Will AI take my job?

Type your occupation and see what every major study actually says about it — each one separately, labelled with what it measures and when. There is no single honest answer to “will AI take my job”, because the four best sources disagree sharply, and that disagreement is the most useful thing in this dataset. 245 occupations, 114,360,850 US workers, every figure traceable to a primary source.

Updated August 1, 2026 · Built by Alexander Gusev · How this dataset is built

245
occupations covered
114M
US workers represented
4
studies, shown separately
0
blended risk scores

Check your occupation

245 occupations covering 114,360,850 US workers. Informal titles work too — use the arrow keys and Enter to pick a result.

Popular:

Pick an occupation above to see what each study says about it, or read every occupation in the table below.

Every occupation in the dataset

Showing 245 of 245 occupations. Sort by any column.

AI exposure estimates by occupation, showing each study separately
Automation shareGDPval
41-20313,897,860$35,41092%36%32%41%not covered
35-30233,854,050$31,200not covered7%0%52%not covered
11-10213,503,020$105,77016%39%14%44%covered
29-11413,379,720$97,550not covered38%6%44%covered
41-20113,089,410$32,88097%36%8%33%not covered
53-70622,950,280$40,24085%6%0%90%not covered
53-70652,833,810$37,330not covered19%0%46%not covered
43-40512,595,750$44,77055%71%70%34%covered
43-90612,464,940$45,01096%50%45%69%not covered
35-30312,270,910$35,23094%22%0%60%not covered
37-20112,209,760$36,84066%3%0%not coverednot covered
53-30322,062,040$58,64079%17%0%65%not covered
43-60141,706,790$47,54096%57%45%71%not covered
15-12521,687,890$135,980not covered45%29%61%covered
49-90711,529,700$49,59064%8%0%76%not covered
13-20111,449,500$83,68094%52%35%45%covered
31-11311,448,910$42,260not covered14%0%38%not covered
43-10111,436,680$69,5001%53%19%57%covered
35-20141,409,890$37,39096%12%1%71%not covered
25-20211,388,390$63,9700%31%10%32%not covered
43-30311,373,680$50,67098%31%31%52%not covered
33-90321,283,470$38,02084%25%0%87%not covered
41-30911,256,010$69,990not covered57%not covered46%not covered
41-40121,238,190$72,08085%71%63%47%covered
35-10121,223,240$44,08063%26%6%42%not covered
41-10111,121,800$48,52028%41%26%45%covered
47-20611,096,780$47,12088%3%3%not coverednot covered
13-10821,066,670$102,320not covered40%not covered52%covered
25-20311,065,210$72,0401%34%29%42%not covered
53-3033983,300$44,86069%25%0%not coverednot covered
43-6013961,610$45,93081%61%36%72%covered
37-3011952,640$39,15095%2%0%69%not covered
13-1071912,430$75,940not covered59%40%42%not covered
43-4171910,180$38,01096%58%43%59%not covered
13-1161899,580$78,76061%58%65%55%not covered
13-1111898,280$101,86013%50%24%45%not covered
35-2021893,600$35,32087%5%0%79%not covered
37-2012860,670$35,51069%6%0%not coverednot covered
11-3031841,710$166,5707%43%39%50%covered
31-9092817,870$45,69030%14%5%28%not covered
43-5071816,870$45,26098%36%0%68%covered
47-1011812,210$79,92017%20%3%71%not covered
