What AI is actually doing to jobs, and what to do about it

Exposure and displacement are two different measurements

Exposure indexes work the same way in every study. Researchers break an occupation into tasks, score each task for how much of it a model could perform, and roll the scores back up. What comes out is task overlap. The International Labour Organization, whose index produced the widely quoted one-in-four figure, writes in the paper itself that exposure does not imply the immediate automation of an entire occupation, only the potential for a large share of its current tasks to be performed using the technology, and that whether this leads to an occupation disappearing depends on adoption decisions and on how workers and firms respond.

That sentence is usually cut from the coverage. Once you keep it, the numbers stop being frightening and start being useful. An exposure score tells you which parts of your day a model can already do a version of. It does not tell you whether your employer has bought the tool, integrated it, retrained anyone, or changed a single job description.

There is a second measurement that matters more and gets quoted less. The IMF pairs exposure with complementarity: how far an occupation is shielded by social, ethical, physical or skill factors. A judge scores high on exposure and high on shielding. A clerical worker scores high on exposure and low on shielding. Two occupations, the same exposure score, completely different risk. Any article that gives you one number without the second one has thrown away the half that predicts what happens to you.

What the exposure numbers actually say

The ILO and NASK working paper published in May 2025 found that one in four workers globally are in an occupation with some generative AI exposure, and that 3.3 percent of global employment sits in the highest exposure gradient. The spread by country income is wide: 11 percent of total employment in low-income countries against 34 percent in high-income countries. Within the top gradient the gender gap widens with national income, reaching 9.6 percent of female employment against 3.5 percent of male employment in high-income countries. Clerical work dominates that top gradient. Of the 13 occupations the study places there, most are clerical: data entry clerks, typists and word processing operators, accounting and bookkeeping clerks, general office clerks.

The IMF measures a different thing and gets a bigger headline number. Roughly 40 percent of global employment is in high-exposure occupations, 60 percent in advanced economies, 40 percent in emerging markets, 26 percent in low-income countries. In advanced economies the IMF splits that into 27 percent of employment in high-complementarity jobs and 33 percent in low-complementarity jobs. The low-complementarity third is the group with real displacement risk. The high-complementarity share is the group where the tool makes the person more productive.

The OECD, writing in 2023 before generative AI adoption had gone far, put 27 percent of jobs in occupations at high risk of automation when all automation technologies are counted, not only AI. That number predates ChatGPT-era tooling and is a reminder that automation anxiety has a long, imprecise measurement history.

The researchers revised their own estimates downwards

This is the detail that should change how you read any AI exposure headline. When the ILO team updated its 2023 index in 2025, the mean scores for clerical occupations went down, not up. The paper explains why: some 2023 task scores reached 0.9, which reflected an overly optimistic assessment of full automation potential, and two years of actually using the tools showed that tasks like taking meeting notes or scheduling appointments still require substantial human effort.

So the direction of travel is not uniformly worse. Some occupational scores rose sharply over the same two years, particularly web and media developers, statistical and database specialists, and software-related occupations, because models gained multimodal and agentic abilities. Others fell as researchers learned what the tools actually do at work rather than in a demo.

Read that pattern against the way this subject is written about. The occupations that got scarier are the highly digitised ones where the work already lived inside software. The occupations where the fear was loudest in 2023, the office and admin jobs, got a downward revision from the people who produced the original number. Neither of those facts made the news cycle, because a downward revision is not a story.

If you are in a clerical role, this does not mean you are safe. Clerical work is still the most exposed group, and the exposure sits in tasks you probably do every day. It means the timeline is longer and messier than the coverage implies, and that leaves room to act.

Where the labour market has actually moved

Measured effects, not projections. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab used ADP payroll records covering millions of US workers through June 2026. Their revised August 2026 paper reports no evidence of widespread, economy-wide job displacement. What it does report is a specific gap: employment of workers aged 22 to 25 in AI-exposed occupations now stands 19 percent below where it would be had it kept pace with their less-exposed peers, and experienced workers show no comparable gap. The mechanism is reduced hiring rather than increased separations, and the declines concentrate in occupations where AI usage substitutes for human tasks. Where the usage complements workers, employment is flat or rising.

The authors call these early descriptive indicators rather than causal estimates, and they note the pattern weakens when you control for education. Take it as what it is: the clearest signal available, pointing at the front door of the labour market rather than at the people already inside.

