How We Estimated AI Risk for 10,000+ Careers (and Why They're Estimates)
Every job in Career Universe has a section called "Will it last?". It shows our estimate of AI and automation risk for the next 5 years and the next 10, and under "Why", one sentence on what changes and what stays human. This post lays out how those ratings were made, what they mean, and what they don't mean.
I'll say it up front: these are estimates. They are judgments made against a written rubric, not measurements and not predictions.
The question we asked
The easy question is "can this job be automated?" We didn't ask it. Almost every job has some tasks a machine could do. A list of those tasks tells you very little about whether anyone will hire a beginner in 2032.
The question for each job was:
Each part of that question matters.
- "AI and robots" means both generative AI (writing, code, images, analysis, customer chat) and physical automation (self-driving vehicles, warehouse robots, drones, automated inspection).
- "Shrink" means fewer openings or worse ones. The rubric asks about the number and quality of openings, not just whether some tasks can be automated.
- "For someone entering it" means we judged it from the newcomer's side. That includes what the rubric calls the entry-level squeeze: when AI does the junior work, companies hire fewer juniors, even if the senior job holds up.
That squeeze matters most for anyone just starting out, whether you're 16 or changing fields at 45. It's also the part early research suggests is showing up first. A Stanford Digital Economy Lab working paper, "Canaries in the Coal Mine?", found that for workers aged 22 to 25 in the jobs most exposed to AI, employment now stands about 19% below where it would be if it had kept pace with less-exposed peers (the August 2026 version of the paper, with data through June 2026). At the same time, the Budget Lab at Yale found no clear disruption to the broader job market in the 33 months after ChatGPT came out. Both can be true at once, and that's exactly why we judged the beginner's position.
Two horizons, one rule
Each job gets two ratings:
- In 5 years: through about 2031.
- In 10 years: through about 2036.
The rule: the 10-year rating can never be lower than the 5-year rating. Risk adds up over time.
The instructions also said robotics and autonomy count for more at 10 years than at 5, because robots and self-driving machines are slower to roll out than software.
What low, medium and high mean
Each rating is one of three words. No percentages, on purpose. A percentage would suggest a precision we don't have.
- Low: AI and robots are tools, and human demand holds steady or grows. This is typical of hands-on work in messy physical settings, work built on deep human trust and relationships, licensed accountability, and live performance.
- Medium: a real share of the tasks gets automated. There are fewer or different openings, especially at entry level, but the role stays and changes.
- High: the core tasks are largely automatable, and a big drop in openings is likely within the time frame.
Low doesn't mean safe. It means that in our estimate, AI looks more like a tool in that job than a replacement for the people doing it.
The anchors that kept us honest
A rubric like this drifts unless it's pinned to examples. So before rating anything, we fixed a set of anchor jobs and their ratings, and every other job was judged against them.
| Anchor jobs | 5 years | 10 years |
|---|---|---|
| Electrician, plumber, registered nurse, surgeon, K-12 teacher, firefighter, physical therapist, wildland firefighter, chef | Low | Low |
| Long-haul truck driver, radiologist, farmworker | Low | Medium |
| Software developer, accountant, pharmacist | Medium | Medium |
| Paralegal, graphic designer, warehouse picker, market research analyst, junior financial analyst | Medium | High |
| Customer service rep, data entry clerk, bookkeeper, copywriter, translator, telemarketer | High | High |
You may disagree with some of these. That's fine, and it's useful. If you think software developer should be low and low, or high and high, you now know which way to shift every rating in the map in your head. That's what the anchors are for. They make the ratings comparable across 29 very different sectors.
A few more rules came with them. Expect a real mix inside every sector, and don't mark everything low. Judge each job on its own instead of copying one answer across a whole field (similar jobs can share an answer when they really are the same). And each job already had an older, single AI rating from the first round of research. That older rating was an input to re-judge honestly, not the answer.
The one-sentence "how"
Every rating has to come with one sentence, 170 characters at most, that says concretely what changes and what stays human. No hype, no doom. On a job card you'll find it by tapping "Why" under "Will it last?".
A few real ones from the map:
- Family medicine physician, low and low: "AI scribes write visit notes and flag lab trends; examining patients, judging messy cases and long family relationships stay human."
- Saturation diver, low and low: "Resident subsea robots take more inspection work, but complex repairs on pipelines and wellheads still need divers' hands."