53-7051774,420$46,42093%3%0%not coverednot covered
47-2111757,220$63,19015%15%0%61%not covered
35-3011756,390$34,34077%14%0%47%not covered
23-1011754,500$159,6704%48%17%31%covered
15-1232717,190$61,860not covered67%47%66%not covered
49-3023704,640$50,62059%7%0%74%not covered
51-1011673,430$74,4502%23%0%36%covered
11-3021670,570$175,1404%56%16%65%covered
33-3051670,520$76,21010%23%12%59%not covered
47-2031670,090$60,58072%9%0%78%not covered
29-2061648,410$64,4006%25%0%48%not covered
35-2011641,070$30,89081%3%0%74%not covered
11-2022637,080$148,2701%48%4%48%covered
25-2022620,090$64,37017%34%30%37%not covered
49-1011617,500$79,8600%27%10%50%not covered
51-9061597,370$48,57098%18%3%33%not covered
11-9111597,080$123,8601%37%7%63%covered
53-7064559,820$36,28038%10%0%80%not covered
35-9011542,750$33,98091%0%0%75%not covered
25-3031524,770$41,670not covered8%not covered23%not covered
15-1211519,530$105,850not covered47%28%84%not covered
39-9011518,910$34,9808%33%1%51%not covered
41-3031489,570$78,6602%52%44%49%covered
41-3021479,100$62,28092%53%32%46%not covered
25-2011478,780$38,1401%31%0%40%not covered
47-2073478,090$59,85095%9%0%48%not covered
35-9021477,450$34,81077%0%0%76%not covered
29-2052471,680$45,75092%43%7%42%not covered
47-2152465,840$63,80035%6%1%70%not covered
43-6011459,910$76,59086%74%23%57%not covered
13-1151458,300$69,2801%59%28%48%not covered
35-2012441,050$37,45083%9%0%43%not covered
49-9041439,640$64,52067%14%2%82%not covered
21-1093437,860$45,93013%38%0%56%not covered
35-9031432,690$31,20097%33%7%32%not covered
13-1041417,070$80,7308%61%12%38%covered
51-4121416,210$53,75094%4%0%not coverednot covered
49-9021409,670$61,01065%14%2%73%not covered
53-3031409,180$38,77098%14%3%70%not covered
43-3021404,060$48,50096%42%19%45%not covered
53-3051402,930$47,920not covered18%0%not coverednot covered
41-2021400,810$41,30097%68%20%43%covered
39-3091397,830$32,15072%37%6%58%not covered
11-2021395,240$166,7901%58%32%38%not covered
23-2011392,880$62,89094%45%29%37%not covered
21-1021392,550$59,5503%30%1%40%covered
43-5061390,160$59,65088%52%9%42%not covered
31-9091387,790$48,07051%10%0%not coverednot covered
33-3012380,500$58,94060%24%0%98%not covered
53-7061380,430$35,83037%0%0%not coverednot covered
11-9021380,360$114,9907%40%12%51%not covered
51-9111379,060$43,22098%8%0%not coverednot covered
17-2051367,840$100,8402%38%1%48%not covered
17-2112365,740$102,4403%61%4%53%covered
13-2051361,980$102,74023%50%57%44%covered
21-1012353,310$64,3301%33%12%37%not covered
33-2011345,990$59,28017%9%0%not coverednot covered
39-9032331,490$36,5601%25%0%54%covered
43-3071329,480$43,03098%33%2%51%not covered
43-5052328,820$60,55068%35%0%52%not covered
11-9032328,330$105,8700%46%5%27%not covered
13-1031324,230$78,00098%49%8%41%not covered
29-1171323,040$132,300not covered32%9%38%covered
29-1051321,970$140,9101%40%9%37%covered
15-1244314,340$99,130not covered51%34%68%not covered
11-9141311,180$69,99081%43%17%44%covered
39-5012305,710$35,79011%5%3%44%not covered
17-2141296,810$104,1101%37%8%45%covered
35-3041293,900$35,36086%11%0%84%not covered
