PaperlayerView exact revision on arXiv

Source-linked AI summary

When My Skill Becomes Agent Skill: How Knowledge Workers Share Their Expertise with AI Systems

Tianqi Song, Zicheng Zhu, Hancheng Cao, Yi-Chieh Lee

arXiv:2610.09608v1cs.HC

TL;DR

The paper asks how AI reuse, which can apply workers’ expertise without their continued involvement, shapes willingness to share. A mixed-methods experiment compares authorization for AI and human reuse, finding lower authorization for AI and a clearer difference in whether workers share than in which categories they release.

  • Problem

    How workers authorize AI reuse of specific materials from their own work, and how those decisions differ by recipient, remains less understood despite AI systems’ ability to apply expertise without workers’ continued involvement.

    Source: §1 · p. 2, §2.2 · p. 4

  • Method

    A mixed-methods, between-subjects experiment combined quantitative survey measures and qualitative open-ended questions to study knowledge-sharing decisions.

    Source: §3 · p. 6

  • Results

    Participants were less likely to authorize AI than human reuse; category-level evidence was inconclusive, making whether to permit reuse the clearest difference.

    Source: §6 · p. 22

  • Takeaways & Limitations

    Workers’ contribution decisions concern not just making expertise available but what its reuse may mean for contributors.

    Source: §5 · p. 19

    The study cannot determine which AI properties—such as scalability, autonomy, or generalization—contributed to the authorization gap.

    Source: §5.4 · p. 22

Abstract

from arXiv · show

Organizations have long sought to make workers' expertise reusable by others. Agentic AI changes the nature of such reuse by enabling AI systems to act on workers' knowledge with limited human involvement. This raises the question of how this shift shapes workers' willingness to share their expertise. We conducted an experiment with knowledge workers who created materials incorporating domain knowledge and decided whether to authorize human or AI reuse. We find that AI and human reuse differ primarily in whether workers choose to share their knowledge, rather than in what they choose to share. Participants' reasoning shifts from prosocial considerations when sharing with humans toward concerns about loss of control, replacement, and downstream governance when sharing with AI. These findings suggest that AI reuse may intensify tensions between organizational knowledge reuse and contributors' interests. We discuss implications for workplace AI and knowledge management policies that preserve workers' rights and agency.

1 Introduction

The introduction asks how workers decide whether and what expertise to share for AI reuse, where knowledge may be applied without their continued involvement. It previews evidence that workers authorize AI reuse less often than human reuse, while their concerns also shift toward control and downstream consequences.

1 Introduction

  • AI reuse raised concerns about applying knowledge sensitively across contexts, protecting professional roles, and governing its use beyond workers’ involvement.Participants also anticipated loss of control and expert influence, while preferring recognition and continued control as conditions for sharing.Source: §1 · pp. 2–3
  • The paper addresses limited understanding of how workers authorize reuse of concrete work-derived materials, choose what to release, and set conditions for subsequent use.Its central question asks how workers decide whether and what to share with AI systems and understand the consequences and conditions of sharing.Source: §1 · p. 2
  • A mixed-methods study assigned 61 professionals to human or AI recipient conditions and elicited occupation-relevant work artifacts before authorization decisions.Participants reviewed an LLM-generated skill preview and chose whether to authorize reuse and which material categories to release.Source: §1 · pp. 2–3
  • Participants were less likely to authorize organizational knowledge reuse by AI than by human professionals, while category-level release-rate differences were not statistically significant.The clearest difference was whether workers permitted reuse at all.Source: §1 · p. 3
  • Among AI sharers, conversations were less often shared than task outputs or reasoning, with boundaries tied to sensitive information and contextual interpretation.AI-condition participants also anticipated greater loss of expert influence and control; continued control and credit were expected to increase willingness to share, alongside varied preferences for acknowledgment and anonymity.Source: §1 · p. 3
  • The authors call for workplace AI design and governance that preserve contributor agency and safeguard against adverse downstream uses.They also point to recognition, continued control, and accommodation of different preferences for credit and anonymity as ways to support voluntary participation.Source: §1 · p. 3

2 Related Work

The related work frames expertise reuse as a problem of codifying knowledge while retaining context, then identifies worker authorization and agency as underexamined in AI knowledge contribution. It organizes the study around sharing decisions, anticipated consequences, and governance conditions.

