Source-linked AI summary
From Future of Work to Future of Workers: Addressing Asymptomatic AI Harms for Dignified Human-AI Interaction
Upol Ehsan, Samir Passi, Koustuv Saha, Todd McNutt, Mark O. Riedl, Sara Alcorn
TL;DR
Asymptomatic craft erosion and gaps in worker protection remain underaddressed in AI-mediated workplaces. A year-long longitudinal study in radiation oncology found that AI’s short-term efficiency gains can accompany skill atrophy and erosion of professional identity.
Problem
AI discourse leaves asymptomatic craft erosion underaddressed, while knowledge workers face gaps in worker protection.
Source: §2.1 · p. 2, §6.1 · p. 15Method
The paper conducted a year-long longitudinal study of AI integration in radiation oncology using mixed qualitative methods.
Source: §3 · p. 3Results
AI assistance brought short-term efficiency gains alongside asymptomatic effects that developed into skill atrophy and wavering professional identity.
Source: §1 · pp. 1–2, §4 · p. 6Takeaways & Limitations
The paper concludes that skill preservation and human dignity should be treated as requirements alongside efficiency, with sociotechnical immunity to detect, contain, and recover from hidden harms.
Source: §8 · p. 16The framework was cross-domain verified and deployed at multiple sites, but findings came from a single domain.
Source: §7 · p. 16
Abstract
from arXiv · showhide
In the future of work discourse, AI is touted as the ultimate productivity amplifier. Yet, beneath the efficiency gains lie subtle erosions of human expertise and agency. This paper shifts focus from the future of work to the future of workers by navigating the AI-as-Amplifier Paradox: AI's dual role as enhancer and eroder, simultaneously strengthening performance while eroding underlying expertise. We present a year-long study on the longitudinal use of AI in a high-stakes workplace among cancer specialists. Initial operational gains hid ``intuition rust'': the gradual dulling of expert judgment. These asymptomatic effects evolved into chronic harms, such as skill atrophy and identity commoditization. Building on these findings, we offer a framework for dignified Human-AI interaction co-constructed with professional knowledge workers facing AI-induced skill erosion without traditional labor protections. The framework operationalizes sociotechnical immunity through dual-purpose mechanisms that serve institutional quality goals while building worker power to detect, contain, and recover from skill erosion, and preserve human identity. Evaluated across healthcare and software engineering, our work takes a foundational step toward dignified human-AI interaction futures by balancing productivity with the preservation of human expertise.
1 Introduction
A year-long study of AI in radiation oncology reveals how short-term efficiency can conceal gradual erosion of workers’ expertise and identity. The paper proposes dignified Human-AI interaction that protects human agency, expertise, and self-worth.
1 Introduction
- Task-centric accounts of AI success emphasize speed, accuracy, and throughput, leaving gaps in frameworks and tools for detecting asymptomatic harms to workers’ agency, expertise, and dignity.The paper shifts attention from job displacement and task automation to long-term effects on professional skills, judgment, and identity.Source: §1 · p. 1
- In a year-long radiation-oncology deployment study (42 participants; 24 interviews, 5 workshops, 52 think-aloud sessions), invisible behavioral shifts hardened into manual-skill atrophy and wavering professional identity.Clinicians feared losing the hands-on skills that made them unique and becoming bystanders in their own practice.Source: §1 · p. 1
- The AI-as-Amplifier Paradox describes how AI can improve doctors’ short-term efficiency while invisibly eroding the expertise and core practices it was meant to support.Hidden drifts can harden into chronic damage and change what it means to be a doctor.Source: §1 · pp. 1–2
- An intervention addressing overreliance and sustaining critical engagement revealed the need for proactive safeguards; Social Transparency features addressed parts of the paradox.The findings informed a multi-level framework for dignified Human-AI interaction.Source: §1 · p. 2
- The paper defines dignified Human-AI interaction as integration that preserves agency, expertise, and self-worth, and proposes sociotechnical immunity to detect, contain, and recover from hidden harms.Its multi-level framework uses early-warning signals, containment actions, and recovery routines, and was evaluated in healthcare and AI-assisted software engineering.Source: §1 · p. 2
- The longitudinal, high-stakes study exposes hidden deterioration that can precede symptomatic failures and argues that human dignity and flourishing belong among AI-mediated work’s success metrics.The authors frame this as a shift from the future of work to the future of workers.Source: §1 · p. 2
2 Background and Related Work
The background contrasts AI’s productivity promise with harms to workers that may emerge unnoticed, and reviews short-term effects alongside growing evidence of longer-term threats to dignity and professional identity. It situates the paper’s worker-centered, longitudinal contribution amid calls for both systemic analysis and practical responses to workers’ lived experiences.