49-3031289,960$61,77073%0%0%77%not covered
51-4041287,050$58,75065%13%0%28%not covered
41-4011284,800$104,92025%57%27%35%covered
13-2072274,330$76,69098%65%19%61%not covered
41-2022270,070$38,63098%34%0%37%not covered
13-2052266,800$105,07058%67%35%42%covered
45-2092265,500$35,660not covered7%2%53%not covered
11-3012263,960$114,130not covered50%0%66%covered
15-2051262,440$120,230not covered59%46%58%not covered
41-1012214,390$87,5208%44%23%43%covered
11-1011204,350$213,9902%35%3%53%not covered
35-1011200,040$62,47010%26%0%40%not covered
27-1024197,830$62,9608%41%37%30%not covered
41-9022193,370$52,83086%41%28%47%covered
15-1212190,650$129,180not covered54%49%76%not covered
19-1042172,340$103,4100%57%4%23%not covered
25-2058163,930$74,260not covered30%10%41%not covered
29-1122162,450$100,3300%39%1%54%not covered
29-1071162,150$135,88014%13%0%51%not covered
51-9124158,740$48,250not covered1%0%97%not covered
33-9092157,550$33,58067%16%0%not coverednot covered
29-2053156,960$45,1304%34%0%59%not covered
11-3013156,180$106,660not covered27%13%55%not covered
33-1012154,610$106,0400%36%0%41%covered
49-3021149,310$54,89091%4%0%not coverednot covered
53-7081147,240$49,69093%23%0%96%not covered
47-4011146,720$74,69063%14%5%60%not covered
27-2012143,120$90,3602%48%9%67%covered
29-1126139,790$82,2807%19%0%51%not covered
35-2015138,650$35,88094%0%0%72%not covered
49-3011138,090$79,87071%6%0%67%not covered
47-2181135,490$55,44090%0%2%63%not covered
25-4022133,790$68,270not covered50%20%51%not covered
53-2031131,650$63,58035%15%9%72%not covered
29-2056129,140$47,3803%15%0%55%not covered
51-8031128,490$60,02061%13%0%not coverednot covered
43-9021127,080$41,34099%50%67%77%not covered
29-1021124,390$170,9500%23%3%37%not covered
47-2211119,770$61,80082%4%0%not coverednot covered
29-2055117,460$64,65034%2%0%not coverednot covered
33-3021114,430$93,79034%30%4%60%not covered
25-1194114,110$63,820not covered45%16%28%not covered
15-1255113,330$104,000not covered68%25%51%not covered
25-2032111,420$66,2701%33%0%34%not covered
11-9121108,690$167,2202%57%6%45%not covered
15-2031108,510$88,9404%63%43%53%not covered
29-1215107,510$244,180not covered38%0%42%not covered
17-1011106,770$99,2802%43%8%46%not covered
13-2053105,420$81,37099%50%6%55%not covered
53-2011103,560$232,14018%23%0%not coverednot covered
39-1014103,190$48,560not covered34%4%47%not covered
53-6031102,010$35,67083%7%0%73%not covered
31-901198,790$58,45054%19%0%38%not covered
17-302395,130$78,19084%33%2%53%not covered
15-125192,230$100,390not covered68%75%63%not covered
27-304191,690$77,9206%65%25%34%covered
29-203290,160$96,59035%29%0%62%not covered
19-204189,250$82,2203%73%5%18%not covered
49-209886,340$60,07082%18%3%19%not covered
11-306184,320$148,0803%47%20%36%not covered
29-113183,900$130,1004%23%9%59%not covered
19-203182,770$91,24010%40%26%69%not covered
25-101182,150$99,080not covered42%31%30%not covered
29-103177,570$76,4000%46%13%48%not covered
17-206176,660$161,74022%32%15%46%not covered
13-208276,480$54,92099%63%13%44%not covered