2.1 Organizational Knowledge Management

  • Knowledge management codifies expertise for retention and transfer, but codification alone does not ensure that others can apply knowledge across different work contexts.Knowledge may include situated judgment and knowing-how, not only documents or formal procedures.Source: §2.1 · p. 3

2.2 Knowledge Contribution to AI in Organizations

  • Research on AI knowledge contribution has paid less attention to workers’ authorization of specific work-derived materials and how those decisions vary by recipient.Workers may release task outputs while withholding the reasoning or conversations behind them.Source: §2.2 · p. 4

2.3.1 Knowledge Contribution as a Social Dilemma.

  • Knowledge sharing presents a social dilemma because workers invest effort while benefits may accrue to colleagues or the organization, and expected returns shape contribution.Prior work identifies reciprocity, reputation, relationships, helping, organizational support, and personal costs as relevant considerations.Source: §2.3.1 · pp. 4–5

2.3.2 Worker Agency in Knowledge Contribution to AI.

  • A less examined agency question is how workers perceive their rights and control over whether and how their expertise is used to create AI capabilities.Source: §2.3.2 · p. 5

2.4 Research Questions

  • The study asks whether and what workers authorize for AI reuse, what benefits and risks they anticipate, and which governance conditions could change willingness to share.The questions focus on recognition, control over knowledge use, and safeguards concerning downstream consequences.Source: §2.4 · pp. 5–6

3 Methods

The methods combine a between-subjects experiment, survey measures, and open-ended questions with occupation-specific simulations that produce work artifacts for review and sharing decisions. Participants were assigned a human or AI recipient, reviewed extracted knowledge, and chose whether and what to share.

3 Methods

  • The study used a mixed-methods, between-subjects experiment combining quantitative survey measures and qualitative open-ended questions.Source: §3 · p. 6

3.1.1 Occupational domains selection. / 3.1.2 Task selection and scenario development.

  • Four white-collar domains—marketing, sales, human resources, and management—were selected for realistic judgment-oriented tasks that could be simulated within study time constraints.Tasks were refined with domain professionals and selected for occupational relevance, expertise, and substantive human input.Source: §3.1.1 · p. 6, §3.1.2 · p. 7

3.1.3 Simulation workflow.

  • The simulations elicited task products, underlying reasoning, and workplace conversations through a scenario response, explanation, and dialogue with simulated coworkers.Source: Figure 2 · p. 7, §3.1.3 · p. 7

3.1.4 Validation check.

  • Participants rated the simulations as realistic (M=5.67, SD=1.07, median=6.00 on a 1–7 scale), and 77.0% reported completing similar tasks in the past year.Source: §3.1.4 · p. 8

3.2 Study Procedure

  • After a pre-survey, participants were randomly assigned to a human or AI recipient, completed a simulation, reviewed extracted knowledge, made sharing decisions, and completed a post-survey.Source: §3.2 · p. 8

3.3 System Design / 3.3.1 Frontend, Backend, and Database.

  • The platform guided participants from occupation-specific source materials through task artifacts to extracted, editable knowledge points that they reviewed before deciding whether to share.For each session, the system recorded outputs, conversation transcripts, extracted knowledge and revisions, sharing decisions, and interaction timestamps.Source: §3.3 · pp. 8–9, §3.3.1 · p. 9

3.3.2 LLM Integration.

  • The LLM generated simulated coworker responses and synthesized professional knowledge; extraction prompts classified discrete points as factual, procedural, or conditional.Participants could revise or remove points in an editable skill preview, and generally made few revisions.Source: §3.3.2 · p. 9

3.4 Participants

The study recruited experienced, full-time professionals for simulated workplace tasks and measured their authorization decisions, perceptions, and governance preferences for knowledge reuse. It combined closed- and open-ended measures with checks of baseline balance and data quality.