On the demand side, Indeed's AI tracker put AI-related job postings at 5.9 percent of all US postings as of June 2026, well past the previous peak of 3.3 percent in 2022. That is real growth and still a small slice of the market. If you are told to reskill because AI jobs are everywhere, the number to hold onto is the other side of that same figure, the roughly 94 percent of postings that do not mention AI at all.

The occupations the projections single out

The US Bureau of Labor Statistics published its 2025 to 2035 employment projections in August 2026. They are worth reading precisely because BLS does not attribute declines to AI in a headline. It publishes occupation-level numbers you can check.

Computer programmers are projected to decline 7 percent between 2025 and 2035, from 110,800 jobs, with about 4,400 openings a year, all of which BLS expects to come from workers transferring out or retiring rather than from growth. Computer support specialists are projected to decline 3 percent over the same period, from 903,100 jobs, though the network support sub-occupation within that group is projected to grow 1 percent. That matters if anyone has told you the help desk is your way into technology.

Against that, data scientists are projected to grow 35 percent, from 275,600 to 371,000 jobs, with a 2025 median wage of $120,230. Software developers, quality assurance analysts and testers are projected to grow 10 percent, from 1,905,400 to 2,090,800.

Those two sets of numbers sit inside the same technology sector and point in opposite directions. The declining roles are the ones where the work is well specified and repeatable. The growing ones require judgement about what to build and whether the output is right. That is the same split the Stanford payroll data found between substitution and complementarity, arrived at from a completely different direction.

Almost half of a typical job is exposed and almost none of it is replaceable end to end

Indeed Hiring Lab scored 2,884 work skills found in US job postings for how much generative AI could change how they are performed, mapping the results onto more than 53.5 million postings. The distribution: 40 percent of skills fall into minimal transformation, 19 percent into assisted, 40 percent into hybrid, and 1 percent into full transformation. Weighted onto actual postings the picture shifts slightly: in a typical US job posting, 46 percent of listed skills fall into hybrid or full transformation, 12 percent into assisted, and 42 percent into minimal. That 46 percent is where the almost-half framing in the heading comes from, and it is Indeed's number rather than ours. Hybrid means the model does the bulk of routine execution and a human manages exceptions, validates output and carries the legal or ethical responsibility.

One year earlier the same analysis found zero skills very likely to be fully replaced. In the 2025 run, 19 skills, or 0.7 percent of those analysed, crossed that line. The direction is real and the level is low.

The occupational spread is where this becomes personal. In a typical software development posting, 81 percent of listed skills are classified hybrid. In a typical nursing posting, 68 percent of listed skills are classified minimal transformation, because the core of the work is physical and interpersonal, and the exposure sits in documentation and communication at the edges.

Hold those two profiles side by side and the practical question stops being whether AI affects your job. It affects almost every job. The question is whether the exposed part of your work is the part you are paid for, or the part that gets in the way of the part you are paid for. For nurses it is mostly the latter. For a junior developer writing well-specified functions, it is the former.

What a foundational AI certification actually is

AWS Certified AI Practitioner is a foundational exam: 65 questions in 90 minutes for 100 USD. AWS describes the intended candidate as someone familiar with, but who does not necessarily build, solutions using AI and machine learning technologies, and lists the target roles as business analyst, IT support, marketing professional, product or project manager, line-of-business or IT manager, and sales professional. Microsoft's Azure AI Fundamentals sits in the same fundamentals tier, but check what its page says now before assuming the two exams are equivalent. Microsoft renumbered it from AI-900 to AI-901 and updated the English version on 15 April 2026. The current certification page says the candidate should have conceptual knowledge of AI solutions in Azure, the foundational technical skills to work with them, knowledge of Python coding syntax and programming techniques, and familiarity with Azure resources. Older summaries of this exam, written against the AI-900 wording, say that data science and software engineering experience are not required. That line is no longer on the page, and any guide still quoting it is describing a version of the exam that has been superseded.

Read those descriptions honestly. AWS is telling you plainly that the credential proves familiarity, and familiarity is a screening asset rather than a hiring reason. No employer creates a requisition for someone who is aware of machine learning concepts. Microsoft now asks for a little more than awareness, which makes AI-901 slightly more work and slightly less of a pure vocabulary badge. Neither difference turns either exam into a job.

Worth noting where the pay-premium claims come from. The AWS certification page cites a November 2023 AWS study for the figures that employers will pay 47 percent more for AI-skilled IT professionals and 43 percent more in sales and marketing. That is a vendor study on a vendor sales page for a vendor exam. It may well be directionally right. It is not independent evidence, and you should not spend money on the strength of it.