19-303375,990$100,580not covered51%6%36%not covered
43-415175,200$46,17098%60%7%62%covered
11-203274,850$146,910not covered57%23%32%not covered
19-409973,910$62,28061%not covered10%51%not covered
43-505173,720$62,13095%28%0%not coverednot covered
27-102571,500$67,1902%36%0%41%not covered
27-401170,230$58,10055%30%2%52%covered
15-125470,190$92,650not covered64%48%55%not covered
15-124269,990$104,620not covered60%33%63%not covered
19-402169,620$57,51030%27%6%59%not covered
47-222168,380$62,78083%0%5%not coverednot covered
17-201167,710$134,9602%52%8%52%not covered
29-121667,150$256,560not covered37%8%42%not covered
13-204164,390$83,51098%56%17%50%not covered
41-904158,430$35,45099%53%29%29%not covered
21-201157,200$60,8101%24%11%40%not covered
41-304155,110$50,16010%56%41%61%not covered
27-309152,060$60,17038%84%43%63%not covered
27-402151,760$44,6602%22%20%51%not covered
17-102250,830$75,44038%43%0%64%not covered
39-601249,240$38,95021%70%19%77%covered
23-209348,580$58,65099%55%2%59%not covered
27-304347,940$76,9104%81%25%29%not covered
53-201247,630$123,22055%22%0%not coverednot covered
13-203147,160$91,64094%64%7%34%not covered
53-403146,440$78,00083%21%0%not coverednot covered
41-902146,100$73,22097%42%26%58%covered
19-305144,230$89,32013%58%10%48%not covered
29-104142,790$136,57014%5%0%63%not covered
53-305441,050$42,100not covered25%0%59%not covered
27-302339,250$62,200not covered63%21%29%covered
17-208138,340$107,1102%50%4%48%not covered
53-502136,850$92,46027%21%0%not coverednot covered
27-204236,1807%13%0%41%not covered
43-401135,940$65,75098%80%17%60%not covered
33-902135,580$51,22031%39%0%73%covered
43-902235,010$49,28081%49%24%55%not covered
43-201134,280$38,63096%48%39%67%not covered
51-201134,020$65,38079%3%5%not coverednot covered
27-102133,490$83,9104%41%4%24%not covered
47-223131,350$53,140not covered11%0%98%not covered
15-204129,030$105,65022%71%21%48%not covered
15-201126,670$130,00021%54%5%34%not covered
51-902126,000$48,54097%8%4%86%not covered
27-403225,610$75,42031%47%22%46%covered
39-403125,100$55,010not covered9%0%47%not covered
23-102324,030$153,99040%25%31%51%not covered
53-202122,510$148,08011%31%0%49%not covered
21-202122,160$52,1003%46%15%40%not covered
27-403121,550$74,99060%18%17%39%not covered
11-201121,470$133,6604%55%17%33%not covered
27-301121,240$47,34010%54%6%36%not covered
27-101419,970$102,0302%48%36%54%not covered
19-301117,790$124,72043%46%24%27%not covered
29-112417,070$105,31034%17%0%68%not covered
19-101315,730$78,8502%63%5%56%not covered
39-501115,000$38,21080%7%0%48%not covered
29-112514,930$61,9600%36%0%69%not covered
45-201114,410$49,94094%29%0%49%not covered
29-118113,660$95,7800%46%0%36%not covered
27-101311,220$55,4904%31%36%47%not covered
49-90819,980$64,120not covered4%0%not coverednot covered
23-10229,210$75,5306%37%24%38%not covered
11-90136,500$89,9005%25%0%43%not covered
43-90814,580$51,12084%60%18%56%not covered
41-90123,780$48,47098%40%0%80%not covered
43-20213,430$41,74097%55%0%72%not covered
15-20212,030$126,7105%69%42%45%not covered