3.4 Participants

  • The final analytic sample comprised 61 participants from four target occupational groups, screened for education, employment, task experience, and prior AI awareness.Participants were assigned to Human (n=31) and AI (n=30) conditions; recruitment was constrained by occupation- and task-experience-specific eligibility criteria.Source: §3.4 · p. 9

3.5 Measures

  • Measures combined rating scales, binary authorization decisions, checkbox selections, and open-ended responses to address the three research questions.The analyses integrated quantitative and qualitative findings when interpreting each question.Source: §3.5 · p. 10

3.5.1 Skill-Sharing Authorization (RQ1).

  • Authorization used an overall yes-or-no decision followed, for those opting in, by selections among six categories of participant-produced materials and AI-extracted knowledge.Participants also considered contribution to an actual professional skill repository and explained which knowledge they would include or withhold and why.Source: §3.5.1 · pp. 10–11

3.5.2 Perceptions of Knowledge Sharing (RQ2).

  • Perceived consequences were measured with six items covering knowledge-use risks, loss of knowledge power, and collective benefits, supplemented by an open-ended question.Items included concerns such as loss of control or context and benefits such as helping others and preserving professional knowledge.Source: §3.5.2 · p. 11

3.5.3 Governance Preferences for Knowledge Sharing (RQ3).

  • Governance measures asked whether consent, control, credit, compensation, oversight, and protections against evaluation or replacement would increase willingness to contribute.Control was added to capture continuing authority over shared knowledge, including review, correction, restrictions, or withdrawal.Source: §3.5.3 · pp. 11–12

3.5.4 Additional measures and data-quality checks.

  • Additional measures characterized baseline balance and used comprehension, attention, and suspicion checks to assess data quality and participants’ awareness of the study focus.Participants generally described the study as examining workplace tasks, suggesting its specific focus on AI knowledge sharing was not apparent to most during the simulation.Source: §3.5.4 · p. 12

3.6 Analysis

The analysis combined quantitative tests for rating scales and authorization choices with reflexive thematic analysis of open-ended responses. Test selection reflected the response format and distributional properties, and quantitative and qualitative findings were integrated by research question.

3.6 Analysis

  • Quantitative and qualitative findings were integrated by research question, with two-sided tests and effect sizes and 95% confidence intervals reported where applicable.The analysis used α = .05.Source: §3.6 · p. 12

3.6.1 1–7 Rating-Scale Responses.

  • Rating-scale outcomes were compared between recipient conditions using Mann–Whitney U tests, with Brunner–Munzel tests additionally used when variances were unequal.Nearly all outcomes departed from normality; the two tests produced consistent conclusions in all cases.Source: §3.6.1 · p. 12

3.6.2 Binary Authorization Decisions.

  • Binary authorization rates were compared between recipient conditions using two-sided Fisher’s exact tests because some expected cell counts were small.Source: §3.6.2 · p. 13

3.6.3 Checkbox Selections.

  • Checkbox selections were compared between conditions and across categories using exact tests, with Benjamini–Hochberg correction for within-category comparisons.The analysis considered the full sample and, separately, category patterns among participants who authorized reuse.Source: §3.6.3 · p. 13

3.6.4 Open-Ended Responses.

  • Open-ended responses were analyzed through iterative reflexive thematic coding and collaborative review rather than inter-rater agreement.Authors compared interpretations, refined candidate themes, and reviewed them against coded extracts and the full dataset.Source: §3.6.4 · p. 13

4 Results

The results examine participants’ willingness to authorize AI or human reuse and their concerns and motivations around sharing. They also cover how governance safeguards, ongoing control, employment protections, and credit shape sharing preferences across the two conditions.

4.1.1 Overview of Skill-Sharing Decision.

  • 23 of 30 participants (76.7%) authorized AI reuse, compared with 30 of 31 (96.8%) authorizing human reuse.The difference was statistically significant; Figure 5a presents the overall binary authorization decision.Source: §4.1.1 · p. 13, Figure 5 · p. 14
  • Authorization was descriptively lower for AI across all six categories, but no category-level difference was significant; written reasoning showed the largest gap (66.7% vs. 87.1%, p = .073).The category results were directionally consistent with the overall decision but did not establish differences in any specific category.Source: §4.1.1 · p. 14, Figure 5 · p. 14
  • Among all participants, releasing no categories was more common in the AI condition, yet release-count distributions did not differ significantly for written materials or extracted knowledge.No-category rates were 23% vs. 3% for participant-authored materials and 27% vs. 13% for extracted knowledge; releasing all three remained common in AI (40% and 53%).Source: Figure 6 · p. 14, §4.1.1 · p. 15