2.1 Future of Work vs. Future of Workers
- AI is framed as improving accuracy, efficiency, and productivity [83], while workers face oversight challenges, de-skilling, and displacement [25] [83]; craft erosion may remain unnoticed until expertise has atrophied.The cited research also links this pattern to automation bias and overreliance, especially under time pressure or when systems appear accurate.Source: §2.1 · p. 2
- Worker-centered scholarship highlights how technology can intensify workloads and power asymmetries, while knowledge workers’ AI management remains understudied despite distinct vulnerabilities and limited collective action.The paper positions its worker-developed framework as addressing the resulting lack of pragmatic tools for navigating erosion of expertise and identity; the cited passage also reports that AI threatens college graduates 5x more than high school graduates.Source: §2.1 · pp. 2–3
2.2 Troubling Signs in Human-AI Interaction and Collaboration
- Studies find that AI assistance can cause unnoticed skill atrophy and dependence, correlate with reduced critical thinking, and hinder learning or judgment [33]; this paper adds longitudinal evidence beyond short-term snapshots.It also examines workers’ everyday adaptations amid systemic conditions and technological change.Source: §2.2 · p. 3
2.3 AI Harms: Impacts to Human Dignity and Professional Identity
- Emerging evidence links AI to hidden costs including diminished control, anxiety, de-skilling, alienation, and weakened professional identity across clinical and educational work.These costs can brew beneath visible productivity gains and short-term impacts.Source: §2.3 · p. 3
3 Study Context & Methods
This year-long mixed-methods study traced routine AI use in radiation oncology, focusing on how clinicians’ relationship with an AI-assisted treatment-planning system evolved and how skill erosion could emerge.
3.1.1 Case Study Background.
- Radiation oncology’s irreversible, multi-stakeholder cancer-treatment decisions formed the high-stakes setting for studying RadPlan, a commercial AI planner deployed at pseudonymous Hospital Alpha.RadPlan generated initial dose plans for clinicians to review and adjust; synthetic data were used in the reconstructed interface to protect proprietary systems and patient information.Source: §3.1.1 · pp. 3–4
3.1.2 The Problem and Intervention:
- Across five sites, RadPlan initially shortened planning cycles by ~15%, improved dose-volume histograms, and raised confidence, but clinicians later voiced intuition rust and unexamined acceptance.In response, the team added Social Transparency: clinician-facing 4W logs showing who did what, when, and why, with privacy safeguards. Its rationale included multi-stakeholder decisions, no simple ground truth, and shifting standards; Figure 2 depicts the intervention.Source: §3.1.2 · pp. 4–5, Figure 2 · p. 4
3.2 Participants: Overcoming Access Challenges and Building Community Trust
- The study recruited and worked with 42 professionals—15 radiation oncologists, 12 medical physicists, 7 dosimetrists, and 8 hospital administrators—representing a significant achievement given the access challenges.Participation was voluntary, with informed consent obtained at each phase.Source: §3.2 · p. 5
3.3 Data Collection & Analysis
- Researchers logged RadPlan use and conducted 52 in-situ think-aloud sessions, five participatory workshops, and 24 interviews across the 12-month deployment.Analysis used iterative grounded-theory coding, interdisciplinary discussions, and triangulation of interview, workshop, and drawing data to develop the study’s themes.Source: §3.3 · p. 5
3.4 Ethics, Privacy, and Reflexivity
- The team used pseudonyms, participant IDs, and removal or minimal editing of identifying details, while reflecting on how its interdisciplinary and diverse perspectives shaped interpretation.The team included HCI, STS, Human-centered AI, NLP, and radiation-oncology expertise; clinician co-authors helped build trust, guide recruitment, and shape the study and interventions.Source: §3.4 · p. 5
4 Findings: From Asymptomatic Effects to Chronic Harms to Identity Commoditization
The findings trace a progression from hidden shifts beneath early AI-enabled performance gains to accumulating damage to expertise, autonomy, resilience, and professional identity. They show how AI can enhance work while eroding the human capabilities it was meant to support.