What each study actually measures

Four sources dominate every conversation about AI and jobs, and they are routinely quoted as if they were answering the same question. They are not. One is a pre-LLM model of computerisation. One rates tasks. One counts real behaviour. One grades finished work products. Reading a number without knowing which of the four it came from is how “47% of jobs at risk” ended up in a decade of headlines.

The Future of Employment · Frey & Osborne, Oxford Martin School · 2013

Modelled probability that an occupation could be computerised over one to two decades, based on nine O*NET bottleneck variables. Predates large language models entirely.

Scale: 0 to 1. Higher means more susceptible to pre-LLM computerisation. · Primary source

GPTs are GPTs · Eloundou, Manning, Mishkin & Rock (Science) · 2023-2024

Share of an occupation's O*NET tasks where an LLM, or software built on one, could cut the time to complete the task by at least half without lowering quality. Exposure, not automation.

Scale: 0 to 1 (the beta measure: fully exposed tasks count 1, tool-assisted tasks count 0.5). · Primary source

Anthropic Economic Index · Anthropic · 2026

Observed behaviour, not potential: millions of real Claude conversations mapped to O*NET tasks, then split into automation (the model does the task) and augmentation (the model helps a person do it).

Scale: Exposure runs 0 to 1; the automation and augmentation shares are percentages that sum to 100. · Primary source

GDPval · OpenAI · 2025

Whether frontier models can produce real occupational deliverables — briefs, models, decks, care plans — well enough that an experienced professional prefers them to a human expert's work in a blind comparison.

Scale: Coverage only. GDPval spans 44 occupations across the nine largest sectors of US GDP; it does not publish a per-occupation risk score. · Primary source

Coverage is uneven, and gaps are shown as gaps

Of the 245 occupations here, Frey & Osborne covers 207, Eloundou et al. covers 244, the Anthropic Economic Index publishes observed exposure for 242, and GDPval benchmarks 43. Where a study does not cover an occupation the tool prints not covered. It never substitutes a neighbouring occupation, a group average, or a plausible-looking guess.

The studies disagree — and the disagreement is the story

Across the 206 occupations where both can be compared, Frey & Osborne's 2013 probability of computerisation correlates with Eloundou et al.'s LLM task exposure at roughly −0.19. Not weakly positive. Negative. The jobs the 2013 model flagged as most at risk are, on average, slightly less exposed to language models than the ones it called safe.

Frey and Osborne 2013 compared with Eloundou et al. 2024 for ten occupationsA dumbbell chart. For manual and clerical occupations the 2013 computerisation probability is far higher than the 2024 language-model task exposure. For writing, legal, teaching and mathematical occupations the ordering reverses.0%25%50%75%100%TelemarketersBookkeeping clerksRestaurant cooksAccountants & auditorsWelders & cuttersParalegalsMathematiciansLawyersWriters & authorsElementary teachers
Frey & Osborne 2013, probability of computerisation Eloundou et al. 2024, share of tasks exposed to LLMs

The reason is mechanical. Frey & Osborne's classifier keyed on manual dexterity, cramped working positions, originality, negotiation and caring for others. Under that model, welding and cooking looked automatable and writing did not. Language models inverted exactly that ordering: they are excellent at drafting, summarising and restructuring text, and they cannot lift anything.

The same reversal shows up in pay. In this dataset the 2013 computerisation probability correlates with median annual wage at about −0.53 — low-paid work was the work at risk. Anthropic's observed AI exposure correlates with wage at about +0.13. The threat moved up the income distribution.

Where the modern measures do agree

Eloundou's theoretical task exposure and Anthropic's observed exposure correlate at about +0.64across 241 occupations. That is the useful signal: what a model could do and what people are doing with it broadly line up, which is a point in favour of both. Within the Eloundou study, human raters and the GPT-4 rater agree at about +0.86, so the residual disagreement is about method, not noise.

Where they part company is instructive. Truck drivers score 17% on Eloundou's task exposure and 0% on observed Anthropic usage — there is no keyboard in the cab. Computer programmers score 68% and 75% respectively, the highest observed exposure of any occupation in the dataset.

Exposure is not automation, and automation is not job loss

Every number in the tool sits somewhere on a chain, and no single study spans the whole chain. Confusing one link for another is what turns a research finding into a scary headline.

From task exposure to job lossA five-step funnel narrowing from task exposure, through demonstrated capability, observed adoption and automation, to job loss. Each step is a separate condition, and every study measures only one of them.Task exposureA model could helpDemonstrated capabilityA model does it wellObserved adoptionPeople actually use itAutomationNo human in the loopJob lossAn employer cuts the role
Widths are illustrative, not measured. The point is that each arrow is a separate condition, and no single study spans the whole chain.