4.1.2 Types of Authorized Materials and Sharing Boundaries.

  • Among authorizers, conversations were often withheld in both conditions, reflecting concerns about personal information, unpredictability, and the context needed for interpretation.In AI, conversation release was 65% versus 87% for written reasoning and task output, a nonsignificant difference; participants in both conditions described conversations as personal or liable to lose context.Source: §4.1.2 · p. 15

4.2.2 Shared Decontextualization Concerns but Different Interpretive Challenges.

  • Participants expected less control over interpretation when sharing with AI (M = 4.70, SD = 1.53) than with humans (M = 3.94, SD = 1.31; p = .032, g = 0.53).Decontextualization concerned participants in both conditions, but AI-related responses questioned systems’ ability to make contextual judgments, whereas human-related responses emphasized missing information and the need for more context.Source: §4.2.2 · pp. 16–17

4.2.3 Prosocial Motivations for AI Sharing Are Often Conditional.

  • Participants rated helping others and preserving professional knowledge similarly across conditions, but cited benefits to others more often as a reason to share with humans (51.6%) than with AI (26.7%).Human participants described helping other professionals as a direct reason to contribute, while recognizing AI’s potential benefits did not necessarily translate into willingness to share through AI.Source: §4.2.3 · p. 17
  • Five of eight AI participants who mentioned prosocial benefits endorsed AI-mediated sharing, often reciprocating knowledge received, but such motives were less frequent and conditional.Source: §4.2.3 · p. 17

4.3 Preferred Governance of Skill Sharing (RQ3)

  • All seven proposed governance conditions received mean ratings above no change in both conditions, indicating participants generally expected them to increase willingness to contribute.Source: §4.3 · p. 17, Figure 9 · p. 18

4.3.1 Ongoing Control Over Knowledge Shared with AI.

  • For withdrawal, expected willingness increased more in AI (M=6.03, SD=0.89) than Human (M=5.26, SD=1.55; p=.022, g=0.60, 95% CI [0.19, 0.99]).AI-condition participants wanted to remain informed about subsequent use and retain the ability to update contributions or withdraw permission.Source: §4.3.1 · pp. 17–18

4.3.2 Mixed Preferences for Credit and Anonymity When Sharing with AI.

  • Credit increased expected willingness more for AI (M=6.07) than Human (M=4.97; p<.001), but participants differed on attribution and anonymity.Some favored accurate credit, while others preferred anonymity; credit need not mean publicly identifying contributors.Source: §4.3.2 · pp. 18–19

4.3.3 Protection Against Adverse Employment Consequences.

  • Both groups expected protections against replacement and performance evaluation to increase willingness, with no significant condition difference; employment concerns appeared more in AI responses (8 vs. 1).Participants sought assurances that contributions would not be used to replace them or evaluate their performance.Source: §4.3.3 · p. 19

5 Discussions

This discussion considers how recipient type and AI-related concerns shape workers’ reuse decisions, alongside context-aware, contestable reuse and organizational governance. It also addresses how workplace constraints may limit explicit release choices and calls for longitudinal field studies of organizational conditions and reuse experiences.

5.1 Boundary Condition of Knowledge-Reuse Authorization for AI

  • Participants were less willing to authorize AI than human reuse in organizational and public settings, but recipient type did not reliably predict which material categories authorizers released.The clearest difference was whether reuse was permitted, rather than fine-grained selection after authorization.Source: §5.1 · p. 19

5.1.1 Decontextualization as a Persistent Knowledge-Reuse Problem.

  • Decontextualization was the most common concern in both conditions: participants feared situated knowledge would be misunderstood or misapplied, a problem not unique to AI reuse.The discussion frames decontextualization as a persistent risk whenever professional judgment is reused beyond its original setting.Source: §5.1.1 · p. 19