That leaves a narrow, real use. If you already hold a cloud or data job and a manager, a partner-tier requirement or an internal ladder wants the badge, 100 USD buys the badge. Our page on getting a certification to raise your pay at your current employer covers how to check that before you pay.

What to do in the next month if your work is in the exposed group

Start with a task audit, not a course. Write down what you did last week in half-day blocks. Mark each block as one of three things: work a model could already draft, work that needs your judgement or your relationships, or work that requires you to be physically present. That is the same method the ILO, IMF and Indeed all use, applied to one person. It takes an hour and it costs nothing.

If most of your week is in the first bucket, your exposure is genuine and your response is to move toward the second bucket inside your current employer, where you already have credibility, rather than to start over somewhere new. Ask to own the exception cases, the quality review, the client-facing part, the thing that goes wrong. The Stanford finding that experienced workers show no comparable employment gap is partly a statement about how hard people with context are to replace.

Use the tools before you buy a certificate for them. Every major model has a free tier. Spending twenty hours applying one to your actual work will teach you more about what it can and cannot do in your job than a 65-question exam on vocabulary, and it gives you something specific to say in an interview.

If you are out of work rather than worried at work, this is the wrong page. Our post on being unemployed for six months deals with the screening problem directly, and if you were laid off after a long tenure, the post on being laid off after 20 years covers the parts of that specific to you.

What this does not fix

This post does not predict whether your specific job will survive, and no source cited here does either. Exposure indexes measure task overlap, not employer decisions, and the ILO, IMF and Indeed all say so in their own methodology sections. The Stanford payroll finding is described by its authors as descriptive rather than causal, weakens when education is controlled for, and covers US ADP clients rather than the whole economy. BLS projections are projections. No number here shows that holding an AI certification changed anyone's employment outcome, because neither AWS nor Microsoft publishes an employment outcome for certificate holders, and the pay-premium figures on the AWS page come from an AWS study rather than an independent one. One claim in an earlier draft of this page did not survive checking: the line that Microsoft's Azure AI Fundamentals page says data science and software engineering experience are not required. That was true of the retired AI-900 wording and is not true of the AI-901 page as it stands today.

Where an exam fits

• AWS Certified AI Practitioner (AIF-C01), only if you already work in a cloud or data role and a manager, employer partner requirement or internal ladder has named the badge
• Azure AI Fundamentals (AI-901), only if your employer runs on Azure and someone there has actually named the badge, and only after you read the current AI-901 page, which now expects Python syntax and Azure familiarity rather than concepts alone
• Power BI Data Analyst (PL-300), only if you already build reports or handle data in your current job, since Microsoft expects PL-300 candidates to be proficient in Power Query and DAX before sitting it
• AWS Certified Data Engineer - Associate (DEA-C01), if you have hands-on pipeline or SQL work behind you and want the associate-level exam that assumes it

