This post is the third 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 third installment clarifies what it means for a job to be “exposed” to AI and why high exposure does not automatically imply automation, job loss, or deskilling. Recent measurement frameworks converge on a core insight: the relevant unit of analysis is not the occupation, but the tasks inside jobs. Exposure metrics therefore estimate where AI could matter, while real-world impact depends on adoption frictions, complementary human tasks, organizational choices, and governance.
Three empirical patterns are especially consistent across prominent sources. First, exposures are often highest in clerical, administrative, and content-intensive work, and can be substantial in many high-education occupations when measurement focuses on language and information processing tasks. Second, observed near-term effects frequently look like augmentation with uneven distribution, where outcomes vary by occupation and experience, with copilots sometimes helping less-experienced workers perform tasks more quickly while more-experienced workers may use AI more selectively, and where risks can shift toward monitoring, work intensification, over-reliance, or quality failures. Third, historical and contemporary evidence warns against interpreting exposure as destiny: in recent data, AI exposure has not been associated with negative aggregate employment outcomes so far, and in some analyses, exposure correlates positively with employment growth.
Exposure is not automation
Across modern frameworks, exposure typically describes the share of tasks in an occupation that could be affected such as tasks whose completion time could plausibly fall with AI assistance, tasks whose outputs could be generated with AI, or tasks that become easier to do at scale (search, drafting, summarizing, classifying, translating). This differs from the stronger claim of automation, which implies that a system can reliably perform a task end-to-end within real workplace constraints (quality thresholds, liability, privacy, context, and integration into wider workflows). The distinction matters because many high-stakes domains require verification and interpersonal judgment that current systems cannot safely replace.
Empirically and theoretically, at least four mechanisms break the exposure → displacement chain:
Task bundling in jobs: Most occupations combine automatable and non-automatable activities. Even if one cluster is heavily exposed, the occupation may shift toward the remaining tasks rather than disappear.
Complementarity and induced demand: Productivity improvements can raise output, expand service demand, or create new complementary tasks (quality assurance, customer-facing work, compliance, domain specialization). The classic task-based literature formalizes this as a balance between displacement and “new tasks” that reinstate labor demand.
Adoption frictions and governance constraints: Work processes are sticky: integration costs, liability, data protection, and institutional rules can delay or limit automation even when technical capability exists. This is visible in cautious adoption in regulated contexts and in the centrality of risk management frameworks for deployment.
Distributional and organizational choices: Organizations can use the same tool as a copilot (augmentation) or as a headcount-reduction lever (substitution), with profoundly different outcomes for wages and autonomy. Survey evidence finds workers strongly anticipate both possibilities.
Tacit knowledge and invisible labor: Many jobs include informal, experience-based, or relational tasks that are difficult to observe or measure in the datasets used to estimate exposure. If this “invisible work” is undercounted, exposure metrics may overstate how much of a job can realistically be automated. Organizational research also suggests employers often lack full visibility into what workers actually do, which can lead to overestimating automation potential and underestimating the role of tacit knowledge in maintaining quality and reliability.
How exposure is measured
Most approaches begin with task datasets and then map tasks to AI capabilities.
The International Labour Organization’s 2025 refined Global Index of Occupational Exposure to Generative AI exemplifies this approach. It builds task-level automation potential scores using a large occupational task taxonomy (29,753 tasks in a national classification mapped to ISCO-08), human survey ratings, expert validation, and model-based predictions, then aggregates into four “exposure gradients” that incorporate both mean exposure and task variability. The ILO’s headline estimate is that about one in four workers globally are in occupations with some degree of GenAI exposure, but the highest exposure category (Gradient 4) covers a much smaller share of global employment (reported at 3.3%), with notable gender and income-group differences in exposure.
A second influential approach evaluates whether LLMs can substantially accelerate tasks. The OpenAI/University of Pennsylvania “GPTs are GPTs” framework defines exposure as tasks for which an LLM could reduce completion time by at least a given threshold, and extends estimates under assumptions about complementary software and tooling. It reports that roughly 80% of the U.S. workforce could have at least 10% of tasks affected, and about 19% could have at least 50% of tasks affected (under their definitions), while emphasizing that these are not adoption forecasts.
Some institutions explicitly separate exposure from automation risk. For example, the OECD reports a skills-and-abilities based risk-of-automation measure (including AI and other automation technologies) and estimates that occupations at the highest risk of automation account for about 27% of employment on average across OECD countries in their sample, while also noting that the set of jobs most exposed to AI can differ from the set most at risk of automation.
Similarly, the International Monetary Fund proposes a cross-country exposure measure and adds a “complementarity” dimension to distinguish jobs where AI is more likely to augment from those more likely to displace. The IMF estimates around 40% of global employment is exposed to AI, rising to about 60% in advanced economies, and suggests that roughly half of exposed jobs in advanced economies could face negative effects depending on complementarity and preparedness.
