Public discourse swings between catastrophism and dismissal. The statistics most cited in both directions measure potential exposure to AI, not any realized change in employment. We went to the actual data.
In public discussion around artificial intelligence (AI) and labor, the narrative swings back and forth between alarmist warnings and glib dismissals, with figures like "300 million jobs at risk" quoted without question. But those statistics refer to the potential for displacement by AI, not the actual changes to employment figures. This project approaches that gap between perception and reality as a humanities question, one about how a society interprets technological change, through the lens of employment statistics and AI exposure scores. Comparing employment data to peer-reviewed scores that measure a job's potential for automation, we explore which jobs are actually undergoing significant change, and what data tells us about how the story of AI "replacement" holds up.
This project examines the intersection of economics and humanity in two ways. First, it asks whether AI replaces more jobs than it creates, as is often claimed. Second, it asks how a society narrates and interprets technological change, who controls the narrative, and how labor is made visible (or rendered invisible) in the process. Two key principles underpin this investigation: data is never neutral and technology doesn't just happen to a society. It is designed and deployed in ways that align with the power structures in which it exists. Together, these ideas allow us to frame the question of AI not just in terms of what it is doing to our jobs but also of the stakes of that process. Who benefits? Who loses? And why?
Our primary framework is Data Feminism (D'Ignazio and Klein). It gives us tools to ask who produces data, whose interests it was designed to serve, and whose labor is rendered invisible in the process of counting. Applied here, it means treating the Bureau of Labor Statistics not as a neutral mirror of the economy but as a system that counts certain kinds of work and systematically misses others: gig workers, informal laborers, undocumented workers. The exposure scores demand the same scrutiny: produced in part by GPT-4 rating each occupation's susceptibility, they are an AI system's account of its own economic reach, and we treat them as objects of analysis as much as sources of evidence.
Our secondary framework is Science and Technology Studies (STS), which treats technologies not as autonomous forces acting on society but as products of the social relations and power structures in which they are built and deployed. The narrative that "AI is taking the jobs" casts technology as an agent with effects that are inevitable and evenly distributed; our data complicates that story at almost every turn. We also draw on Boyd Davis, Vane, and Kräutli, who argue that the ethical commitment of a humanities visualization is not to maximize trust in the data but to reveal where that trust should be withheld, a principle we apply by flagging what our dataset cannot tell us as clearly as what it can.
Putting these principles into practice means starting with the data, and being honest about its limits. Our analysis joins two open-access datasets on shared Standard Occupational Classification (SOC) codes. The first is the U.S. Bureau of Labor Statistics' Employment Projections: 2024–34 and Occupational Employment and Wage Statistics, which cover current employment, median wages, and ten-year projected growth for hundreds of occupations. The second is the "GPTs Are GPTs" dataset (Eloundou et al., published in Science), which scores each occupation's exposure to generative AI. After cleaning and merging the data in Python, crosswalking the exposure file's eight-digit O*NET-SOC codes to the Bureau's six-digit SOC codes, we test one legible question: are the occupations most exposed to AI the ones projected to shrink (replacement) or to grow (augmentation)?
We merged two open-access datasets on Standard Occupational Classification (SOC) codes to test whether AI exposure predicts employment change. All cleaning was done in Python using pandas.
Exposure Score. Eloundou et al. publish several measures; we use the β score (E1 + 0.5·E2), which counts a task as exposed when AI access at least halves the time to do it. For each occupation we average the human-annotator and GPT-4 β ratings into a single 0–1 value, keeping both so their disagreement can be examined.
SOC Code Crosswalk. The exposure file uses eight-digit O*NET-SOC codes; BLS uses six-digit SOC codes. Each exposure record was mapped to its six-digit SOC parent, averaging where several O*NET detailed occupations roll up to one BLS occupation (for example, home-health and personal-care aides).