5.1.2 Extending Social Exchange Theory to Knowledge Contributions to AI.

  • AI-condition participants more often raised professional-replacement concerns, whereas human-condition participants more often cited prosocial motives; AI accounts also included prosocial or reciprocal reasoning.The findings suggest AI reuse changes the relative importance of familiar knowledge-sharing considerations rather than displacing them.Source: §5.1.2 · p. 20
  • Workers may value collective benefits while assigning greater weight to potential losses of professional value, influence, or control, intensifying tensions between organizational gains and contributors’ interests.The discussion describes this as an existing tension in knowledge sharing that AI reuse may intensify.Source: §5.1.2 · p. 20

5.2.1 Preserving Context While Enabling Knowledge Reuse.

  • Organizational authorizers released workplace conversations less often than task outputs or written reasoning; systems should specify source, transformation, recipient, destination, and authorized purpose.Reuse boundaries should distinguish adaptation within an approved domain from expansion to new purposes, populations, or consequential decisions.Source: §5.2.1 · p. 21

5.2.2 Making Knowledge Reuse Traceable and Contestable.

  • Workers wanted notification, revocable or limited permissions, and accountability; transparency should enable continuing influence through inspection, correction, or restriction, while separating audit provenance from identity disclosure and credit.Participants also noted that identity disclosure could create social exposure, making anonymity or pseudonymity preferable for some.Source: §5.2.2 · p. 21

5.3 Policy Implication: Institutional Governance for AI Knowledge Management

  • Organizations should establish binding reuse rules, involve workers or representatives when purposes materially change, assign responsibility for high-impact uses, protect against adverse employment consequences, and specify benefit sharing.Worker voice can help make reuse boundaries enforceable and revisable.Source: §5.3 · pp. 21–22

5.4 Limitation and Future Work

  • The study’s limits include a modest, occupationally narrow sample, a broad AI framing that bundled system properties and uses, and a controlled simulation unlike actual workplaces.Future research should use larger, more diverse samples and vary technical uses and governance conditions independently.Source: §5.4 · p. 22
  • Workplace expectations, mandatory participation, or automatic data collection may limit workers’ choices, making contribution more reluctant or constrained than this study captures.Participants had explicit choices about whether and what to release.Source: §5.4 · p. 22
  • Longitudinal field studies are needed to examine how organizational conditions and reuse experiences shape workers’ participation over time.Such studies should also examine contribution quality and the downstream uses contributions support.Source: §5.4 · p. 22

6 Conclusion

Participants were less likely to authorize AI reuse than human reuse, with the clearest difference being whether they shared at all. The findings point to preserving workers’ context, control, and attribution choices in organizational knowledge systems.

6 Conclusion

  • Participants were less likely to authorize AI reuse than human reuse, while evidence of category-level differences was inconclusive.The clearest difference concerned whether workers permitted reuse at all.Source: §6 · p. 22
  • Both groups raised concerns about knowledge being taken out of context, while AI participants expressed greater concerns about interpretive control and possible professional replacement.Source: §6 · p. 22
  • Workers emphasized continued control over subsequent use, while individual preferences varied between receiving credit and remaining anonymous.Source: §6 · p. 22
  • Knowledge-management systems should preserve contribution context and provenance and support review, correction, restrictions, and prospective withdrawal.Systems should also accommodate different attribution preferences.Source: §6 · p. 22
  • Treating workers as continuing participants in decisions about knowledge use can help protect their interests and support sustainable sharing with organizational AI.Source: §6 · p. 23

A.1 Participant Characteristics and Randomization Checks

The 61 retained participants were split across human and AI recipient conditions, and the reported checks found no statistically detectable condition differences across participant characteristics.

A.1 Participant Characteristics and Randomization Checks

  • The analytic sample comprised 61 complete participants: 31 in the Human condition and 30 in the AI condition.All retained participants reported full-time employment, and the reported patterns were consistent with successful allocation across recipient conditions.Source: §A.1 · p. 25, §A.1.1 · p. 25
  • Age was compared with two-sided independent-samples t tests; other numeric variables used Mann–Whitney U tests, categorical variables Fisher’s exact tests, and Hedges’ g was signed AI minus Human.Source: §A.1 · p. 25
  • No statistically detectable condition differences appeared in demographic characteristics, professional background, assigned task domain, or pre-task scale composites.Source: §A.1.1 · p. 25

A.1.2 Pre-Survey Control Variables.

