Why AI Exposure and Employment Projections Can Rise Together

Why are exposure and employment projections not contradictory?

Exposure and employment projections answer different questions, so they can point in different directions without conflict. Exposure asks whether AI capabilities could assist with or complete some tasks associated with an occupation. It does not show that employers have adopted those tools, that a particular worker uses them, or that jobs will disappear. In Artificial Intelligence (AI) Exposure Categories, the U.S. Bureau of Labor Statistics (BLS) explicitly says its categories are not employment forecasts, adoption probabilities, or worker replacement estimates. A projection instead estimates the occupation’s future headcount after demand, productivity, technology, and other forces are considered together. The BLS Web Developers and Digital Designers outlook makes the distinction concrete: it classifies web developers as very highly exposed relative to other occupations, yet projects U.S. web-developer employment to rise from 87,400 in 2025 to 90,600 in 2035, a rounded 4% increase. The same outlook says expanding e-commerce may support demand while improving tools and AI may soften growth. That is a net projection with countervailing pressures, not evidence that AI caused growth or that individual developers are secure. Read exposure as a prompt to inspect tasks; read projected growth as conditional market context. Neither settles a worker’s outcome.

Sources: Artificial Intelligence (AI) exposure categories — U.S. Bureau of Labor Statistics; Web Developers and Digital Designers — Occupational Outlook Handbook

What does an AI exposure category actually measure?

A BLS AI exposure category is a relative comparison among occupations, not a measure of how much of one worker’s week an AI system can handle. BLS describes theoretical exposure as the possibility that AI could assist with or complete some work. Its current-evidence measures use observed AI interactions mapped to occupational tasks, but do not directly establish that workers in the occupation used AI on the job. These signals do not show employer adoption or staffing changes. (U.S. Bureau of Labor Statistics, “Artificial Intelligence (AI) Exposure Categories.”) BLS combines five external sources. Three are theoretical: one relates AI capabilities to occupational abilities, while two assess whether language models could reduce time spent on occupational tasks. Two are observed-use measures, mapping Claude activity and Microsoft Copilot activity to occupational tasks or work activities. These inputs do not measure an identical concept. They assess different things: potential capability or product-specific activity. Their combination is not a direct audit of every workplace or a count of each employee’s exposed hours. BLS maps results to its occupational classification, based on the 2018 Standard Occupational Classification system, then ranks occupations within each source. Converting different raw scores to percentile ranks from zero to one makes them comparable on a relative scale: each rank indicates how an occupation compares with others covered by that source. BLS takes a median across the three theoretical ranks and another across the two observed-evidence ranks. A clustering algorithm groups the resulting dimensions into low, moderate, high, and very high categories. “Very high” therefore means relatively more exposure than other occupations in the classification, not an absolute threshold or a stated percentage of tasks that can be automated. Some sources do not cover every occupation. BLS imputes missing ranks using other available source data and the occupation’s two-digit occupational group. Of 4,155 possible occupation-source combinations, 211 were imputed, affecting 75 occupations. BLS says the categories depend on the occupation universe, source choices, crosswalks, imputation, normalization, and clustering. The theoretical inputs describe capabilities available no later than mid-2023; observed sources may lean toward early adopters. It is a scoped comparison, not a timeless measure of workplace practice. For a worker, a key limit is that the category does not distinguish automation from augmentation. A tool may assist a task, complete part of it, or change the review work around it; the category cannot say which design an employer will choose. BLS explicitly says exposure is not an employment forecast, adoption probability, productivity or wage forecast, or worker-replacement estimate. Use the label to identify recurring duties worth inspecting, then check what is assisted or completed, how much review is needed, and whether responsibilities or output expectations have changed. Exposure can direct attention; it cannot calculate personal odds of losing a job.

Sources: Artificial Intelligence (AI) exposure categories — U.S. Bureau of Labor Statistics

What does the current web-developer example show?