4.1 Emergence of the AI-as-Amplifier Paradox: From Enhancement to Erosion
- The AI-as-Amplifier Paradox is that AI-enhanced performance can erode the expertise it supports, progressing from subtle effects to chronic harms and identity commoditization.Participant narratives depict this as an expertise erosion cascade: visible success metrics can mask worsening effects over time.Source: §4.1 · p. 6, §4 · p. 6, Figure 4 · p. 8
4.2 Asymptomatic Effects: Hidden Shifts Beneath the Surface
- About 15% faster planning and slightly better plans initially masked reduced vigilance and intuition, as AI reliance curtailed exploration without immediate errors.Routine cases invited faster acceptance, while some less experienced staff risked false confidence even in complex cases; asymptomatic shifts required behavioral proxies to detect.Source: §4.2 · pp. 7–8
4.3 Chronic Harms: Erosion of Skill, Autonomy, Resilience Over Time
- By about six months, chronic harms included slower manual planning, diminished confidence and autonomy, and a thinning human safety buffer that made success increasingly brittle.Some workers independently practiced without AI to maintain skills, but these countermeasures were not institutionalized and complacency remained the general trend.Source: §4.3 · p. 9
4.4 Identity Commoditization: Threats to Professional Identity and Dignity
- Practitioners feared becoming AI overseers or rubber stampers, hollowed out rather than replaced, with reduced expertise threatening professional identity and dignity.They linked dignity to having meaningful, respected work and warned against becoming bystanders in their own practice.Source: §4.4 · pp. 9–10
- Participants also worried that managers might treat clinicians as interchangeable, lower-skilled AI supervisors and that AI would reshape professional training and the meaning of expertise.These were longer-term anxieties; the study reports that no such workforce action occurred during the study.Source: §4.4 · p. 10
5 Framework Development & Evaluation: Towards a Dignified Human–AI Future
This section develops and assesses a worker-centered framework for detecting and countering AI-related expertise erosion, using scenario-based workshops in healthcare and software engineering. The workshops examine how its practices can make erosion legible and protect human expertise while meeting institutional quality needs, and underscore the importance of embedding them in routine workflows.
5.1 Social Transparency: A Turning Point on Navigating the Amplifier Paradox
- Social Transparency used peer rationales and 4W context to expose AI limits, calibrate reliance, interrupt autopilot, and keep human expertise visible, while motivating a more proactive framework.Small design frictions did not solve the erosion problem, but helped practitioners question, verify, and correct AI outputs and recognize the value of their own expertise.Source: §5.1 · pp. 10–11
5.2.2 Individual Worker Level (maintaining agency and expertise).
- The Worker level prompts practitioners to protect meaningful craft, maintain fading skills, and use signals such as repeated instant approvals to trigger AI-off review.Participants described weekly manual contouring as quality assurance that kept skills alive, and framed everyday resistance through legitimate institutional practices.Source: §5.2.2 · p. 11
5.2.3 Technology Level (building erosion-resistant systems).
- The Technology level calls for erosion-resistant AI that surfaces uncertainty, adds useful friction, automates draining rather than valued work, and calibrates reliance using human corrections.Participants proposed using recurring AI disagreements to expose brittle areas and adjust model confidence.Source: §5.2.3 · pp. 11–12
