This post is the fourth in a series titled “AI and the Future of Work,” hosted on the Notre Dame-IBM Technology Ethics Lab’s blog. In this series, we will examine the implications of AI for jobs and society: how tasks and roles might change, how organizations can adapt, how workers are affected, and what governance frameworks are being considered to create and support beneficial and responsible outcomes.
As AI tools move from pilot projects to routine workplace infrastructure, the core issue is not whether AI will “replace” work in the abstract, but whether organizations configure automation and augmentation across workflows. Rather than a binary choice, these approaches exist along a spectrum: some tasks may be fully automated with end-to-end machine execution, others may be AI-assisted with human oversight, and others deliberately remain human-led. These design choices determine where responsibility sits and how work is experienced. This distinction matters ethically because the same AI capability implemented in different ways can produce sharply different outcomes for worker autonomy, dignity, fairness, and distributional justice.
configure automation and augmentation across workflows. Rather than a binary choice, these approaches exist along a spectrum: some tasks may be fully automated with end-to-end machine execution, others may be AI-assisted with human oversight, and others deliberately remain human-led. These design choices determine where responsibility sits and how work is experienced.
This blog post develops a practical and values-informed approach to workplace AI. It begins by clarifying automation and augmentation as points along a continuum and defining deskilling and reskilling mechanisms. It then connects these design choices to normative labor values, and finally explores how institutions might operationalize “human oversight” through concrete governance models.
Automation and augmentation
In many conversations about AI at work, experts talk about automation as if it's inevitable. But there’s a difference between using AI to replace people and using it to support them. Whether AI leads to job loss or job transformation depends on how it’s introduced, how tasks are organized, what the tools are used for, and who makes the decisions. Rather than treating automation and augmentation as a binary, it is more accurate to understand them as points along a spectrum of design choices within workflows. A single job may include multiple tasks, some of which are fully automated, others partially automated with human oversight, and others deliberately preserved for human judgment and interaction. In practice, organizations are configuring mixtures of both across task bundles.
Automation is best understood as end-to-end task execution by an AI system within real operational constraints: accuracy and reliability thresholds, privacy and security requirements, liability boundaries, integration with systems of record, and robust handling of exceptions. Tasks that are most amenable to automation tend to share certain characteristics: they are rule-based or easily programmable, involve structured inputs and outputs, produce low-variance and highly predictable outcomes, and operate in environments where edge cases can be anticipated or tightly managed. Automation is also often pursued for tasks that are repetitive, cognitively or physically burdensome, or even hazardous for human operators. These characteristics make full delegation to AI systems both technically feasible and organizationally attractive.
Augmentation describes AI use that improves speed, quality, or accessibility of work while preserving human responsibility for outcomes. In practice, augmentation often looks like drafting, summarizing, recommending, or detecting, followed by human verification and contextual judgment. This is consistent with expert analyses suggesting that, for generative AI in particular, the dominant near- to medium-term effect is often task-level transformation rather than occupational automation. Augmentation can also support reskilling when it is implemented as a learning system rather than a replacement system.
Reskilling and upskilling are reinforced when humans are trained not only to verify and calibrate AI output quality, identify error modes, and handle escalation, but also to develop relational capacity with AI systems such as learning how to effectively prompt, steer, and collaborate with these tools to achieve desired outcomes. This form of human–AI partnership goes beyond technical proficiency and becomes a distinct competency in how work is performed. At the same time, workers can transition toward tasks that the organization is intentionally protecting for humans, such as relationship work, accountability-bearing judgment, complex negotiation, and governance.
Importantly, the difference is not semantic. Automation reallocates control and accountability; augmentation redistributes cognitive load and can either widen or narrow skill gaps depending on how the system is introduced, who is trained, and how performance is measured.
Deskilling and reskilling
Discussions of AI in the workplace often focus on performance risks, such as hallucinations or bias in system outputs. But a governance lens that centers work must also account for how these systems reshape the skills required to perform tasks over time. As AI tools take on portions of cognitive work or restructure workflows, they redistribute expertise, potentially eroding some skills while amplifying or creating others. The central question, then, is not only whether AI systems perform accurately, but whether the resulting changes in human capability are intentional, desirable, and beneficial in the long run.
Deskilling is commonly defined as a reduction or erosion of existing skills due to technology-mediated changes in how tasks are performed, often through the routinization of judgment, fragmentation of tasks, or the replacement of human discretion with scripted or model-driven outputs.
In AI-mediated settings, this process operates across multiple, interrelated dimensions. First, technical deskilling refers to the erosion of hands-on or domain-specific capabilities when systems take over execution (for example, reduced practice in diagnostic interpretation or independent analysis). Second, cognitive deskilling captures the offloading of core mental processes, including interpretation, critical thinking, problem-framing, creativity, and learning through iteration or failure, as workers increasingly rely on AI-generated outputs rather than generating or interrogating them. Third, structural or normative deskilling reflects broader shifts in responsibility and professional judgment, where decision-making authority becomes embedded in systems, potentially weakening habits of accountability, deliberation, and ethical reasoning. Framing deskilling across these dimensions clarifies that the issue is not only the loss of technical proficiency, but also the reconfiguration of how individuals think, learn, and exercise judgment within AI-mediated work systems.