BLS currently places web developers in its very-high relative AI-exposure category and projects U.S. employment to grow 4% between 2025 and 2035, from 87,400 to 90,600 jobs. Both statements can be true because they describe different things. The exposure category compares occupations by how closely some tasks match AI capabilities and observed AI activity; the projection estimates a future count of workers in one occupation. BLS says its exposure category is not an employment forecast and high relative exposure does not necessarily mean employment will decline. The paired figures illustrate coexistence, not a contradiction or a causal finding. The figures come from separate BLS products. Its exposure categories combine five external measures, three theoretical and two based on observed AI activity mapped to occupational tasks or work activities. BLS converts source scores to relative percentile ranks and groups occupations into four categories. “Very high” means that, compared with other occupations, a larger fraction of tasks could be assisted or completed by AI and models have been observed performing some of them. Those interactions do not show that workers in the occupation used AI on the job. It does not distinguish automation from assistance. It is a relative task-applicability signal, not a count of affected hours or an adoption forecast. The employment estimate concerns the occupation’s net headcount across the United States, not the number of tasks AI may change. BLS’s Occupational Outlook Handbook reports about 87,400 web-developer jobs in 2025 and projects 90,600 in 2035, a rounded 4% increase. Its explanation names forces pointing in opposite directions: expanding e-commerce may support demand, while improving tools and increasing AI use may raise productivity and soften employment growth. “May” matters. BLS does not quantify each force’s contribution, say AI creates the projected jobs, or attribute the positive balance to AI. The projected increase reflects broader assumptions and influences, with AI described as a possible brake. BLS lists web-development duties including writing and testing code, meeting with clients or management, choosing applications for site needs, monitoring traffic, and checking performance. AI assistance could affect some repeatable digital steps, while demand for websites and decisions about what a site must do may move differently. This helps explain how opposing forces might coexist; it is not a measured breakdown of tasks or proof that judgment-heavy work is protected. The practical reading is narrow: high relative exposure and positive projected headcount are compatible. Growth does not establish that a particular developer’s role is secure, and exposure does not mean it will disappear. Neither national measure describes a specific employer’s adoption, staffing plans, or local hiring. A worker’s next question is which recurring tasks are changing and whether the workplace is changing output expectations, review responsibilities, or staffing. These figures prompt that investigation; they do not settle its result.

Sources: Artificial Intelligence (AI) exposure categories — U.S. Bureau of Labor Statistics; Web Developers and Digital Designers — Occupational Outlook Handbook

How can demand and productivity push in opposite directions?

A productivity gain changes the labor needed for a given amount of work; it does not tell us how much work customers will want. If an AI-assisted workflow lets a developer complete a basic site task in fewer hours, an employer might reduce hours, take on more projects with the same staff, lower prices, or redirect effort toward work that still needs judgment and checking. Each choice has a different employment consequence. An occupation-wide projection reflects the net direction after such responses and other forces combine, rather than counting only tasks that could be automated. The OECD Employment Outlook 2023 describes several channels that can operate together. Automation can displace workers from tasks a system performs instead. AI can complement people on other tasks, allowing them to do the same work more efficiently. Cost savings may induce demand for a product or service, increasing labor demand in work that remains or in related jobs. New tasks can also arise around using, maintaining, or improving technology. The OECD emphasizes that the net labor effect depends on which channels dominate. Savings create a productivity-induced demand effect only when they lead to greater demand for goods or services; efficiency is not itself proof of additional hiring. BLS uses this balance in its current outlook for web developers and digital designers. It says employment is projected to grow as e-commerce expands and businesses seek to compete online. In the same discussion, BLS says improved web-development tools and greater AI use may soften that growth: workers may become more productive, and people in other occupations may take on some basic web-development tasks. These are countervailing forces in one occupation-level outlook, not a finding that AI caused growth. BLS does not quantify how much projected demand comes from e-commerce or how much AI may reduce labor needs. Consider the mechanism, not a forecast about a particular firm: if basic page construction takes less effort, a business could commission more frequent updates or additional digital work because costs have fallen. That demand might create work for developers, or the same volume might be completed by fewer workers. More complex requirements, integration, accessibility, security, and checking could also change the work mix, but BLS does not measure those mechanisms separately. The key condition is whether new demand and complementary work are large enough to offset labor saved on automated tasks. If not, output could rise while employment grows more slowly, stays flat, or falls. Productivity is a change in capacity, not a synonym for job growth. A firm can use extra capacity to produce more, shorten delivery time, handle a backlog, reduce workload, or reduce staffing. The response depends on demand, costs, competition, and work organization; neither the OECD framework nor BLS's national projection identifies an individual employer's choice. For a worker, ask: are time savings producing more projects and different responsibilities, or mainly reducing the hours and people assigned to the same output? The projection records a conditional net balance; local evidence about workload, output, staffing, and task allocation helps show which side is appearing nearby.