As of Anthropic's sixth Economic Index report (June 2026), real Claude.ai conversations in May 2026 split 51.4% augmentation to 48.6% automation — a narrow majority still has a person in the loop, and the balance has been drifting back towards augmentation, not away from it. Anthropic's own labour-market analysis (March 2026) found no systematic rise in unemployment among highly exposed workers since late 2022, with tentative evidence that hiring of younger workers has slowed in exposed occupations. That is a real and worrying signal — but it is a signal about entry-level hiring, not about existing jobs disappearing.

For the specific question of what has already been announced and attributed to AI, see our AI job displacement tracker, which follows the actual reductions rather than the forecasts.

What the macro forecasts add — and what they cannot tell you

The occupation-level studies above are the ones that can say anything about your job. The famous headline numbers are economy-wide, and they are useful only as context.

Economy-wide AI and employment forecasts
SourceHeadlineWhat it is really saying
IMF, 2024~40% of global employment exposed; ~60% in advanced economiesExposure, split roughly half into complementarity and half into displacement risk
Goldman Sachs, 2023300 million full-time jobs exposed globallyTwo thirds of US and European jobs exposed to some automation; a quarter to half of that workload substitutable
McKinsey Global Institute, 2023Up to 30% of US hours worked automatable by 2030Hours, not headcount; generative AI moved the estimate from 21.5% to 29.5% and implies ~12M extra occupational switches
WEF Future of Jobs, 2025170M roles created, 92M displaced by 2030A net gain of 78M, with 39% of workers' existing skills expected to change
PwC AI Jobs Barometer, 202662% average wage premium for AI skillsBuilt from a billion job adverts across 27 markets; a story about pay and skill demand, not headcount

Notice what these have in common: none of them counts jobs that have gone. They count exposure, hours, adverts and scenarios. If you want the occupation-level version of the same question, our companion pages cover the jobs AI is most likely to replace and which jobs are safest from AI.

How to read your own result

High task exposure, low observed automation

The most common pattern for professional work in 2026. It means the tooling can already do a lot of what you do, and people are mostly using it as a collaborator. Practically: the tasks in your “most often done with AI” list are the ones where speed is about to stop being a differentiator, and the ones in your “least” list are where your leverage now sits.

Low task exposure, low observed automation

Typical of physical, hands-on and in-person work. Language models are not the risk here; robotics, demand and wages are. Frey & Osborne remains the more relevant historical reference for these occupations even though it is thirteen years old, which is why the tool still shows it.

High exposure with an automation-skewed split

Where the automation share sits well above the augmentation share, the work is already being delegated end to end rather than collaborated on. Treat that as the strongest in-dataset signal, and check the sample caveat: occupations with a tiny share of observed usage have small samples, so their splits swing.

What to actually do with this

  • Work at the task level, not the job level. No occupation in this dataset is uniformly exposed; the spread within a job is almost always wider than the gap between jobs.
  • Use the task lists as a to-do list. The high-penetration tasks are the ones to hand over first — that is where the measured time savings are.
  • Watch the augmentation share, not just the exposure score. Falling augmentation in your occupation is the earlier warning sign.
  • Discount any tool, including this one, that gives you a single percentage. If it does not tell you which study, which year and what it measured, it is guessing.

Method, sources and limitations

Every value in the tool was pulled from a primary source and stored under the study that reported it. Nothing is averaged, rescaled or interpolated across studies.