Upskilling and reskilling are distinct mechanisms: while upskilling focuses on expanding capabilities to perform current responsibilities more effectively, reskilling focuses on preparing workers for different roles as tasks and job structures change. However, deskilling is rarely a direct property of the model and is typically an emergent property of workflow design and incentives. Three recurring pathways are especially relevant:
-
Dequalification through default outputs: When workers are evaluated primarily on throughput, AI-generated drafts can become the default, and the human role collapses into fast approval or light editing. Over time, this reduces opportunities to practice core skills such as writing, reasoning, and synthesis.
-
Automation bias and overreliance: If organizations treat AI suggestions as presumptively correct, workers may defer rather than deliberate. This can erode domain reasoning and situational awareness, which raises both safety and accountability risks.
-
Task fragmentation and loss of craftsmanship: As AI systems absorb central tasks such as drafting, classification, and first-pass interpretation, humans may be left with narrow edge cases to handle. This can create brittle skill pipelines where junior workers never develop foundational competence because early-career practice space is automated away.
Ethical analysis for workplace AI
A governance approach that focuses only on technical performance—accuracy, reliability, efficiency—risks overlooking the ethical core of workplace AI: the quality and meaning of human work, and the fairness with which benefits and burdens are distributed. At stake are both short- and long-term productivity metrics, but also questions of dignity, autonomy, and justice in how AI systems reshape roles, expectations, and labor conditions.
Human dignity, agency, and autonomy
In workplace contexts, dignity, agency, and autonomy concepts capture distinct but related dimensions of human-centered work. Dignity refers to the conditions under which work respects the inherent worth of the individual, including non-degrading treatment, fair evaluation, and the preservation of meaningful roles. Autonomy concerns the degree of control individuals have over how they perform their tasks, including discretion in decision-making, pacing, and problem-solving. Agency, by contrast, refers to the capacity to act intentionally within a system—to make choices, influence outcomes, and exercise judgment—even within constraints. While autonomy is about the space for decision-making, agency is about the capacity to act within that space.
In AI-mediated workplaces, these dimensions can be affected in different ways. Dignity may be undermined when work becomes overly surveilled or reduced to mechanistic outputs; autonomy may be constrained through rigid, system-driven workflows or choice architectures; and agency may be weakened when workers are expected to defer to system outputs without meaningful opportunities to question, override, or reinterpret them. OECD analyses of AI in workplaces explicitly treat agency and dignity as ethical risk domains and warn that AI systems can undermine them through excessive monitoring, constrained choice architectures, and opaque performance evaluation.
From an oversight perspective, respecting dignity and autonomy means designing roles where humans have meaningful control, rather than symbolic approval. When people are asked to formally approve decisions they didn’t shape or fully understand, they bear responsibility without real authority. This creates accountability gaps and can lead to moral injury, especially when systems cause harm and human overseers are held liable for outcomes they couldn’t influence.
Fairness and non-discrimination
Workplace AI can produce disparate impacts through biased training data, proxy variables, and uneven performance across groups, particularly in hiring, scheduling, promotion, testing/credentialing, and performance scoring. OECD work emphasizes fairness and bias as central ethical risks and highlights the importance of transparency, explainability, robustness, and accountability.
NIST’s risk management framing similarly centers trustworthiness characteristics (including fairness and harmful bias management, validity and reliability, accountability, and privacy-enhancement) as measurable properties that organizations should manage across the AI lifecycle.
Distributional justice
Distributional justice concerns who benefits from productivity gains and who bears transition and monitoring costs. The IMF’s policy analysis anticipates substantial AI exposure in advanced economies and warns that without preparedness AI can widen inequality even when aggregate productivity rises.
In practice, distributional injustice can appear even in augmentation deployments: workers may absorb the burden of verification, exception handling, and reputation risk, while productivity benefits accrue elsewhere in the organization. Governance must therefore evaluate not only whether AI improves output, but whether it does so in a way that preserves worker bargaining power, mobility, and the possibility of advancement.
Human oversight governance that works
Human oversight must be treated as a structured and resourced system. This includes clearly defined governance roles, sociotechnical mechanisms for intervention, adequate training for those tasked with oversight, and accountability pathways when things go wrong.
A useful foundation comes from the human oversight typologies outlined in the European Commission’s trustworthy AI guidelines, which distinguish between three levels of involvement: human-in-the-loop (where human approval is required before the system acts), human-on-the-loop (where a human can intervene during or after the system operates), and human-in-command (where humans retain ultimate authority over system goals and deployment). These distinctions help clarify the depth of control humans have—and the ethical adequacy of that control—across different use cases.
-
Human-in-the-loop (HITL) oversight means that a human intervenes in every decision cycle. This model is appropriate when each output carries significant consequences, requiring individualized judgment and accountability. Examples include clinical documentation sign-off in healthcare, final review of legal filings, and decisions related to hiring, promotion, or termination. Here, automation supports, but does not replace, professional discretion.