Sources: Artificial intelligence and jobs: No signs of slowing labour demand (yet) — OECD Employment Outlook 2023; Web Developers and Digital Designers — Occupational Outlook Handbook

Why can one occupation contain exposed and less-exposed work?

An occupation is a bundle of tasks, and an exposure label summarizes that bundle rather than describing every worker’s day. Within one role, some activities may be more technically applicable to generative AI than others; a single average can therefore conceal pressure concentrated in particular duties. The International Labour Organization’s *Generative AI and Jobs: A 2025 Update* makes that variation part of its method: it assesses nearly 30,000 tasks, then classifies occupations using both mean task exposure and the variability of task scores. Two occupations with comparable averages can differ in whether exposure is spread broadly across tasks or concentrated in only some of them. The index models potential exposure, not adoption or redundancy. Web development illustrates the mix without requiring a claim about a fixed share of automatable work. The U.S. Bureau of Labor Statistics describes developers as creating and maintaining websites, handling technical performance and capacity, and sometimes creating site content. Its duties also include meeting clients or managers about needs, testing applications and navigation, coordinating with teammates over information and layout, monitoring traffic, and solving problems. Producing a first draft of code or content is different from deciding whether a site meets a client’s requirements, works under expected traffic, remains usable across devices, and behaves correctly after a change. This occupational description does not measure any developer’s task shares or employer AI use. Judgment, communication, and accountability may remain necessary, but tools can alter the information workers use, the speed expected, the number of projects assigned, or the amount of checking required. A developer who uses an AI-generated component still needs to establish whether it fits the design, integrates safely with the site, and passes relevant tests. The ILO’s 2025 brief says most jobs are more likely to be transformed than made redundant because human input remains necessary in many tasks. This is a global modeled conclusion, not a guarantee for a particular job. For a worker, the useful next step is to inventory recurring duties rather than map a broad job title onto an average. Record what is drafted, checked, explained, coordinated, and ultimately signed off; then note where AI is already used, where its output needs substantial correction, and who bears responsibility if it fails. The map cannot predict employment, but can focus attention on changing review procedures, output expectations, or staffing decisions. Use the occupational average as a prompt to investigate your own task bundle.

Sources: Generative AI and Jobs: A 2025 Update; How might generative AI impact different occupations?; Web Developers and Digital Designers — Occupational Outlook Handbook

What assumptions and time horizons bound a projection?

A ten-year employment projection is a conditional national estimate, not a precise prediction of what a particular employer will do or a clean measure of AI’s causal effect. The Employment Projections: 2025–2035 Technical Note says BLS builds a potential scenario for structural economic change under specific assumptions. It does not try to anticipate future business-cycle activity. If the assumptions are not realized, actual employment will differ. That boundary matters when a worker reads a positive occupation outlook beside an exposure measure: the projection describes one modeled balance over time, not a promise that each task, workplace, or region will expand. BLS links its occupational estimates to projections for the labor force, the macroeconomy, and industry output and employment. The technical note assumes full employment in the projected year and treats productivity and technological progress as continuing in line with historical experience. In effect, it extends relationships observed in historical data into the projection period. This makes the result a structured scenario grounded in past patterns, but also means that an unusually fast change can weaken its fit. BLS explicitly says that if technology advances much faster than it has historically, those relationships may not hold and the methods may not yield reasonable results. The note addresses AI directly. Faster productivity growth could raise output; if that increase were uniform across industries, BLS says it would not change projected employment under its full-employment framework. If productivity effects differed by industry, they could affect employment estimates, but BLS says it lacks data to base such differential impacts on. For the 2025–2035 projections, it therefore uses technological progress in line with historical patterns rather than adding speculative adjustments. BLS considers research on new technologies and adjusts conservatively when evidence supports a change. It describes AI’s potential labor-market effects as highly uncertain and says the precise impact ten years ahead cannot be predicted. Time horizon matters. A decade-long structural projection is not a near-term hiring count, employer adoption schedule, or local vacancy forecast. BLS releases projections annually to incorporate new data, research, and analysis, so a future vintage may differ as evidence and assumptions change. The occupational outlook page gives national context for web developers; it also links to state and area data, a more relevant starting point for checking regional conditions. A worker weighing a move can treat the national projection as one scenario anchor, then inspect current local vacancies, the skills they request, and concrete workplace changes such as altered review duties, output targets, or staffing. These checks cannot make the projection certain, but they tie a decision to local conditions and constraints.

