---
id: "lib-2026-10-07-when-my-skill-becomes-agent-skill-how-knowledge"
type: "library"
title: "When My Skill Becomes Agent Skill: How Knowledge Workers Share Their Expertise with AI Systems"
lang: "EN"
org: "NUS / Emory"
source_type: "Paper"
date: "2026-10-07"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-07-when-my-skill-becomes-agent-skill-how-knowledge/"
original_url: "https://arxiv.org/abs/2610.09608"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Randomized mixed-methods experiment from NUS and Emory HCI researchers that measures real authorization decisions, though on a small sample of 61 US professionals in a paid simulation."
summary_basis: "full-text"
original_minutes: "40"
tags: ["enablement", "future of work", "tacit knowledge", "agent skills"]
frames: ["tacit-to-explicit"]
---

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

Tier A (Strong evidence): Randomized mixed-methods experiment from NUS and Emory HCI researchers that measures real authorization decisions, though on a small sample of 61 US professionals in a paid simulation.

Summary based on the full text.

## In short

A study by Tianqi Song, Zicheng Zhu, Yi-Chieh Lee (National University of Singapore) and Hancheng Cao (Emory), posted on arXiv on 7 October 2026. It tests whether knowledge workers are as willing to let AI agents reuse their expertise, packaged as "agent skills", as they are to let human colleagues reuse it. They are not: workers were less likely to authorize AI reuse at all, while what they chose to share looked similar once they agreed. Reasons shifted from helping colleagues to worries about loss of control, replacement and how the knowledge would be used later.

## Key numbers

- 76.7% vs 96.8%: share of professionals who authorized reuse of their materials by AI systems versus by human professionals (p = .026)
- 60.0% vs 90.3%: authorization for an actual public repository, AI versus human recipients (p = .008)
- 6.07 vs 4.97: expected boost to willingness from being credited for reuse, AI versus human condition, 1-7 scale (g = 1.30)

## What it reports

- 61 full-time, college-educated US professionals in HR, people management, sales or marketing, randomly assigned to an AI or human recipient.
- Organizational repository: 23 of 30 (76.7%) authorized AI reuse versus 30 of 31 (96.8%) for humans; OR = 0.11, p = .026.
- Actual public repository: 18 of 30 (60.0%) for AI versus 28 of 31 (90.3%) for humans; OR = 0.16, p = .008.
- Among those who shared with AI, conversations were released less often (65%) than written reasoning or task output (87% each); category differences between conditions were not significant.
- Expected loss of expert influence: 4.07 (AI) versus 3.10 (human), p = .020. Expected loss of control over interpretation: 4.70 versus 3.94, p = .032.
- Replacement worries in open answers: 16.7% of AI-condition participants versus 3.2% in the human condition. Benefit to others as a reason to share: 26.7% versus 51.6%.
- Governance: the ability to review, correct, restrict or withdraw raised expected willingness more for AI (6.03 versus 5.26, p = .022); credit even more (6.07 versus 4.97, p < .001, g = 1.30).

## How they got it

Participants, recruited via CloudResearch and paid US$10 for about an hour, completed an occupation-specific task drawn from O*NET, wrote their reasoning and talked with two simulated coworkers. An LLM extracted a skill preview, which they reviewed. They then decided whether and what to share with their assigned recipient, and rated consequences and governance conditions. Simulations were rated realistic (5.67 of 7).

## Limits

- 61 participants in four roles limits precision; null category-level results do not show the absence of differences.
- "AI" was described broadly, so the study cannot separate training, retrieval and autonomous agent use.
- A paid study with free choice likely overstates willingness compared with workplaces where collection is mandatory or automatic.

## Why it matters for companies

Companies building agent skills from employee know-how should expect a consent gap: in this study roughly one in four workers who would share with colleagues refused when the recipient was AI. The levers that moved willingness were concrete and cheap: credit for reuse, and the ability to review, update and withdraw contributions after sharing. Informal conversations are the material people hold back first, which matters for systems that mine chat or meeting records. Treating skill capture as data collection without these controls risks the kind of backlash the authors cite at Meta.

Original: https://arxiv.org/abs/2610.09608