5.2.4 Organizational Level (governing dignified Human-AI collaboration).
- The Organizational level uses policies and oversight to contain harms, protect human authority, set humane automation boundaries, and reinvest saved time in practice and peer learning.The immunity triad links early-warning indicators, circuit breakers such as AI-off reviews, and recovery routines such as manual-first drills and override logs.Source: §5.2.4 · p. 12
5.3.1 Methods & Analysis:
- Two participatory, scenario-based workshops assessed feasibility and affordances in healthcare and software engineering, with 17 participants new to or experienced in AI-assisted work.The healthcare workshop included 7 participants new to the study, while the cross-domain workshop included 10 AI-assisted programming practitioners; 12 hours of data were analyzed.Source: §5.3.1 · p. 12
5.3.2 Framework Walkthrough (within-domain): RadPlan.
- In healthcare and software-engineering walkthroughs, participants used the framework to detect quiet AI-reliance drift, add human safeguards, and recover through logged traces and skill practice.The evaluation used participant-driven scenario design with 7 healthcare participants and 10 software-engineering participants; the walkthroughs applied framework questions to RadPlan and AI-assisted billing code.Source: §5.3.1 · p. 12, §5.3.2 · p. 13, §5.3.3 · p. 14
5.3.4 Framework Evaluation Findings.
- Participants repurposed signals such as instant approvals and zero-edit merges into AI-off reviews, rationale requirements, and traceable records that served both quality and worker-protection goals.Override logs and peer annotations could document the human expertise behind decisions while satisfying institutional monitoring requirements.Source: §5.3.4 · p. 14
- Participants worried protective routines would not last unless lightweight instrumentation fit existing workflows, including QA-dashboard drill rates, rotating craft champions, and periodic override-log reviews.Suggested reviews were one override log per clinical round or sprint.Source: §5.3.4 · p. 14
- Workshops showed the framework turning private anxieties about expertise erosion into collective evidence and routine metrics into protection without open confrontation.Participants described shared vocabulary as infrastructure for coordination and negotiation, and as a united voice to management.Source: §5.3.4 · p. 14
- The authors call for future evaluation of the framework in additional domains.Appendix A.2 provides best practices on metrics and an evaluation plan.Source: §5.3.4 · p. 14
6 Discussion & Implications
The discussion argues that AI can both strengthen and erode workers’ capabilities, and proposes dignified Human-AI interaction that balances organizational goals with worker protection. It calls for evaluating and designing AI to preserve expertise, agency, and professional purpose over time.
6.1 Addressing the Missing Middle: The Framework’s Dual Architecture for Knowledge Worker Protection
- Knowledge workers face a “missing middle” of precarity, and the framework offers tactical tools for survival now while building groundwork for future collective resistance.The group includes professionals who may exceed $100K unemployment-assistance thresholds, face NDAs, and work where unionization is difficult or outlawed.Source: §6.1 · p. 15
6.1.1 Dual-Purpose Design: Quality Standards as Worker Protec-
- Social Transparency’s 4W records served quality control while documenting AI limitations; participants reported 23 missed dose-constraint violations and used the evidence to prevent a staff reduction and support advocacy.The 4Ws record who did what, when, and why.Source: §6.1.1 · p. 15
6.1.2 Threading the Pragmatic Needle Between Survival and Trans-
- The framework turns professional excellence into collective protection by repurposing managerial goals, creating “uneasy alliances” that help workers navigate AI harms without direct confrontation.It is a pragmatic response to AI already in use, using professional norms such as quality and compliance to protect expertise while building infrastructure for future organizing.Source: §6.1.2 · p. 15
6.2 Research & Design Implications
- Human-Centered AI research should measure long-term expertise retention, professional agency, and skill resilience alongside performance, making dignity an explicit workplace AI success criterion [49].Longitudinal studies can assess whether AI sustains workers over time, not merely what it adds to task performance.Source: §6.2 · p. 15
- Designers should build erosion resistance and sociotechnical immunity into AI through human-only task lists, Social Transparency, seamful design, and safeguards that preserve human judgment and skills.The proposed safeguards include cognitive forcing functions, uncertainty communication, provenance trails, and AI-off drills; humane automation targets relevant tasks without displacing human roles.Source: §6.2 · pp. 15–16
6.3 The Economic Case for Dignified Human-AI Interaction: Preventing Catastrophic Costs
- Preserving human expertise is a business risk-control strategy: over-automation can leave teams unable to diagnose failures, recover capability, or meet oversight needs.The paper cites Knight Capital’s $460 million loss, Boeing’s MCAS deaths, and Zillow Offers’ $500 million loss as failures, and argues for sensing drift, containing automation creep, and reinvesting efficiency in skill recovery.Source: §6.3 · p. 16
7 Limitations and Future Directions
The study offers a foundational framework, but its findings come from one domain despite cross-domain verification and multi-site deployment. Future work should test and refine the framework across domains, populations, methods, and institutional arrangements, while the authors extend their empirical design work with critical reflection on AI’s implications for workers.