Joining the Projections. Each occupation's 2024 employment and projected 2024–34 percent change were taken from the BLS Employment Projections and Occupational Outlook Handbook (2024–34 release, 28 Aug. 2025) and joined to the exposure scores on the six-digit SOC code.
Visualization Sample. Plotting all 900-plus scored occupations would be unreadable, so the charts present a thirty-occupation cross-section chosen to span the full exposure range and the major occupational sectors, from data entry to the building trades. Every value shown is traceable to the two public sources above.
Transparency is the essence of ethical data work. According to Boyd Davis, Vane, and Kräutli, each step from "data-cleaning" to "rendering in visualizations" should be considered "a significant transformation," and not a neutral process (Boyd Davis et al. 524). Every crosswalk, average, and selection was a decision, and recording those decisions allows readers to audit our judgment rather than trust it.
There is no such thing as a neutral dataset. Hepworth and Church write that visualizations are "inherently rhetorical, and therefore bias-laden visual artifacts" (Hepworth and Church). Choosing what to count is an argument already. The exposure scores carry a bias of their own: they were generated in part by GPT-4 rating how reachable each job is by tools like GPT-4, and the machine and the human annotators do not always agree. GPT-4, for instance, rated medical transcription as far more automatable (β = 0.88) than the human annotators did (β = 0.23). We treat the score as an object of analysis, not a verdict. Our most serious limitation, though, is who gets counted. Neither of the sources counts any job that is not considered official by the government. This omission reflects a long tradition of the "omission of non-dominant social groups" in the source data (Boyd Davis et al. 524). It means that our analysis probably underestimates the effects of AI on the workers who are the most vulnerable. We treat that absence as a finding rather than a footnote, since the responsibility of honest data work is "to reveal the extent to which [the data] may be untrustworthy" (Boyd Davis et al. 524).
Each circle is one occupation. Size reflects current number of workers. Hover for details. If high exposure reliably caused job loss, every circle on the right side of the vertical line would sit below zero. They do not.
Scatter plot summary: across thirty occupations, AI exposure score on the horizontal axis shows no consistent relationship with projected employment change on the vertical axis (Pearson r = −0.19). High-exposure occupations fall both above and below the zero line. Data entry keyers (0.70) and telemarketers (0.67) are projected to decline sharply, while data scientists (0.67), web developers (0.78), and writers and authors (0.84), all comparably exposed, are projected to grow.
| Occupation | Sector | AI exposure | Proj. change | Employment 2024 |
|---|---|---|---|---|
| Writers & Authors | Legal & Creative | 0.84 | +4% | 135,400 |
| Web Developers | Computing & Data | 0.78 | +8% | 86,000 |
| Statisticians | Computing & Data | 0.75 | +9% | 32,200 |
| Data Entry Keyers | Office & Admin | 0.70 | -26% | 141,600 |
| Data Scientists | Computing & Data | 0.67 | +34% | 245,900 |
| Telemarketers | Office & Admin | 0.67 | -22% | 67,400 |
| Software Developers | Computing & Data | 0.66 | +16% | 1,693,800 |
| Customer Service Reps | Office & Admin | 0.64 | -5% | 2,814,000 |
| Tax Preparers | Business & Finance | 0.62 | +5% | 90,600 |
| HR Specialists | Business & Finance | 0.62 | +6% | 944,300 |
| Secretaries & Admin Assts | Office & Admin | 0.61 | -2% | 1,944,000 |
| Medical Records Specialists | Health Information | 0.59 | +7% | 194,800 |
