This post is the fifth 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.
This fifth installment in the series examines how AI is reshaping two deeply consequential domains of modern work: who gets hired and how work is managed once individuals enter the workplace. As organizations increasingly integrate AI-driven systems into recruitment pipelines and day-to-day management practices, these technologies continue to become embedded in decisions that shape access to opportunity, conditions of work, and professional trajectories. This post takes a grounded approach, focusing on what these systems can do in practice, how recurring patterns of bias and potential harm can emerge across different stages, and why these risks are not always visible at the point of deployment.
From human judgment to algorithmic decision-making
Hiring has arguably never been a neutral or purely meritocratic process. Decades of research have documented how human decision-making in recruitment is shaped by heuristics, implicit bias, informal networks, and structural inequalities. Interviewer impressions can be inconsistent, similarly qualified candidates may be evaluated differently, and access to opportunity is often mediated by factors unrelated to job performance.
The introduction of AI into hiring and workplace management is therefore not occurring in a vacuum. These systems are often adopted in response to perceived limitations of human judgment: to introduce consistency, process large applicant pools, and reduce reliance on informal or subjective criteria. This framing matters because the relevant comparison is not between biased AI and unbiased humans, but between different systems of decision-making that produce different types of errors, trade-offs, and distributions of opportunity.
Why hiring and management make AI ethically high-stakes
The EU’s AI Act classifies AI used for recruitment and worker management/monitoring as high-risk, explicitly noting impacts on livelihoods, career prospects, and workers’ rights. Hiring and algorithmic management are high-stakes because small statistical differences can convert into life-changing outcomes at scale: who sees an opportunity, who is shortlisted, who gets coached, who is penalised, who is promoted, or who is terminated.
A second reason is that AI in hiring frequently begins before the interview. Recommendation and targeting systems can shape who even learns about a role, and which applicants are encouraged (or discouraged) from applying. Research and enforcement actions around ad delivery and targeting have shown that digital systems can produce skewed exposure even without an employer explicitly selecting protected traits as optimization can learn proxies and reproduce structural patterns.
A third reason is that algorithmic management can convert measurement into pressure. The International Labor Organization defines algorithmic management as systems using tracked data and other inputs to organize, assign, monitor, supervise, and evaluate work, sometimes using AI but other times using simple rules-based automation. That distinction matters: we can see serious worker-rights impacts even when no machine learning or other AI system is involved. In practice, these systems often translate continuous measurement into performance thresholds and behavioral incentives. Workers may be required to meet dynamically adjusted productivity targets (such as items processed per hour, call handling times, or delivery speeds), with automated alerts, warnings, or penalties triggered when metrics fall below expected levels. In some settings, real-time dashboards and monitoring tools create constant visibility into performance, encouraging workers to prioritize what is measured, even when it conflicts with quality, safety, or professional judgment. This can also produce anticipatory pressure: workers may adapt their behavior preemptively to avoid negative flags, reduce breaks, or conform to system-defined optimal patterns, even in the absence of direct managerial intervention.
While much of the discussion focuses on risks, it is important to recognize why organizations adopt these systems in the first place. Compared to human decision-making, AI systems can introduce consistency across evaluations, reduce reliance on informal or ad hoc judgments, and process large applicant pools that would otherwise be screened using coarse heuristics or under significant time pressure. In some contexts, structured and validated tools may reduce certain forms of individual-level variability or bias, particularly where human decision-making is highly subjective or inconsistent.
However, these potential advantages are not inherent to AI systems. They depend on how systems are designed, trained, and deployed. The same features that enable scale and consistency can also reproduce bias systematically, making errors less visible and more difficult to contest.