Sector snapshots of exposure and impact
The sector analyses below distinguish: (1) task-level exposure (what AI could affect), (2) current evidence of substitution vs. augmentation (what is happening), and (3) ethical risk distribution (who bears risk and why).
Law
Legal work spans information-intensive tasks (such as research, summarization of precedents, drafting contracts and briefs, document review, and compliance monitoring) and socially embedded tasks (such as client counseling, negotiation, courtroom pleading, and accountability for advice). Exposure is therefore expected to be high for document-heavy work and lower for high-liability judgment and interpersonal components.
Task-based measures can diverge by definition. In the ILO global index, “Legal and related associate professionals” appear with moderate exposure levels (example mean exposure values reported near ~0.39 for related categories, with “Legal professionals n.e.c.” in a lower/minimal exposure grouping), reflecting the index’s task taxonomy and the fact that many legal tasks are not purely generative-text production.
In contrast, LLM-acceleration frameworks that focus explicitly on language tasks identify some legal-support roles as extremely exposed. “Legal Secretaries and Administrative Assistants” are listed among occupations labeled “fully exposed” in one GPT-powered software scenario within the “GPTs are GPTs” taxonomy.
Direct causal evidence for “AI replaces lawyers” remains limited; more robust evidence points to augmentation in specific contexts when tools are combined with training and oversight. A field study of legal-aid settings reported by Reuters found most participating legal aid attorneys who were given access to multiple AI tools perceived productivity gains and planned continued use, while also reporting concerns about accuracy and privacy.
Legal AI concentrates risks among clients least able to audit outputs. Core risks include: hallucinated citations or misstatements of law, confidentiality and privilege risks when sensitive data is shared with vendors, bias and homogenization in drafting and argument templates, and access inequities if better-resourced firms have better tools and governance.
Healthcare
Healthcare mixes high-trust interpersonal work (clinical judgment, communication, physical examination, procedures) with documentation, coding, scheduling, and information retrieval. This creates an exposure profile in which administrative and documentation tasks are a primary target for near-term generative AI leverage, while much direct care remains less automatable given embodied and contextual constraints. At the same time, other forms of AI, including imaging, diagnostic, and decision-support systems, are already highly deployed in certain specialties, meaning exposure varies substantially depending on the type of AI considered.
In the ILO global GenAI index, many health professions and hands-on roles appear relatively low in exposure compared to clerical work, reflecting the fact that the index focuses on generative AI affecting language and information-processing tasks. This should not be read as low exposure to AI overall, as several medical specialties, particularly those relying on imaging, pattern recognition, or clinical decision support already show high exposure to other forms of AI. Regulatory data further illustrate this variation: among AI-enabled medical devices authorized by the U.S. Food and Drug Administration in 2023, the large majority were concentrated in radiology (79%), with smaller shares in cardiovascular (9%), neurology (5%), gastroenterology/urology (4%), and anesthesiology (2%), reflecting particularly high exposure to image-recognition and diagnostic AI in certain specialties.
At the same time, LLM task-based indices can flag high exposure for healthcare-adjacent information roles and for documentation-related task clusters.
Available evidence most strongly supports augmentation through reduced documentation burden. The American Medical Association reports survey findings that physicians identify administrative burden reduction as the leading opportunity for AI, with a majority emphasizing automation of non-patient-facing work.
Peer-reviewed and open-access syntheses of AI “scribe” systems report promising effects on workflow and documentation outcomes, including improvements in efficiency and clinician experience, typically under continuing human oversight.
Healthcare concentrates ethical risks among patients whose data is most sensitive and among clinicians who can be held responsible for AI-mediated errors, whether the system involves generative tools, diagnostic models, or embedded medical-device software. Key risk axes include: privacy (PHI exposure), bias (unequal performance across subpopulations), automation bias (overreliance on suggested diagnoses), and deskilling if documentation and interpretation tasks are offloaded without redesigning training.
Creative work
Creative occupations blend ideation, iterative production, and client and audience alignment. Generative tools target drafting and editing text, generating image/audio/video assets, storyboarding, variations, localization, and rapid iteration. Yet value is often anchored in tastes, originality signaling, social reputation, and client trust.
LLM exposure indices often place writing and translation among the most exposed occupations. In the “GPTs are GPTs” results, interpreters/translators and writers/authors rank among highest-exposure categories under multiple measurement variants.
The ILO refined index similarly flags linguistic and media-adjacent occupations as exposed in higher gradients (for example, translators appear in a higher exposure gradient), while also noting that exposure does not imply full automation and that most jobs will likely be transformed rather than eliminated.