| Bookkeeping Clerks | Office & Admin | 0.56 | -6% | 1,613,400 |
| Medical Transcriptionists | Health Information | 0.55 | -5% | 43,900 |
| Accountants & Auditors | Business & Finance | 0.54 | +5% | 1,579,800 |
| Insurance Underwriters | Business & Finance | 0.54 | -3% | 127,000 |
| Paralegals | Legal & Creative | 0.49 | 0% | 376,200 |
| Financial Analysts | Business & Finance | 0.48 | +6% | 368,500 |
| Graphic Designers | Legal & Creative | 0.46 | +2% | 265,900 |
| Lawyers | Legal & Creative | 0.45 | +4% | 864,800 |
| Registered Nurses | Care & Education | 0.36 | +5% | 3,391,000 |
| Elementary Teachers | Care & Education | 0.31 | -2% | 1,422,700 |
| Physical Therapists | Care & Education | 0.27 | +11% | 267,200 |
| Home Health & Personal Care Aides | Care & Education | 0.17 | +17% | 4,347,700 |
| Electricians | Trades & Manual | 0.15 | +9% | 818,700 |
| Nursing Assistants | Care & Education | 0.12 | +2% | 1,441,500 |
| Cooks, Restaurant | Trades & Manual | 0.11 | +15% | 1,460,200 |
| Plumbers | Trades & Manual | 0.08 | +4% | 504,500 |
| Construction Laborers | Trades & Manual | 0.04 | +7% | 1,457,000 |
| Janitors & Cleaners | Trades & Manual | 0.01 | +2% | 2,447,700 |
Ranked by projected employment change 2024–34. If the displacement narrative were correct, all bars would point left. Several point right.
Ranking summary: among the ten most AI-exposed occupations, projected employment change ranges from negative twenty-six percent for data entry keyers to positive thirty-four percent for data scientists. Both growing and declining occupations appear within the same high-exposure group; seven of the ten are projected to grow.
| Occupation | AI exposure | Proj. change |
|---|---|---|
| Data Entry Keyers | 0.70 | -26% |
| Telemarketers | 0.67 | -22% |
| Customer Service Reps | 0.64 | -5% |
| Writers & Authors | 0.84 | +4% |
| Tax Preparers | 0.62 | +5% |
| HR Specialists | 0.62 | +6% |
| Web Developers | 0.78 | +8% |
| Statisticians | 0.75 | +9% |
| Software Developers | 0.66 | +16% |
| Data Scientists | 0.67 | +34% |
Bars show average AI exposure per sector (left axis). Line shows average projected employment change (right axis). Sorted highest to lowest exposure. The pattern reveals where augmentation and displacement diverge.
Sector summary: Computing and Data carries the highest average AI exposure and the strongest projected growth, while Office and Admin Support carries nearly as high an average exposure but is the only sector with a net projected decline. Care and Education and the Trades carry the lowest average exposure and steady positive growth.
| Sector | Avg AI exposure | Avg proj. change | Occupations |
|---|---|---|---|
| Computing & Data | 0.72 | +16.8% | 4 |
| Office & Admin | 0.63 | -12.2% | 5 |
| Health Information | 0.57 | +1.0% | 2 |
| Business & Finance | 0.56 | +3.8% | 5 |
| Legal & Creative | 0.56 | +2.5% | 4 |
| Care & Education | 0.24 | +6.6% | 5 |
| Trades & Manual | 0.08 | +7.4% | 5 |
If AI were simply taking jobs, the data would line up neatly: the more exposed an occupation is to AI, the faster it ought to shrink. Chart 1 puts that expectation to the test, plotting each occupation's projected employment change from the Bureau of Labor Statistics against its AI-exposure score from Eloundou et al. The expected slope never appears. The points scatter across the entire range of exposure, with high- and low-exposure jobs alike landing both above and below the zero line; across these thirty occupations the correlation between exposure and projected change is essentially zero (Pearson r = −0.19, so exposure accounts for under four percent of the variation in where a job is headed). That scatter, the absence of any clean relationship between exposure and decline, is the finding the rest of this project builds on.