Where AI shows up across the recruitment pipeline
Most organizations using AI in their hiring pipeline buy a stack of tools, each claiming to reduce time-to-hire or improve talent quality while introducing its own measurement choices and potential biases. Public-facing categories commonly include:
-
Sourcing and advertising: Tools that target job adverts, recommend candidates, or “rediscover” past applicants can shape who is in the funnel in the first place. As studies have shown, even when advertisers cannot explicitly target protected characteristics, delivery optimisation can still skew outcomes, raising fairness and transparency questions about who is reached and why.
-
CV parsing and ranking: Applicant tracking systems can extract features from CVs (such as education, job history, skills) and produce a score or rank. If models are trained on historical hiring outcomes, they can learn patterns that reflect past bias or structural inequality, even if protected traits are removed, because proxies remain (such as career breaks, postcode, institution, gaps in employment, volunteer activities, professional memberships, etc).
-
Assessments and “fit” scoring: Pre-employment tests can include psychometrics, game-based assessments, coding tests, and automated scoring of written responses. The ethics hinge on validity (does it measure job-relevant skills?), accessibility (does it disadvantage candidates with disabilities?), and whether the tool is used as advice or as a gatekeeper.
-
Interview automation: "One-way" video interviews and automated analysis have been an area of particular scrutiny. At least one major vendor removed an automated facial analysis component from its platform following reassessment of the feature's evidentiary basis and concerns raised by researchers and privacy advocates, representing an example of the field responding to criticism through product revision.
-
Screening-out rules at scale: Sometimes the AI is not subtle at all; it is a high-throughput automated rule. For instance, systems that automatically filter out candidates with gaps in employment history can disproportionately exclude caregivers, individuals with health conditions, or those from non-linear career paths.
-
Background checks and risk signals: Tools that scrape online profiles, infer “risk”, or flag anomalies can produce opacity and due process issues, especially when candidates cannot see or correct errors. These systems also raise data-protection questions about lawful basis, minimisation, relevance, and retention.
Importantly, The U.S. Department of Justice has warned that AI-driven screening can violate disability discrimination law when tools screen out applicants with disabilities or fail to provide reasonable accommodations, including when employers rely on third-party vendors.
Algorithmic management and workplace monitoring
If hiring is the gateway, algorithmic management is the day-to-day reality: how tasks are assigned, how performance is measured, and how decisions about rewards and sanctions are made. The Organization for Economic Co-operation and Development has documented how widespread these systems have become. In an OECD employer survey summary on prevalence, a very high share of managers in the United States reported adoption of at least one tool used to instruct, monitor, or evaluate workers, and adoption is substantial across the surveyed countries.
Algorithmic management often includes automated scheduling and shift allocation; real-time productivity tracking (keystrokes, activity logs, call handling times, warehouse scans); performance scoring and benchmarking against targets; automated nudges, warnings, and escalation; “dynamic” pay or piece rates in platform and logistics contexts.
The empirical record on electronic monitoring and worker well-being is mixed but consistently warns against treating these technical applications as benign. A meta-analysis on electronic monitoring found small but measurable associations with lower job satisfaction and higher stress, and noted that targets and feedback regimes can intensify negative effects.
In the United Kingdom context, the Information Commissioner's Office has emphasised that monitoring at work should be purpose-limited, proportionate, and compliant with data protection law, noting increased prevalence and technological expansion. The point is not that monitoring is always forbidden, but rather that monitoring is a higher risk form of processing that must be justified, designed carefully, and explained clearly.
How bias and harm happen in practice
Disputes about “algorithmic bias” often stagnate because people imagine bias as a single bug you can remove. In practice, bias and harm in hiring and management arise from interacting mechanisms documented across multidisciplinary research:
-
Historical-data inheritance: Models trained on past hiring or “top performer” labels can reproduce patterns of exclusion, occupational segregation, or biased evaluation, especially if performance metrics already reflect unequal opportunity (who got mentorship, which assignments were offered, whose mistakes were forgiven).
-
Proxy measurement: Even when protected characteristics are excluded, other features can act as proxies (such as education pathways, location, gaps, accent, disability-related signals). This is why “we didn’t collect race/gender” is often not a fairness guarantee.