Where each field comes from

  • Employment and median wage: the BLS Occupational Employment and Wage Statistics May 2025 national estimates, released 15 May 2026, retrieved through the BLS public data API.
  • Frey & Osborne: the 702-occupation appendix of The Future of Employment (2013), keyed on 2010 SOC codes.
  • Eloundou et al.: the occupation-level beta exposure released alongside GPTs are GPTs, both the human-rater and GPT-4-rater columns.
  • Anthropic: observed exposure and task penetration from the labour-market-impacts release, and the automation/augmentation split and usage share from the May 2026 data in the Anthropic Economic Index, published 26 June 2026.
  • GDPval: coverage only, from the 44 occupations in OpenAI's benchmark. In its first round the best model matched or beat the human expert's deliverable on roughly 48% of tasks, but that is a benchmark-wide figure and is not shown per occupation here.
  • Task lists: O*NET 30.0 task statements joined to Anthropic's published per-task penetration scores.

Known limitations

SOC revisions are not bridged

Frey & Osborne used the 2010 SOC. Occupations renumbered or created since then show as not covered rather than being re-mapped. Registered Nurses, for example, appear in the 2013 paper under the old code 29-1111 at 0.9%, but this dataset uses the current code 29-1141 and leaves the field empty. Bridging the codes would require judgement calls we would rather not hide inside a number.

Observed usage is one vendor's traffic

The Anthropic figures describe conversations with Claude, not with all AI systems, and skew heavily towards technical and professional users. Occupations with very low usage have small samples, so their automation splits are indicative only.

US-only, and nothing here is a forecast

Employment and wage data are United States national estimates. Exposure scores describe capability and current behaviour; none of them predicts what any employer will do.

Why we built this

Planetary Labour builds an autonomous go-to-market engine — software that runs the marketing work itself, publishing content and building distribution without a team operating it. We are on the automation side of this question, which is precisely why we would rather publish the disagreement between studies than a comforting or alarming single number.

What our own data keeps showing is the pattern in the tool above: the tasks that go first are the repetitive, well-specified ones, and the work that remains is judgement, taste and accountability. That is the shape of the argument in our manifesto and across the rest of our future of work writing.

Frequently asked questions

Will AI take my job?

Nobody can give you a single honest probability, because the studies that claim to measure this are measuring different things. Frey & Osborne modelled pre-LLM computerisation risk in 2013. Eloundou et al. measure what share of your tasks a language model could speed up. The Anthropic Economic Index measures what people actually do with AI today. Use the tool above to see all of them for your occupation, and treat wide disagreement between them as the real answer: high task exposure with low observed automation is the most common pattern in 2026.

What is the difference between AI exposure and AI automation?

Exposure means a task could be done faster or better with AI. Automation means the AI does the task end to end, with no person in the loop. Job loss requires a third thing again: an employer choosing to convert that saved time into fewer roles rather than more output. Anthropic's May 2026 data splits real Claude.ai usage 48.6% automation to 51.4% augmentation, so a narrow majority of measured AI use is still a person working with a model rather than instead of one.

Is the famous 47% of jobs at risk figure still accurate?

That number comes from Frey & Osborne's 2013 Oxford paper and refers to the share of US employment in occupations they judged highly susceptible to computerisation over the following one to two decades. It predates large language models entirely, and it was built around manual dexterity, cramped workspaces and routine physical work. Across the 206 occupations where it can be compared with the LLM-era estimates in this dataset, the 2013 figure correlates at roughly minus 0.19 with LLM task exposure — the two measures point in nearly opposite directions.

Which jobs have the highest AI exposure right now?

In Anthropic's observed-usage data the highest exposure sits with computer programmers, customer service representatives, market research analysts and marketing specialists, sales representatives, and financial and investment analysts. These are desk jobs whose output is text, code or analysis. Occupations built on physical presence — chefs, machinists, diesel mechanics, wind turbine technicians — record almost no observed AI use at all.

How accurate are AI job risk calculators?

Most of them are not calculators at all. They take one study, usually Frey & Osborne 2013, and present its single number as a personal probability. That hides both the age of the data and the disagreement between sources. A defensible tool shows each study separately, labels what it measures and when, and says 'not covered' rather than inventing a figure. That is the standard this page holds itself to.

Continue reading