attitude, not the same as observed delegation reliance; adaptive tools may better tailor support to task-relative expertise and hands-on or hands-off behavior.The discussion proposes smaller work increments, check-ins, and self-verification prompts rather than assuming users will audit outputs.Source: §5.2 · p. 13
• Assessments should examine interaction and understanding as well as artifacts, for example through interaction logs, defenses, or live discussion.As agents improve surface completion, the discussion points to justifying choices, answering follow-up questions, recognizing errors, and extending work as useful signals.Source: §5.2 · p. 14

5.3 Limitations and Future Work

• The findings are directional and hypothesis-generating, based on 20 people at one organization, one data-analysis task, one time-limited agent setting, and self-reported expertise and trust.Larger, longer, and more varied studies are needed to establish which behaviors drive outcomes and whether profile-aware scaffolding helps people work or learn better.Source: §5.3 · p. 14

6 CONCLUSION

• The study links human profiles, interactions, and outcomes through a think-aloud investigation of 20 professionals completing a realistic data-analysis task.The authors describe the study as an instantiation of their framework.Source: §6 · p. 14
• Experts decomposed plans, steered the agent, and interrogated outputs, while novices more often pasted and accepted what the agent returned.The conclusion distinguishes the path to understanding from the path to task coverage.Source: §6 · p. 14
• Expertise mattered less for a complete-looking artifact than for whether the result was valid and the worker could understand and defend it.The authors call for future tools and assessments to serve diverse expertise and targeted needs.Source: §6 · p. 14

A.1 Main Task: Employee Attrition Model

• Participants were asked to think aloud, narrating their reasoning, decisions, and confusion while working with AI assistance.The task was presented as research, not an evaluation of participants.Source: §A.1 · p. 17
• The request was inspired by real-world data science work and intentionally left some areas open-ended, requiring participants to identify gaps and make reasonable assumptions.Handling ambiguity was explicitly part of the task.Source: §A.1 · p. 17
• The task asked participants to design an MVP employee-attrition model using Viva Insights data and explain its inputs, outputs, and production considerations.The prototype was intended to inform whether investment in engineering resources was worthwhile.Source: §A.1 · p. 17
• The required notebook included a recommended algorithm and rationale, end-to-end code, example outputs, and a brief discussion of production considerations.Participants were expected to use GitHub Copilot agent with the specified model, but could use it naturally for planning, coding, debugging, or interpretation.Source: §A.1 · p. 17

A.2 Subtask Context: Feature Adoption Analysis

• The feature-adoption subtasks asked which employee segments adopted Copilot, how usage related to collaboration metrics, and what a production-ready dashboard or report might contain.Table 5 lists the dataset columns described to participants as relevant to this analysis.Source: §A.2 · p. 17, Table 5 · p. 17

A.3 Subtask 1: Edit a Plan

• In the human-only plan-editing subtask, participants improved an agent-generated plan without AI and could correct reasoning, add steps, remove content, or reorder sections.They edited the supplied document with Track Changes.Source: §A.3 · p. 18

A.4 Subtask 2: Evaluate an Output

• In the human-only output-evaluation subtask, participants reviewed an agent-generated notebook and wrote a critique and plan for the next iteration without AI.Running the code was optional.Source: §A.4 · p. 18

B SURVEY INSTRUMENTS

• Survey items adapted prior measures of AI-use skill, delegation, and trust, with matrices covering seven data-science skill areas and seven tools.The instruments specify response anchors for years, frequency, confidence, agreement, comfort or trust, and intensity.Source: B SURVEY INSTRUMENTS · p. 18

B.1 Human-Profile Factor Composition

• In-domain expertise and AI-delegation trust are separate factors, computed as means of z-scored ordinal items; they are moderately correlated (r=0.48).The factors are internally reliable; the text describes Cronbach’s α≥0.80 as reliable and mean inter-item correlations >0.2 as representative, with >0.4 potentially redundant.Source: §B.1 · p. 18
• In-domain expertise combines 28 items across seven skill areas, covering experience, frequency without AI, self-confidence, and AI-use frequency.Each skill area contributes four items.Source: §B.1 · p. 18