Chart 2 narrows the view to the ten most exposed occupations, the ones that should be most at risk if exposure alone decided their fate. They move in opposite directions. Data entry keyers (0.70) are projected to lose twenty-six percent of their jobs over the decade and telemarketers (0.67) twenty-two percent, while customer service representatives (0.64), a field of nearly three million, slip five percent. Yet sitting in the very same tier are data scientists (0.67), projected to grow thirty-four percent (the largest gain anywhere in our data); software developers (0.66) at sixteen percent; and statisticians and web developers at single-digit growth despite carrying some of the highest exposure scores we measured. The sharpest illustration is a single pair: data entry keyers and data scientists sit at almost identical exposure, 0.70 and 0.67, and the Bureau projects one to shed a quarter of its jobs and the other to add a third. On its own, then, an exposure score reveals almost nothing about where an occupation is headed. It tells us only that AI can perform some of the work, not what happens to the people doing it once AI is in place.
What does determine the outcome is whether workers can take AI up as a tool or are displaced by it, and that tends to track pay, education, and how heavily the job relies on human judgment. Data scientists (0.67) are projected to grow thirty-four percent and software developers (0.66) sixteen percent, even though their exposure scores sit right alongside those of the declining clerical jobs. In roles like these, AI absorbs the routine work (the drafting, the summarizing, the boilerplate code) and leaves the worker to handle judgment, design, and the decisions that still require a person. The technology becomes a lever rather than a substitute. This is precisely the distinction Eloundou et al. draw between exposure and displacement: a high score means AI can reach into the work, not that the worker is about to be replaced. Whether that reach becomes augmentation or replacement is decided by organizations, by how they redesign roles, what they cut, and whom they keep, far more than by the technology itself. That is the central lesson Science and Technology Studies brings to a dataset like this one: a tool's effects come from the social setting it enters, not from the tool alone.
Chart 3 shows the same pattern one level up, by sector. Computing and Data carries the highest average exposure of any sector we tracked, yet every occupation in it is projected to grow. Office and Admin Support sits just below it on exposure (telemarketers, data entry keyers, secretaries, bookkeeping and customer-service clerks) and is the only sector with a net projected decline. Health Information splits cleanly down the middle: medical records specialists, who increasingly manage AI-assisted systems, are projected to grow seven percent while medical transcriptionists fall five. Meanwhile the lowest-exposure sectors keep adding jobs: home health and personal care aides (projected to grow seventeen percent on more than four million workers), restaurant cooks (fifteen percent), physical therapists (eleven), and the building trades all grow, because their work depends on the physical presence and human contact that AI cannot reproduce. The disruption is real, then, but it concentrates in a single band of work: routine, information-handling jobs.
It is here that the data begin to reveal something the projections alone keep hidden. Ask who actually holds the high-exposure, shrinking jobs in the lower-right quadrant of Chart 1, and the answer is anything but random. By the Bureau's own labor-force data, secretaries and administrative assistants are roughly nine in ten women, bookkeeping and accounting clerks nearly as many, data entry keyers about eight in ten, and customer service representatives close to two-thirds (BLS, "Employed Persons by Detailed Occupation and Sex"). The high-exposure jobs on the growing side of the line (software developers, web developers, data scientists) run the other way, roughly four-fifths men and far more likely to require a college degree. This is exactly what Data Feminism is built to make visible. Its call to "examine power" asks who benefits from a technology and who absorbs its costs; its call to "make labor visible" insists that the work being cut be named rather than folded into a total. Seen this way, the sorting is neither neutral nor accidental: the losses fall hardest on workers who were already the most precarious and the least valued. AI is not inventing a new inequality so much as reproducing an old one through a new mechanism.
We should be just as clear about what the data cannot do. The BLS projections are forecasts, educated estimates rather than confirmed outcomes. More importantly, they count only formal, classified jobs; gig workers, independent contractors, and the undocumented never appear, and those are precisely the workers with the least protection when AI enters their workplace. The dataset's silence, in other words, does not merely miss the inequality; it deepens it. Boyd Davis, Vane, and Kräutli argue that an honest humanities visualization is not the one that leaves a reader most confident but the one that shows where confidence should be withheld, and we have tried to hold to that: these gaps are findings in their own right, a record of whose work the system was built to count.