7 Limitations and Future Directions
- Although cross-domain verification and multi-site deployment informed the framework, the findings came from a single domain, so future work should test it across domains, populations, regulatory environments, and AI tools.Suggested contexts include finance, cybersecurity, professional training, and students acquiring AI literacy.Source: §7 · p. 16
- Future mixed-methods studies could quantify skill-degradation rates, recovery timelines, and immunity thresholds, while examining power dynamics when AI vendors and end users span institutions.Source: §7 · p. 16
- Drawing on Agre’s Critical Technical Practice, the authors aim to pair empirical design work with self-reflective critique of AI’s implications for workers.They seek to learn from and with the HCI and AI communities.Source: §7 · p. 16
8 Conclusions
The radiation-oncology study traces AI’s effects from asymptomatic erosion to chronic harms and identity commoditization, while showing that this trajectory is not inevitable. Its worker-developed framework uses professional norms and institutional quality and compliance goals to protect dignity and expertise alongside efficiency.
8 Conclusions
- In radiation oncology, the AI-as-Amplifier Paradox progressed from asymptomatic effects to chronic harms and identity commoditization, but this erosion is not inevitable.The study frames this progression as a risk that can be addressed before harms ossify.Source: §8 · p. 16
- Social Transparency and a multi-level dignified Human-AI interaction framework build sociotechnical immunity to detect, contain, and recover from AI’s hidden harms.The framework aims to address harms before they ossify.Source: §8 · p. 16
- The future of work should treat human dignity and skill preservation as first-class requirements alongside efficiency.The conclusion calls for centering workers rather than workflows as AI transforms work.Source: §8 · p. 16
- Developed with workers, the framework uses institutional quality and compliance logics through professional norms to protect the structurally vulnerable “missing middle.”This dual-purpose design addresses knowledge workers who are too privileged for conventional collective action yet vulnerable to AI’s erosive effects.Source: §8 · p. 16
A Appendix · A.1 Further details on Workshops and Interviews · A.1.1 Workshop Timeline and Activities (WS1–WS5).
Across a 12-month RadPlan deployment, five workshops with four professional roles traced changing experiences of AI and co-constructed the Dignified Human-AI Interaction Framework.
A.1.1 Workshop Timeline and Activities (WS1–WS5).