-
Accessibility and disability discrimination: Screening tools can disadvantage candidates with disabilities when they rely on speed, speech patterns, eye contact, or rigid interaction formats.
-
Feedback loops and chilling effects: People change behavior when they know they are being watched or scored. In hiring, perceived bias can reduce trust and deter applicants. In management, constant monitoring can shift work toward what is measured (not what matters).
A recurring challenge across these mechanisms is that many of these risks are not visible at the point of deployment. Systems may appear to function effectively based on aggregate performance metrics, vendor-provided validation, or internal testing that does not capture downstream impacts. Proxy variables may not be immediately identifiable, feedback loops may take time to materialize, and affected individuals often lack visibility into how decisions are made. As a result, harms can emerge gradually and remain difficult to detect, attribute, or contest within organizational settings.
What determines whether AI expands or constrains opportunity
The effects of AI in hiring and workplace management are not fixed. The same class of systems can either expand access to opportunity or reinforce existing inequalities depending on how they are designed and governed. Three sets of conditions are particularly important.
Technical design choices shape what is measured and optimized. Decisions about training data, feature selection, model validation, and performance metrics determine whether systems capture job-relevant criteria or encode proxies for structural disadvantage.
Organizational practices influence how systems are used in context. Procurement decisions, levels of human oversight, reliance on automated outputs, and the ability of workers or candidates to contest decisions all affect how risks materialize in practice.
Regulatory and governance frameworks set the boundaries within which these systems operate. Requirements around transparency, auditing, documentation, and accountability can shape incentives and create mechanisms for detecting and addressing harm.
What some jurisdictions are starting to require
In Europe, the EU AI Act classifies AI used for recruitment, selection and worker management and monitoring as high-risk systems. It also highlights that deployers (including employers) play a critical role because they understand context of use and can identify risks not foreseen at development time. The same regulation includes transparency-related expectations relevant to workplace settings, including obligations to inform individuals when they are subject to certain high-risk AI systems and specific workplace information requirements when employers deploy high-risk systems affecting workers.
In the UK, the UK Department for Science, Innovation and Technology has published guidance on responsible AI in recruitment, framing the problem as partly a procurement and assurance challenge: organisations need mechanisms to evaluate tools and to align deployment with the UK’s broader AI governance principles. This is notable because it treats responsible recruitment AI as a lifecycle practice: selection, contracting, deployment, and monitoring, rather than a one-off ethics review.
At the city level, New York City’s Local Law 144 is an example of a compliance model for automated employment decision tools: it requires a bias audit within a defined time window, public posting of results, and notice to candidates and employees. The law’s definition covers computational processes that produce simplified outputs (score/classification/recommendation) used to substantially assist or replace discretionary decision-making.
Additionally, the Office of the New York State Comptroller issued an audit in December 2025 assessing enforcement of Local Law 144 and reported weaknesses including complaint-routing issues, limited outreach, and gaps between agency review and potential non-compliance identified upon re-review. It also notes civil penalties are available for violations.
Final Thoughts
What emerges across hiring and algorithmic management is not a story about isolated tools, but about the gradual construction of decision-making infrastructures that shape access to work, conditions of work, and trajectories within it. The ethical challenge is therefore not reducible to fixing bias in a single model or improving accuracy in a single stage of the pipeline. Rather, it is about understanding how multiple sociotechnical systems—often procured, layered, and only partially understood—interact to structure opportunities and constraints for workers. In this context, accountability cannot remain diffuse. Organizations deploying these systems are not merely passive users of vendor technologies; they are active participants in configuring how these systems operate in practice, what signals they prioritize, and how much discretion remains with human decision-makers. As AI continues to move deeper into employment contexts, the central question is not whether AI will shape work—it already does—but whether the conditions under which it operates are structured to expand opportunity, preserve worker dignity, and ensure meaningful accountability.
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.