-
Human-on-the-loop (HOTL) refers to supervisory oversight through periodic monitoring, sampling, or auditing rather than constant, per-instance control. This model suits high-volume contexts where errors can be flagged and escalated reliably. Examples include automated quality inspection pipelines in manufacturing, and AI-suggested replies in customer support, where human agents monitor for anomalies but are not required to intervene at every step.
-
Human-in-command (HIC) emphasizes that humans retain strategic authority over whether and how a system is used in the first place. It is especially critical for tools that influence rights, dignity, or power asymmetries such as algorithmic management systems, workplace surveillance, or automated hiring screeners. HIC governance ensures that deployment itself is a deliberative choice, not a default, and that workers and managers alike can understand, contest, or refuse harmful uses.
A workable oversight system requires clearly defined protocols that specify who is responsible for monitoring, when and how intervention should occur, what thresholds trigger human review, and how decisions are documented and audited. These protocols must be embedded in daily workflows and supported by training. Without operational clarity, oversight becomes fragmented or symbolic, therefore increasing the risk of harm, ambiguity in accountability, and erosion of trust among workers affected by AI-driven decisions:
-
Verification protocols: Specify what must be checked (facts, citations, safety-critical claims, privacy leakage), at what rate (100% in some contexts, sampling in others), and with what competence (trained reviewers with authority to override), and against which quality thresholds (clearly defined performance benchmarks such as accuracy rates, acceptable error margins, or “minimally viable” standards for deployment). Making these thresholds explicit is critical to avoid implicit or shifting definitions of what counts as “good enough,” particularly in contexts where errors carry asymmetric or compounding risks.
-
Escalation pathways: Define triggers (low-confidence outputs, conflicts with authoritative sources, out-of-distribution cases, user complaints) and the escalation ladder (second reviewer, specialist, temporary rollback, vendor incident process). These mechanisms are especially important in GenAI systems where confabulation and groundedness failures are well documented.
-
Accountability rules: Oversight must specify who is responsible for harms and who has power to intervene. Algorithmic management literature highlights that automated managerial decisions can create unclear accountability boundaries, particularly when firms rely on vendor systems. These challenges are likely to intensify as organizations adopt more complex AI ecosystems, including multi-agent systems and third-party agentic tools that interact across workflows. In such environments, decision-making becomes distributed across multiple systems and actors, further diffusing responsibility and complicating traceability. OECD’s evidence base on algorithmic management notes both potential consistency benefits and documented detrimental effects on workers, underscoring the need for clear, enforceable governance structures that can operate under these increasingly layered and interconnected conditions..
Considerations for dignity-preserving deployment
If AI is treated as workplace infrastructure, a central design question is how to balance productivity gains with the preservation of dignified work. The considerations below are framed as ways organizations might navigate the trade-offs between automation and augmentation, drawing on deskilling/reskilling dynamics and risk management approaches reflected in National Institute of Standards and Technology frameworks and labor-policy evidence from International Labour Organization, Organization for Economic Co-operation and Development, and International Monetary Fund.
-
Job redesign before headcount redesign: What would it look like to introduce AI as an explicit component of job design, rather than layering it onto existing roles and expecting workflows to adjust organically? In this approach, organizations would define in advance which tasks remain human-led (e.g., judgment, relationship management, accountability), which are AI-assisted, and which may be automated but still require structured verification. This framing aligns with evidence that AI more often reshapes task composition than eliminates occupations outright.
-
Training and reskilling pathways that preserve ladders: How might training be structured if the goal is not only tool adoption but the preservation of skill development pathways over time? Beyond technical proficiency, this could include verification capabilities, awareness of model limitations, escalation protocols, and role-based accountability. It also raises the question of whether organizations treat workforce development as a short-term adjustment or as an ongoing strategic function, particularly in light of research emphasizing the importance of sustained investment in learning systems.
-
Dignity metrics: What indicators would be needed to evaluate AI systems not only in terms of performance, but also in terms of their impact on the quality of work? If evaluation is limited to productivity or cost metrics, effects such as deskilling, reduced autonomy, or opaque decision-making may be interpreted as efficiency gains. Expanding measurement frameworks could involve tracking dimensions such as autonomy retained (decision-making discretion), verification load (time spent checking outputs), escalation health (effectiveness of raising and resolving concerns), fairness signals (variation in outcomes across groups), and surveillance intensity (scope and consequences of data collection). The challenge, then, is not only identifying these indicators but determining how they are operationalized and acted upon within governance systems.
Final thoughts
Workplace AI does not dictate a singular future. The same underlying system—a generative model, a recommendation engine, a classifier—can expand human capability and reduce drudgery, or it can accelerate deskilling, intensify surveillance, and displace responsibility without recourse. What determines the difference is not the technology alone, but the design and governance choices that shape how it is used: how tasks are restructured, how oversight is implemented, and how benefits and risks are distributed across workers, organizations, and sectors.
Perla Khattar is a J.S.D. candidate at the University of Notre Dame Law School and a Research Fellow at the Notre Dame-IBM Tech Ethics Lab. In her research, she analyzes emerging technologies such as synthetic data, neurodata, quantum computing, and AI through a privacy lens, focusing on how law and policy can anticipate technological disruption while safeguarding individual rights.