Sources: Employment Projections: 2025–2035 Technical Note; Web Developers and Digital Designers — Occupational Outlook Handbook

Why are projected openings different from projected growth?

Projected growth and projected openings answer different questions. Growth is the expected net change in the number employed in an occupation between two dates. Openings estimate positions that may need filling, including jobs left by workers who transfer to another occupation or leave the labor force. Replacement hiring can continue even when net employment changes little; it does not mean the occupation is adding an equal number of new jobs. A vacancy count is not a direct measure of expansion or a sign that a particular employer is hiring.

The current U.S. Bureau of Labor Statistics Occupational Outlook Handbook makes the distinction visible in its web-work figures. It projects about 13,600 openings each year, on average, for the combined group of web developers and digital designers from 2025 to 2035. BLS says many openings are expected because workers transfer to other occupations or exit the labor force, including through retirement. Separately, the detailed projection estimates web-developer employment will rise from 87,400 in 2025 to 90,600 in 2035, a rounded 4 percent increase. The combined group is projected to grow 5 percent overall. Those figures have different denominators: the annual opening estimate covers both developers and digital designers, while the 4 percent figure is for web developers alone. They should not be presented as competing estimates or treated as interchangeable.

A projected opening is not automatically a newly created role caused by demand for more web work. Some openings replace departures; net growth captures the expected balance after additions and reductions in employment. Neither count isolates AI’s contribution. In the same outlook, BLS says e-commerce expansion may support demand, while improving tools and increased AI use may soften growth by raising productivity and enabling some workers in other occupations to handle basic web-development tasks. These forces may pull in different directions; BLS does not assign each a measured share of openings or growth.

For someone considering entry or a move, openings provide context about expected hiring needs, but cannot establish local competition, the mix of replacement and expansion roles, or the tasks a new hire will perform. Read the estimate alongside net growth, then check regional projections and current vacancies for the location and specialty that matter to you. Treat task changes as a separate question: hiring may continue while routine work, review duties, or skill mix shifts. These national occupation-level figures promise neither a job nor evidence that AI created openings or will eliminate them.

Sources: Web Developers and Digital Designers — Occupational Outlook Handbook

What realized evidence can and cannot settle yet?

Observed productivity or AI use can show that a particular task changed in a particular setting. It cannot, by itself, settle whether employment across a whole occupation will rise or fall over a decade. The gap matters because experiments, workplace records, exposure measures, and national projections observe different things: task performance, actual use, possible task applicability, and net headcount. They are evidence to connect, not interchangeable verdicts. The International Labour Organization review, “AI and jobs. A review of theory, estimates, and evidence,” brings together randomized controlled trials, field experiments, digital traces, and partial survey evidence. Its abstract reports productivity gains of roughly 20 to 60 percent in controlled trials and 15 to 30 percent in field experiments, while emphasizing that results depend on context. It also reports that novice workers tend to gain more from language models on simple tasks, but findings on complex tasks are mixed. Those figures describe performance in reviewed settings; they do not mean employers can reduce staff by the same proportion, or that the saved time becomes more output, shorter hours, or new hiring. The review identifies limits that matter here: experiments often use simple tasks, cover a limited variety of models, and do not establish whether exposure becomes adoption, substitution, productivity gains, or changes in expertise. The same review records signals that complicate a purely optimistic reading. Digital traces show substitution between people and machines in writing and translation, alongside rising demand for AI; it describes mild evidence of lower demand for novice workers, with stronger declines in some recent survey, platform-payment, or administrative-data studies. These are not a single occupation-wide employment estimate. Different data sources capture different tasks, worker groups, and periods, and the review notes gaps in coverage. Still, the findings show why positive productivity results cannot be treated as proof that every worker benefits: gains can coexist with reduced demand for some kinds of work or experience. A separate counterpoint comes from the OECD working paper “Artificial intelligence and the changing demand for skills in the labour market.” Its analysis of vacancies reports that the share in highly AI-exposed occupations asking for at least one emotional, cognitive, or digital skill rose by eight percentage points over time. But an establishment panel in the same report finds evidence that demand for those skills is beginning to fall. The two results are not necessarily inconsistent: vacancy composition over time and changes within establishments are different measures, and both concern skill demand rather than total employment. The paper therefore supports a narrower conclusion: exposed occupations can change internally, and the direction may differ by skill and measurement window. It does not provide a current U.S. web-developer headcount forecast. Read these studies alongside the Bureau of Labor Statistics’ warning that its relative exposure categories do not predict employment, productivity, adoption, or replacement. Exposure flags plausible task pressure; a projection combines broader assumptions into a conditional headcount outlook. Realized evidence makes the mechanisms less abstract, but does not collapse these measures into personal odds. The defensible verdict is coexistence with uncertainty: take task change seriously, while seeking employer and local-market evidence before deciding what move fits your constraints.