A subtler limit runs through the comparison itself, and it is worth naming directly. The Bureau's 2024–34 projections were built from information available in June 2024, near the very beginning of the generative-AI rollout, and by the Bureau's own account they "approach AI in the same manner as other technologies," extrapolating from "relationships found in historical data" and assuming technological progress stays "in line with historical experience." The Bureau even cautions that "if technology advances much more rapidly than it has historically… BLS projection methods are unlikely to yield reasonable results" (BLS, "Employment Projections: Frequently Asked Questions"). In other words, the institution charged with forecasting the future of work has chosen, deliberately and transparently, not to model generative AI as a sharp break from the past. That makes our two measures fundamentally different kinds of claim: the exposure score is a forward-looking, GPT-4-assisted estimate of what AI could reach, while the projection is a conservative, trend-based account that may well lag a fast-moving technology. Read through Data Feminism, the gap between them is not noise to apologize for but a finding in its own right: even the most authoritative public account of American labor is a situated, methodologically cautious construction, and the distance between what AI is forecast to do and what the official numbers will yet admit is precisely the space where this debate now lives.
The popular story casts technology as the protagonist and workers as bystanders, swept along by something inevitable. Our data tell a different one. AI's effects cannot be read off an exposure score; they are uneven, shaped by context, and they trace lines of inequality drawn long before any of this began. So, to return to where we started: are the most AI-exposed jobs the ones disappearing? Not reliably. And who works in the ones that are? Overwhelmingly the people this economy has undervalued for a long time. The real question was never whether AI took the jobs. It is the quieter, harder one: whose jobs, and why theirs?
DIGHUM 100 · Scott Caddy · UC Berkeley · Summer 2026
Robin directed our group's research focus and research question, significance framing, and theoretical introduction. Robin acted as our group's final line of defense, ensuring all components aligned in the development of a singular argument. Robin is a DIGHUM 100 student at UC Berkeley who is interested in the connection between labor history and technology.
Etasha built and designed our group's project site, ensured accessibility standards, and cited all images and multimedia elements; she also supported the writing of the analysis and ethical considerations. Etasha is a DIGHUM 100 student at UC Berkeley who is interested in inclusive digital design.
Richard did all data work, including downloading the BLS Employment Projections 2024–34 and Eloundou et al. GPTs-are-GPTs dataset, and the crosswalking of the eight-digit O*NET-SOC code to the Bureau's six-digit SOC code in Python with pandas. Richard is a DIGHUM 100 student at UC Berkeley with a background in economics.
Niki built and embedded all visualizations for our group's project using Python and Plotly, transforming our merged dataset into charts showing the relationship between AI exposure and occupation sector growth and shrinkage. Niki is a DIGHUM 100 student at UC Berkeley interested in data justice and feminist digital humanities.
Yoonha wrote the data analysis addressing the research question, connecting visualizations to the theoretical framing and discussing what the data tells us, what it doesn't tell us, and why. Yoonha is a DIGHUM 100 student at UC Berkeley interested in the effect of narratives about technology on labor and policy.
Abdullah wrote the ethical considerations and annotated bibliography for the project. Abdullah is a DIGHUM 100 student at UC Berkeley interested in social implications of artificial intelligence.
An investigative data-journalism series documenting how gig and warehouse workers are hired, managed, and fired by algorithms with little or no human oversight, from Postmates and Instacart to Amazon. Its relevance for our purposes is that it represents a case study of actual AI effects on the labor market rather than potential AI exposure, showing what displacement means for individuals. For our project it is useful as a comparison since it grounds our data analysis in lived experience, and poses the Data Feminist question of whose termination is counted in official statistics and whose is omitted entirely. It informs our decision to foreground the exclusion of informal and gig workers from the BLS dataset in the limitations section.