Pre-task measures were used to characterize the sample and assess baseline balance; task exposure and content review were also examined, with review behavior unlikely to explain sharing decisions.

A.1.2 Pre-Survey Control Variables

  • Pre-task scale composites were used to characterize participants and assess baseline balance between conditions.The scales ranged from 1 to 7 and were summarized by condition.Source: §A.1.2 · p. 26, Table 6 · p. 26
  • Engagement patterns were broadly similar across occupational domains, suggesting comparable workload and exposure across task designs.Participants spent an average of 52.32 minutes on the full study and 34.19 minutes in the task application.Source: §A.1.3 · p. 27
  • The AI condition had a higher mean number of edited points, but this result was driven by four participants and arose among multiple uncorrected comparisons.The authors advise interpreting the result cautiously.Source: §A.1.3 · p. 27
  • Participants generally accepted extracted knowledge points without substantial revision, and content-review behavior was unlikely to explain later sharing decisions.The conditions were comparable in the proportion removing or editing any point.Source: §A.1.3 · p. 27

A.1.4 Post-Survey Task Evaluation.

Participants generally evaluated the occupational tasks as realistic and reported moderate-to-high self-efficacy across domains.

A.1.4 Post-Survey Task Evaluation

  • Participants generally rated the occupational tasks as realistic.Source: §A.1.4 · p. 28
  • Participants reported moderate-to-high self-efficacy across the task domains.Source: §A.1.4 · p. 28
  • A majority of participants in each domain had completed a similar task before, suggesting the scenarios were recognizable relative to their professional experience.Source: §A.1.4 · p. 28

A.2 Measurements

In a public-repository decision with real consequences, participants were less likely to authorize AI reuse than human reuse. AI recipients also affected how much participants shared, while reported concerns and preferred governance conditions differed by recipient.

A.2.1 Public-Repository Sharing as an Ecological Validity Check

  • Public-sharing permission was granted by 18/30 AI-condition participants (60.0%) versus 28/31 Human-condition participants (90.3%), a significant difference.The reported odds ratio was 0.16 (95% CI [0.05, 0.68], p = .008).Source: §A.2.1 · p. 28
  • Authorization rates were lower in the AI condition for all six material categories; the overall direction matched organizational sharing, with more pronounced category-level differences.Three category differences remained significant after correction; differences for conversations, factual knowledge, and procedural knowledge did not.Source: §A.2.1 · pp. 28–29, Figure 10 · p. 29
  • Written reasoning, task output, and decision rationales were each released less often to AI than human recipients.Written reasoning: 13/30 (43.3%) versus 26/31 (83.9%); task output: 12/30 (40.0%) versus 24/31 (77.4%); decision rationales: 13/30 (43.3%) versus 23/31 (74.2%).Source: §A.2.1 · p. 29
  • Participants released fewer knowledge categories to the public repository in the AI condition (M = 2.60) than in the Human condition (M = 4.35).The organizational-sharing comparison was in the same direction but smaller and not statistically significant.Source: §A.2.1 · p. 29
  • Among participants who authorized public sharing, no individual category differed significantly by recipient condition; overall, recipient type primarily affected whether and how much participants released, not which categories they selected.The conditional category analyses were interpreted descriptively.Source: §A.2.1 · p. 29

A.2.2 Full Item-Level Results for RQ2

  • Compared with human recipients, AI recipients elicited expectations of greater loss of influence from expertise and less control over how knowledge would be interpreted.No significant condition differences appeared for the other four anticipated consequences.Source: §A.2.2 · p. 29, Table 10 · p. 30

A.2.3 Full Item-Level Results for RQ3

  • Credit and control were rated as more important for increasing willingness to contribute knowledge in the AI condition than in the Human condition.Compensation and the three workplace-specific safeguards showed no significant condition differences.Source: §A.2.3 · pp. 29–30, Table 11 · p. 30
Loading 2610.09608v1…