Sources: AI and jobs. A review of theory, estimates, and evidence; Artificial intelligence and the changing demand for skills in the labour market; Artificial Intelligence (AI) exposure categories — U.S. Bureau of Labor Statistics

How should a worker turn the two signals into a next move?

Use exposure to decide what to inspect, and employment projections as broad context. Neither tells you what your employer has adopted or your position. The U.S. Bureau of Labor Statistics explains in “Artificial Intelligence (AI) exposure categories” that its categories do not estimate adoption, productivity, job loss, or worker replacement. Its “Web Developers and Digital Designers” outlook is a national projection, not an assessment of any developer’s workplace. Ask which duties are changing and what response fits your constraints?

Start with an inventory of recurring work: drafting, coding, research, client clarification, testing, approvals, or maintenance. Mark where AI might assist or complete a step, then note what still needs judgment, verification, context, or accountability. This map is not a claim that BLS measures your task hours. BLS says its exposure classification compares occupations and does not directly observe whether workers used AI on the job. A high relative label is a reason to investigate, not to assign yourself a risk score.

Next, look for evidence of workflow change at work. Has a tool been approved for a recurring task? Have turnaround expectations, volume, staffing, review, or responsibility for errors changed? If a tool is in use, compare one bounded task before and after, including time spent checking, correcting, and handling exceptions. Faster first drafts alone do not establish better output, reduced labor needs, or increased demand. No current workflow change is evidence about adoption today, not a guarantee about tomorrow.

Then check the market you could reach. Review current vacancies in your region , noting repeated duties, experience, and credentials. Compare those signals with your employer’s direction and the national BLS outlook. BLS projects growth for web developers while saying improved tools and AI may soften it; context, not a local forecast. Projected growth does not prove a specific opening exists or that your specialty will remain unchanged.

Compare three scales of response. An in-role upgrade might mean learning to evaluate AI-assisted work or taking responsibility for workflow quality. An adjacent move could build on your domain knowledge while shifting toward different duties. A larger change requires more retraining, time, and financial flexibility. For each path, record learning cost and time, salary floor, location limits, health needs, and family responsibilities. Prefer the smallest move that addresses observed change ; revisit it if workplace evidence or local demand shifts.

A focused conversation can improve the evidence: ask which task is expected to change, how quality will be checked, and who remains accountable when a tool is wrong. The free task-level checker at /ai-job-risk-checker can help organize your task inventory and suggest first actions. Its change-pressure signals are not a validated probability of displacement. Treat them as a prompt for investigation, then decide from your actual tasks, workplace practice, reachable opportunities, and constraints.

Sources: Artificial Intelligence (AI) exposure categories — U.S. Bureau of Labor Statistics; Web Developers and Digital Designers — Occupational Outlook Handbook

What is the practical conclusion?

High AI exposure and projected employment growth can both be true because they describe different things. The U.S. Bureau of Labor Statistics’ AI exposure categories compare occupations on potential task applicability; they do not estimate adoption or job loss. Its Occupational Outlook Handbook projects web-developer employment to rise 4% from 2025 to 2035. The same entry says expanding e-commerce may support demand while improving tools, including AI, may soften growth. This shows coexistence, not that AI creates jobs or protects developers. A national projection cannot tell you what will happen at your employer or in your local market.

Look at your work: which recurring duties are changing, whether your employer has changed workflow, staffing or review duties, and what local vacancies require. If they point to change, compare an in-role upgrade, adjacent move or larger transition with your salary needs, location, learning time, cost, health and family responsibilities. Choose a bounded step; neither signal is a personal verdict.

The free task checker can organize your task review. Its change-pressure signals are not a validated probability of displacement. If you need to compare constrained paths, the optional career roadmap supports scenario planning, but cannot guarantee employment or income. Revise your interpretation when credible workplace or local-demand evidence changes, not simply because an occupation has high relative exposure.

Sources: Artificial Intelligence (AI) exposure categories — U.S. Bureau of Labor Statistics; Web Developers and Digital Designers — Occupational Outlook Handbook

Questions readers ask