The database and visualization project devoted to uncovering the hidden history of Indigenous slaves based on archival records. Although it is not related to the same subject matter as our project, it is a great example of a DH project handling official data on dominant groups and recovering the data on marginalized communities. In a similar way to Stolen Relations, our project works with official data and shows limitations in representing certain segments of workers in it; hence, the project informs our decision to consider BLS exclusion of gig and informal workers as one of our key findings instead of a minor issue.
A DH project that uses computational text analysis of historical American newspapers to trace the biblical citations in them and generate visualizations that make visible the patterns not noticeable to close reading. Its relevance for our project is that it exemplifies a good DH project in terms of transparency: Mullen provides public access to his code, data, and methodological decisions, and he discusses limitations of his dataset in a very clear way. This project serves as a warrant for our commitment to data transparency, and an inspiration for creating visualizations revealing what the data cannot say.
Data Feminism is the main theoretical framework we use in our project. According to D'Ignazio and Klein, data collection, analysis, and visualization are always non-neutral processes reflecting the power relations in the institutions and among individuals who produce them. Principles such as "examine power," "make labor visible," and "consider context" are applied to our project in two ways: in our analysis of the BLS dataset, and in the way we create our website. Data Feminism is the most explicit theoretical framework in our project in the sense that it makes explicit our discussion of workers' exclusion from official statistics, and the fact that AI's projected costs fall primarily on women and workers of color.
Our secondary dataset and theoretical source is a peer-reviewed paper in Science. According to the paper, the concept of "exposure" should be distinguished from "displacement": high exposure to AI technologies does not necessarily mean job loss, but rather suggests augmentation potential. This distinction is important for our research question and finding regarding the non-linear relationship between exposure and employment projection depending on the occupation. Our engagement with Eloundou et al.'s paper is critical: we note that exposure scores were partly generated by GPT-4 itself, and pose the Data Feminist question of whether AI system is able to evaluate its own economic impact objectively.
Boyd Davis, Vane, and Kräutli investigate the conditions under which humanities visualizations can and cannot be trusted, and argue that the ethical goal of visualization is not the maximal user trust but the disclosure of the extent to which data can be untrustworthy. This framework is used by us directly in our design approach, since our charts distinguish projections from the confirmed facts, and Data and Methods section documents all our cleaning decisions. In our project, Boyd Davis et al. become the warrant for the commitment to transparency as the key element of our project, and for the explanation of the process of data cleaning and rendering as "significant transformations."
In this article, Hepworth and Church investigate the biases that are built into data visualizations in digital humanities projects and argue that visualizations are rhetorical artifacts carrying biases in them. This theoretical reading is relevant for our project because our key claims depend on our visualizations, and this framework compels us to treat each chart as a constructed rhetorical argument rather than a transparent visualization of the data. In our project, Hepworth and Church inform Data and Methods section directly: we justify our decision to document every cleaning and design decision using the theory of visualization as rhetoric. They inform our design decisions as well: for example, the idea of visualization ethics and the omission of non-dominant groups in source data allows us to treat the BLS dataset's exclusion of gig and informal workers as one of the findings. It helped us to make concrete design decisions: distinction between projection and confirmed outcome, and labeling of data limitations.
This is our primary dataset. Employment Projections report provided by BLS gives ten-year projections of employment growth and decline for more than 800 occupations based on Standard Occupational Classification codes. This is the official and authoritative source of labor market forecasting in the US. Our project uses it as the source of the data on employment trajectory that we visualize and analyze. We also analyze BLS dataset's exclusion of gig and informal workers as the indicator of power relations and visibility of workers in the labor market. The agency's methodological FAQ further documents that the projections treat AI like any other technology and assume historically typical rates of change; we cite it in the analysis when weighing how far these forecasts can anticipate generative AI.