Unlike many sectors where evidence is mainly short-run productivity experiments, creative labor markets already show measurable substitution pressure in some segments, especially in platform-mediated freelance markets where price competition is intense. A large-dataset study of online labor demand reports substantial declines in postings for automation-prone writing and coding-related freelance jobs following the introduction of ChatGPT, and additional declines in demand for graphic design and 3D modeling following releases of image-generation tools.
Complementary evidence from an Upwork-based analysis frames these platform markets as “early detection” environments for AI labor effects because contracts are renegotiated frequently and tasks are short-horizon. A Brookings policy analysis summarizing related work reports modest but statistically meaningful declines in contracts and earnings in more GenAI-exposed freelance categories.
Creative AI has acute distributional risks: uncompensated training on creators’ work (IP and consent) and homogenization and erosion of stylistic diversity.
Administration and clerical work
Administrative work is task-dense: correspondence, scheduling, data entry, record keeping, HR support, payroll, customer communication, and compliance documentation. These tasks are frequently text-based, repetitive, and standardized, making them structurally exposed to both generative AI and agentic AI.
The ILO refined index finds clerical occupations remain among the most exposed globally. Examples in the highest exposure gradient include data entry clerks (reported mean exposure ~0.70), typists/word processing operators (~0.65), accounting and bookkeeping clerks (~0.64), general office clerks (~0.60), and personnel clerks (~0.60).
Similarly, LLM-centric measures identify clerical communication roles and administrative legal support as highly exposed; in the “GPTs are GPTs” table of high-exposure occupations, multiple correspondence and reporting roles are near the top, and “Legal Secretaries and Administrative Assistants” appear as “fully exposed” in one scenario.
Administrative work also has some of the clearest evidence for immediate augmentation. A field study of a generative AI conversational assistant in customer support found sizable productivity gains on average, with especially large gains for novice or less-skilled workers, and mixed effects for the most experienced workers. Controlled experiments on professional writing tasks similarly show productivity gains and reduced inequality in output quality/time when participants have access to generative AI tools.
At the same time, this sector is unusually vulnerable to organizational substitution decisions because tasks are modular and often benchmarked by throughput metrics. Substitution pressure is therefore more plausible here than in high-liability professional judgment roles, even if the initial deployment is a copilot.
Administrative AI concentrates risks around surveillance and work intensification, bias in performance evaluation and HR decision-making, and privacy and security in internal communications. The OECD’s worker and employer surveys surface concerns about work intensity and data collection even alongside reported productivity benefits. Beyond GenAI copilots, algorithmic management systems raise psychosocial risk and monitoring pressure at scale: an EPRS/European Parliament summary reports substantial prevalence of algorithmic management exposures among EU workers and projects further increases, with risks including continuous monitoring and reduced autonomy.
Policy and organizational responses that make exposure non-destiny
Job redesign and reskilling
OECD survey summaries emphasize that training and worker consultation are associated with better worker outcomes when AI is introduced. From a task perspective, reskilling is most effective when paired with job redesign: shifting workers toward tasks where humans retain advantage (judgment, relationship management, exception handling, accountability) rather than merely accelerating throughput.
Procurement, evaluation, and monitoring limits
Institutionalizing safeguards often begins at procurement. NIST’s GenAI profile explicitly calls for updating procurement and vendor assessments to include IP, privacy, security, and related risks, and it frames ongoing monitoring as a necessary component of risk management. In parallel, algorithmic management research underscores the wellbeing risks of continuous monitoring and performance pressure, motivating enforceable limits on surveillance intensity and data use.
Regulation and collective bargaining
Formal regulation increasingly treats workplace AI as a rights-and-livelihood issue. The European Union AI Act text explicitly identifies employment AI systems as high-risk and references risks of discrimination and impacts on livelihoods and workers’ rights. In the U.S., the U.S. Equal Employment Opportunity Commission highlights how AI can produce intentional discrimination or disparate impact, and signals enforcement relevance across the employment lifecycle. The U.S. Department of Justice similarly warns that hiring technologies can result in disability discrimination even when the employer uses third-party tools.
Final Thoughts: Exposure Is a Call to Governance, Not to Panic
The question of which jobs are most exposed to AI is less about forecasting replacement and more about designing a governance response. Exposure metrics help clarify where AI tools could influence workplace tasks, but they do not predict outcomes. Whether exposure leads to augmentation, transformation, or displacement depends on how institutions, organizations, and workers act in response.
What history teaches us—and what the early evidence confirms—is that the impact of AI on jobs is not technologically predetermined. Exposure is shaped by policy, organizational choices/culture, design decisions, and power. This means the ethical challenge is not to halt technological progress, but to govern it in ways that protect human dignity and agency, support adaptation, and avoid deepening inequality.
In short: exposure is not destiny. But it is a signal that tells us where to focus our attention, our safeguards, and our governance efforts.
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.