Common questions

Sources

1. 01Share of workers globally in an occupation with some generative AI exposure, share of global employment in the highest exposure gradient, and the spread between low-income and high-income countries One in four workers globally are in an occupation with some GenAI exposure; 3.3 percent of global employment falls in the highest exposure category; overall exposure is 11 percent of total employment in low-income countries against 34 percent in high-income countries
   International Labour Organization and NASK, ILO Working Paper 140, Generative AI and Jobs: A Refined Global Index of Occupational Exposure (Gmyrek and others, 20 May 2025) checked 2026-08-28
2. 02Gender split in the highest exposure gradient in high-income countries, the occupations in that gradient, and the ILO's own caveat on what exposure means 9.6 percent of female employment against 3.5 percent of male employment sits in Gradient 4 in high-income countries; of the 13 occupations in Gradient 4 the majority are clerical, including data entry clerks, typists and word processing operators, accounting and bookkeeping clerks and general office clerks; the paper states that such exposure does not imply the immediate automation of an entire occupation, and that 2023 task scores reaching 0.9 reflected an overly optimistic assessment of full automation potential
   International Labour Organization and NASK, ILO Working Paper 140, full working paper PDF hosted by co-publisher NASK checked 2026-08-28
3. 03Share of employment in high-exposure occupations by economy group, and the split between high-complementarity and low-complementarity jobs in advanced economies Almost 40 percent of global employment is exposed to AI; 60 percent in advanced economies, 40 percent in emerging markets, 26 percent in low-income countries. In advanced economies, 27 percent of employment is in high-complementarity jobs and 33 percent in low-complementarity jobs
   IMF, presentation by Florence Jaumotte on Staff Discussion Note SDN/2024/001, Gen-AI: Artificial Intelligence and the Future of Work (Cazzaniga and others) checked 2026-08-28
4. 04OECD estimate of jobs at high risk of automation counting all automation technologies, published before widespread generative AI adoption 27 percent of jobs are in occupations at high risk of automation when all automation technologies including AI are considered
   OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market checked 2026-08-28
5. 05Measured employment gap for young workers in AI-exposed occupations in US payroll data No evidence of widespread, economy-wide job displacement; employment of workers aged 22 to 25 in AI-exposed occupations stands 19 percent below where it would be had it kept pace with less-exposed peers, with no comparable gap for experienced workers; the effect operates primarily through reduced hiring rather than increased separations, using ADP payroll data covering millions of US workers through June 2026
   Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Stanford Digital Economy Lab, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, revised August 2026 checked 2026-08-28
6. 06Share of US job postings mentioning AI or generative AI terms 5.9 percent of postings as of June 2026, past the prior peak of 3.3 percent in 2022
   Indeed Hiring Lab, US Labor Market Snapshot, June 2026 checked 2026-08-28
7. 07Distribution of US work skills by degree of potential generative AI transformation, the distribution within a typical job posting, and the split between software development and nursing postings Of 2,884 skills assessed and mapped onto more than 53.5 million US job postings, 40 percent fall into minimal transformation, 19 percent assisted, 40 percent hybrid and 1 percent full transformation; 19 skills, or 0.7 percent, were rated very likely to be fully replaced, against zero a year earlier; in a typical job posting the split is 46 percent hybrid or full, 12 percent assisted and 42 percent minimal; 81 percent of skills in a typical software development posting are classified hybrid, while 68 percent of skills in a typical nursing posting are classified minimal transformation
   Indeed Hiring Lab, AI at Work Report 2025: How GenAI is Rewiring the DNA of Jobs checked 2026-08-28
8. 08Projected employment change for computer programmers in the United States Decline of 7 percent from 2025 to 2035, from 110,800 jobs, with about 4,400 openings projected each year, all expected to result from workers transferring out or leaving the labour force; 2025 median pay $100,390
   US Bureau of Labor Statistics, Occupational Outlook Handbook checked 2026-08-28
9. 09Projected employment change for computer support specialists in the United States Decline of 3 percent from 2025 to 2035, from 903,100 to 878,800 jobs; within the group, computer network support specialists are projected to grow 1 percent and computer user support specialists to decline 3 percent
   US Bureau of Labor Statistics, Occupational Outlook Handbook checked 2026-08-28
10. 10Projected employment change for data scientists in the United States Growth of 35 percent from 2025 to 2035, from 275,600 to 371,000 jobs; 2025 median pay $120,230
    US Bureau of Labor Statistics, Occupational Outlook Handbook checked 2026-08-28
11. 11Projected employment change for software developers, quality assurance analysts and testers in the United States Growth of 10 percent from 2025 to 2035, from 1,905,400 to 2,090,800 jobs; 2025 median pay $134,040
    US Bureau of Labor Statistics, Occupational Outlook Handbook checked 2026-08-28
12. 12AWS Certified AI Practitioner exam format, cost and intended candidate, and the origin of the pay-premium figures quoted on the same page Foundational category, 65 questions, 90 minutes, 100 USD; intended for individuals who are familiar with, but do not necessarily build, solutions using AI and ML technologies, with example roles including business analyst, IT support, marketing professional and sales professional; the page attributes its claim that employers will pay 47 percent more for AI-skilled IT professionals and 43 percent more in sales and marketing to a November 2023 AWS study
    Amazon Web Services checked 2026-08-28
13. 13What Microsoft currently says a candidate for Azure AI Fundamentals should already know, and the exam's current number The certification page states the candidate is at the beginning of a career in AI solution development and should have conceptual knowledge of AI solutions in Azure, the foundational technical skills to work with them, knowledge of Python coding syntax and programming techniques, and familiarity with Azure resources; the English version was updated on 15 April 2026 and the exam is numbered AI-901, with no retirement date. The page no longer carries the older AI-900 line that data science and software engineering experience are not required
    Microsoft Learn, Microsoft Certified: Azure AI Fundamentals checked 2026-08-28
14. 14What Microsoft says a PL-300 candidate should already be able to do The Power BI Data Analyst Associate page states that candidates should deliver actionable insights by working with available data and applying domain expertise, work with business stakeholders to identify requirements, and should be proficient at using Power Query and Data Analysis Expressions (DAX)
    Microsoft Learn, Microsoft Certified: Power BI Data Analyst Associate checked 2026-08-28

Related