- Paralegal, medium and high: "AI drafts filings, summarizes depositions and organizes documents, the core junior work; fewer paralegals likely needed per lawyer."
I think that sentence is more useful than the rating. A word like "medium" tells you to look closer. The sentence tells you where.
Two kinds of automation
It helps to separate two clocks.
Generative AI moves fast and reaches desk work first: writing, summarizing, coding, routine analysis, customer chat. Robots and self-driving machines move slower and reach physical work later - driving, warehouse picking, inspection, some farm work.
That's why the 10-year rating leans harder on robotics, and why some physical jobs sit at low for 5 years and medium for 10. Two of the anchors show it. Long-haul truck driver and farmworker are both rated low for the next 5 years and medium for the next 10.
What the ratings came out to
By horizon
Across all 10,000+ careers in the map, our estimates landed here:
| Horizon | Low | Medium | High |
|---|---|---|---|
| Next 5 years (to about 2031) | 62% | 33% | 5% |
| Next 10 years (to about 2036) | 44% | 39% | 17% |
About three in ten jobs move up one band between the two horizons: 1,800+ from low to medium, and 1,200+ from medium to high. None jump two bands.
By sector
The share of jobs rated low at 10 years runs from 79% in Skilled Trades, 69% in Human & Social Services and 68% in Personal Care & Services, down to 14% in Computing & Technology and 10% in Finance, Banking & Insurance.
By kind of work
The pattern is starker by kind of work. Of the jobs we classed as physically heavy, 80% are rated low at 10 years. Of desk jobs, 5% are. Of jobs where fully remote work is common, 1% are rated low and 71% are rated high.
Read those numbers as a description of our judgments, not a forecast about the economy. They show how we applied the rubric. They would tell you something about the future only if the rubric is right, and that is exactly the thing we can't promise.
What this method is not
- It's not a probability. "High" doesn't mean 80% of those jobs will vanish. It means the core tasks look largely automatable and we expect a big drop in openings within the time frame.
- It's not a BLS or O*NET figure. These ratings don't come from the Bureau of Labor Statistics or O*NET. They're ours.
- It's not computed from the research indexes. Well-known studies from OpenAI and the University of Pennsylvania, Brookings, the ILO and Anthropic mostly measure exposure: how much of a job's tasks AI could touch. We didn't compute our ratings from their scores. Ours ask a different question, about openings for newcomers.
- It's not peer-reviewed. The ratings were written with AI research assistants against a strict schema, checked by validation scripts and spot-checked by hand, by a very small team.
- It's not fixed. When the evidence moves, the ratings should move. We update the data from time to time and fix reported errors.
How to report a rating that looks wrong
Pick a job you know well. Search for it in the map, open it, and tap "Why" under "Will it last?" to read the "how" sentence.
If it looks wrong, email privacy@tarvacore.ai with:
- the job title,
- the rating you'd give at 5 years and at 10 years,
- one sentence on why.
People who do a job see things a rubric can't. That address comes straight to me.
Questions people ask
How were the AI-risk ratings decided?
Every job was rated against the written rubric in this post: one fixed question, two horizons, three possible ratings, and a set of anchor jobs so "low" means the same thing in every sector. Every rating needed a one-sentence reason saying what changes and what stays human.
Were the ratings made by AI?
Partly, yes. The ratings were written with AI research assistants against a strict schema, checked by validation scripts and spot-checked by hand, by a very small team. That's one more reason to treat them as estimates and to tell us when one looks wrong.
Why doesn't the map show a percentage?
Because a percentage would suggest a precision we don't have. Three words, low, medium and high, plus a sentence on how, is about as exact as an honest estimate can be.
How often are ratings updated?
There's no fixed schedule. We update the data from time to time and fix reported errors.
What if I disagree with a rating?
Tell us. Email privacy@tarvacore.ai with the job title, the 5-year and 10-year ratings you would give, and one sentence on why.
Your next step: pick a job you know well. Open it in the map, read its "how" sentence under "Will it last?", and tell us if we got it wrong.
Sources: Brynjolfsson, Chandar and Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence", Stanford Digital Economy Lab, 2025, revised August 2026. Gimbel, Kinder, Kendall and Lee, "Evaluating the Impact of AI on the Labor Market: Current State of Affairs", The Budget Lab at Yale, October 2025. Counts and shares are from our own estimates in the map. Parts of the research behind Career Universe were written with AI research assistants and checked by hand.