- WS1 mapped planning with and without RadPlan, capturing early efficiency optimism alongside isolated concerns about shortcuts creeping in.Participants annotated decision points and drew how AI changed their work and which judgments still felt uniquely human.Source: §A.1.1 · p. 20
- WS2 refined a Social Transparency mock-up as participants described intuition rust and unease about accepting RadPlan’s first suggestion without understanding it.Participants specified what 4W events and peer rationales to show and where transparency cues should appear without overwhelming the planning screen.Source: §A.1.1 · p. 20
- WS3 used 4W traces and earlier drawings to distinguish asymptomatic effects, such as vigilance drift, from chronic harms, including skill loss and narrowing autonomy.The exercises consolidated the expertise erosion cascade and produced initial Worker, Technology, and Organization concern clusters.Source: §A.1.1 · p. 20
- WS4 organized drawings and ST traces through the Sense–Contain–Recover triad and generated routines to strengthen sociotechnical immunity within existing workflows.Examples included AI-off alternate-planning drills, rationales for high-risk overrides, and living “do-not-automate” lists.Source: §A.1.1 · p. 20
- WS5 refined framework questions for embedding in existing routines, emphasizing Humane Automation, saved time, and early warning signs of skill atrophy.Participants reviewed scenarios and ST traces, with Worker-level prompts intended to help junior and senior staff notice warning signals.Source: §A.1.1 · p. 20
- Participants endorsed a visual expertise-erosion arc in which RadPlan initially amplified efficiency while quietly eroding expertise and identity over time.Drawing exercises and panel sequencing produced the arc depicting the AI-as-Amplifier paradox.Source: §A.1.1 · p. 20
A.1.2 Semi-structured Interviews.
Across a 12-month RadPlan deployment, 24 semi-structured interviews tracked participants’ experiences from baseline through routine use and, for some, the ST overlay. Interviews explored workflow, skill and oversight, social transparency, and the effects of AI integration on expertise, autonomy, and dignity.
A.1.2 Semi-structured Interviews.
- 24 interviews were conducted across months 2–3 (8), 5–7 (10), and 10–11 (6), capturing baseline practice, evolving use, and ST-overlay experiences where applicable.The guide remained stable but flexible and ran in parallel with workshops.Source: §A.1.2 · p. 20
- Interviews examined roles and workflow, changes in planning behavior and intuition, AI limitations, and participants’ informal verification and oversight practices.Prompts covered alternate plans, double-checking suggestions, early signs such as faster approvals or skipped exploration, and strategies to prevent skill rust; participants also described relying on peers, checklists, or judgment.Source: §A.1.2 · pp. 20–21
- After ST deployment, additional prompts examined how the 4W log shaped trust, disagreement, collaboration, and the balance between AI output and human judgment.All interviews also explored shifts in expertise, autonomy, and meaning at work, and suggestions for preserving skills and dignity alongside efficiency; sessions lasted 45–60 minutes, were recorded with consent, and transcribed using pseudonyms with identifying clinical details removed.Source: §A.1.2 · p. 21
A.1.3 Data Privacy and Governance.
Hospital Alpha’s privacy, governance, and IRB constraints prevent releasing artifacts that could expose proprietary interfaces or identifiable clinical information.
A.1.3 Data Privacy and Governance.
- Hospital Alpha’s privacy, governance, and IRB requirements bar releasing raw artifacts that expose proprietary interfaces or identifiable clinical traces.Figure 2 therefore shows a reconstructed interface with synthetic data, and workshop artifacts referencing internal systems or specific cases are also restricted.Source: §A.1.3 · p. 21
A.2 Metrics and Evaluation Plan
The framework pairs measures of reliance, skill retention, and dignity with a phased evaluation and iterative rollout. Pilot comparisons and feedback should guide refinement, adaptation, and gradual scaling across teams and domains.
A.2 Metrics and Evaluation Plan
- Track appropriate reliance, de-skilling indicators, and dignity signals, including trust aligned with AI accuracy, manual skill checks, autonomy, and willingness to critique AI.The proposed indicators also include error rates without AI and self-reported engagement.Source: §A.2 · p. 21
- Pilot deployments should compare framework-using teams with teams without it, tracking skill retention and decision quality over time.Organizations should use metric outcomes to refine questions and practices.Source: §A.2 · p. 21
- Roll out iteratively: start in one department, let champions tailor starter questions, assess after 3–6 months, then adapt and scale gradually.Test usability across industries with minimal adaptation, keeping the framework practical for routine check-ins, design reviews, and training while still surfacing deeper issues.Source: §A.2 · p. 21