Where AI reaches the U.S. workforce
How the 2023-era study models potential AI task speedups for 825 U.S. occupations covering about 155 million workers. These results describe modeled exposure, not observed adoption or the probability of job loss.
- 33U.S. average modeled task exposure, out of 100
- 46%of covered workers are in high or very high exposure bands (72 million)
- Computers & mathhas the highest modeled exposure at 65; cleaning & grounds is lowest at 4
- +28points higher for the best-paid fifth of workers than the lowest-paid fifth
What does this mean for your career?
The scores describe tasks that researchers judged potentially faster with language-model assistance under assumptions from 2023. They do not show how much work is automated today or what fraction of workers will be displaced.
A high-exposure occupation can still be adding jobs. AI adoption, regulation, employer decisions, training and human oversight all affect what happens next. Compare the modeled signal with published pay and employment outlook before drawing conclusions.
Check your own occupation → · Compare career alternatives → · Explore transferable skills →
Modeled AI exposure by career field
Computers & math, Office & admin, Business & finance and Engineering have the highest modeled task exposure in the source data. Fields such as cleaning & grounds, construction and farming score lower. These are occupational averages, not predictions of job losses.
- Computers & math65100% of workers high or very high
- Office & admin5796% of workers high or very high
- Business & finance53100% of workers high or very high
- Engineering4896% of workers high or very high
- Science4790% of workers high or very high
- Legal4798% of workers high or very high
- Arts & media4669% of workers high or very high
- Management45100% of workers high or very high
- Sales4374% of workers high or very high
- Education3631% of workers high or very high
- Social services3338% of workers high or very high
- Healthcare3359% of workers high or very high
- Protective service249% of workers high or very high
- Personal care2211% of workers high or very high
- Health support171% of workers high or very high
- Transportation160% of workers high or very high
- Food service149% of workers high or very high
- Repair140% of workers high or very high
- Production130% of workers high or very high
- Farming110% of workers high or very high
- Construction110% of workers high or very high
- Cleaning & grounds40% of workers high or very high
How many workers are at each level
Higher pay, more exposure
In this research model, occupations with higher median pay also tend to have more tasks judged potentially accelerable. The highest-paid fifth scores 46 against 18 for the lowest-paid fifth. This is not evidence that higher-paid workers are already being displaced.
- Lowest paid$31k to $37k
- Lower paid$37k to $46k
- Middle$46k to $62k
- Higher paid$62k to $84k
- Highest paid$84k to $559k
Study scenarios: chatbot alone versus additional AI tools
The original study modeled both direct chatbot use and a wider scenario that assumes additional AI-enabled software. The second scenario covers more potential tasks; it is not measured real-world adoption. Legal has the widest modeled gap, from 11 to 84.
- Computers & math40 → 91
- Office & admin32 → 81
- Business & finance18 → 88
- Engineering19 → 78
- Science20 → 73
- Legal11 → 84
- Arts & media21 → 72
- Management10 → 80
- Sales22 → 64
- Education16 → 56
- Social services16 → 50
- Healthcare11 → 54
- Protective service15 → 34
- Personal care14 → 29
- Health support10 → 24
- Transportation8 → 24
- Food service8 → 21
- Repair6 → 22
- Production6 → 21
- Farming4 → 18
- Construction2 → 19
- Cleaning & grounds2 → 7
The highest- and lowest-scoring large occupations
Jobs with at least 100,000 U.S. workers. Pick any job to see its full check.
| Job | AI risk | Typical pay | Workers |
|---|---|---|---|
| Office Support Worker Exposure check | 78 | $45,670 | 192k |
| QA Tester Exposure check | 75 | $104,300 | 187k |
| Legal Secretary Exposure check | 73 | $55,570 | 156k |
| Executive Assistant Exposure check | 70 | $76,590 | 460k |
| Data Entry Clerk Exposure check | 70 | $41,340 | 127k |
| Web Designer Exposure check | 69 | $104,000 | 113k |
| PR Specialist Exposure check | 69 | $74,750 | 283k |
| IT Specialist Exposure check | 67 | $116,580 | 435k |
| Insurance Claims Clerk Exposure check | 67 | $49,230 | 214k |
| Software Developer Exposure check | 66 | $135,980 | 1.7M |
| Job | AI risk | Typical pay | Workers |
|---|---|---|---|
| Short Order Cook Exposure check | 0 | $35,880 | 139k |
| Busser Exposure check | 0 | $33,980 | 543k |
| Dishwasher Exposure check | 0 | $34,810 | 477k |
| Concrete Finisher Exposure check | 0 | $57,020 | 206k |
| Diesel Mechanic Exposure check | 0 | $61,770 | 290k |
| Tire Technician Exposure check | 0 | $37,710 | 108k |
| Lineman Exposure check | 0 | $95,320 | 131k |
| Meat Trimmer Exposure check | 0 | $38,300 | 146k |
| Fast Food Cook Exposure check | 1 | $30,890 | 641k |
| Janitor Exposure check | 1 | $36,840 | 2.2M |
How to read this
- The score is a study-based estimate of the share of tasks that might be at least twice as fast using AI at comparable quality, under the researchers’ assumptions. It is not the chance of losing a job.
- Potential exposure is different from observed AI usage, productivity gains, employment losses, and future industry demand. None is established by this score alone.
- Ratings reflect 2023-era model assumptions. This is a historical research baseline, not a continuously updated 2026 capability test.
- AvgPay groups the 0–100 scores into Low (below 15), Moderate (15–34), High (35–59) and Very high (60+). These thresholds are editorial presentation bands, not empirically calibrated risk or unemployment probabilities.
- 58 of 825 jobs are estimated from closely related jobs, because the study does not list them separately.
- Field and pay averages count each job by its number of workers, using BLS OEWS employment.
Source: Eloundou, T., Manning, S., Mishkin, P. and Rock, D. GPTs are GPTs: Labor market impact potential of LLMs. Science (2024); working paper arXiv:2303.10130 (2023). Read the study · Data (MIT license) · Our method