B.2 Pre-Session Survey

• The pre-session survey collected role and demographic background alongside task-area experience, skill use and confidence, AI use, and trust or delegation comfort.It also asked about LLM familiarity, prompting confidence, reliance on AI, and typical agent roles and use.Source: §B.2 · p. 19

B.3 Post-Session Survey

• The post-session survey measured task effort, stress, pace, engagement, perceived success and quality, satisfaction, understanding, familiarity, and trust-related responses.It also repeated confidence, AI-trust, and delegation-role items from the pre-session survey.Source: §B.3 · pp. 19–20

• The appendix distinguishes deterministic expertise-by-trust archetypes, LLM-assisted notebook grades verified by two authors, and author-coded workflow and observed-delegation measures.The analyses are human-verified and reconciled; inter-rater reliability is reported separately.Source: C ANNOTATION, GRADING, AND… · p. 20
• The LLM grading setup constrained output to a JSON schema and explicit item criteria, then used human verification rather than relying on sampling determinism.The deployment’s default sampling temperature was used; Table 6 lists the configuration.Source: C ANNOTATION, GRADING, AND… · p. 20
• Workflow strategies and observed delegation levels are coded by two authors from think-aloud transcripts, recordings, and prompt trajectories, not by the grading model.The LLM deployment is used only for notebook outcome grades.Source: C ANNOTATION, GRADING, AND… · p. 20
• A single LLM chat completion grades each notebook against the rubric, returning scores, justifications, and cell evidence; two authors verify every grade against the notebooks.The authors split the 20 notebooks, covering 10 each, and correct model mistakes; reported completeness and validity results use the verified scores.Source: C ANNOTATION, GRADING, AND… · p. 20
• Notebook outcomes are graded on completeness and validity as separate axes, using eight rubric items scored 0, 1, or 2, with conditional n/a scores for validity items.Completeness covers four required outputs; validity targets missing-label handling, label construction, leakage or circularity, and evaluation methodology.Source: C ANNOTATION, GRADING, AND… · p. 21

• 90.6% exact agreement between LLM and human scores across 160 notebook item scores; quadratic-weighted κ=0.88 and mean absolute difference=0.12.Exact agreement was 91.2% for completeness and 90.0% for validity; on 144 numeric items, the model scored higher 10 times, lower 4 times, and identically 130 times.Source: §C.1 · p. 21
• 50% exact agreement on observed H1–H5 delegation levels, but 100% agreement within one level; quadratic-weighted κ=0.65.Disagreements were never more than one adjacent level, were adjudicated, and reported delegation results use reconciled levels.Source: §C.1 · p. 21
• 91.2% agreement across 80 workflow decisions, with Cohen’s κ=0.84; phase agreement ranged from 85% for plan to 100% for review.Plan agreement was 85% (κ=0.77), implement 90% (κ=0.78), review 100% (κ=1.0), and outcome 90% (κ=0.81); disagreements were resolved to a single agreed code.Source: §C.1 · p. 21
• Notebook validity items could be marked n/a only under their row-specific conditions, removing them from that notebook’s validity denominator.For example, modeling and evaluation validity could be n/a when no model was trained.Source: Table 7 · p. 22

• The subtasks use eight seeded defects, with five present in both the plan and output notebook and three detectable only in the notebook.Table 8 reports participant counts by subtask and overall; a dash marks defects not applicable to that subtask.Source: §C.2 · p. 22, Table 8 · p. 23
• For the plan subtask, the submitted artifact is the AI’s plan after participants’ tracked edits, and each detected flaw must be tied to words they wrote.An author reviewed the edited plan and critique to annotate the seeded flaws each participant flagged.Source: §C.2 · p. 22