Karpathy on AI Exposed Jobs and the Future of Work

The US Job Market Visualiser at karpathy.ai/jobs is one of the most useful tools of 2026. It was made by Andrej Karpathy, who used to be the Director of AI at Tesla and is a founding member of OpenAI. This interactive treemap looks at 342 jobs that make up 143 million jobs in the US using official Bureau of Labour Statistics (BLS) data.

It’s not an academic paper or an official economic report. It’s a useful, open-source tool for developers and analysts to make visualisations. The best part? A personalised “Digital AI Exposure” score (0–10) based on a large language model (LLM). This score tells you how much AI (especially digital tools like LLMs) can change or do each job.

This in-depth guide explains what Karpathy’s project shows, which jobs are most likely to be replaced by AI, which jobs are the safest, what it means for the future of work, whether you should panic, and what you can do to stay irreplaceable.

What Is Karpathy’s AI Exposed Jobs Report About?

Karpathy’s visualiser makes boring BLS data into a beautiful, colour-coded treemap:

The size of the rectangle is equal to the number of people who work in that field.
Colour = selected metric (you can switch between expected growth, median pay, education requirements, or AI exposure)

The AI Exposure layer gives each job a score based on the official BLS description of it. The goal is simple: figure out how much digital AI can change a job by automating it or making it much more productive.
Important numbers from the tool:

Overall weighted average AI exposure: 4.9 out of 10 (some sources say the unweighted average is about 5.3 out of 10)
People who make more than $100,000 a year have an average exposure of 6.7/10. White-collar knowledge work is the most affected.
People who make less than $35,000 a year only get an average of 3.4 out of 10.

GitHub has the whole project, including scrapers, the exact LLM prompt, and the raw scores. Anyone can run the pipeline again or make their own scores, like for robotics or offshoring risk.

How AI Exposure Scores Are Calculated

Karpathy’s pipeline:

Scrapes every BLS occupation page
Cleans the descriptions
Feeds each one to an LLM (Gemini Flash) with a detailed rubric
The model returns a score 0–10 + rationale

The rubric focuses on one core question: “How much of this job is purely digital and screen-based?”

Scoring anchors (direct from the project):

0–1 (Minimal): Purely physical, unpredictable environments (e.g., roofer, landscaper)
2–3 (Low): Mostly physical or real-time human interaction
4–5 (Moderate): Mix of physical + knowledge work
6–7 (High): Knowledge work with some human judgment
8–9 (Very High): Almost entirely digital (writing, coding, analysis)
10 (Maximum): Routine digital tasks AI already handles perfectly (e.g., data entry)

Important disclaimer from Karpathy himself:

“A high score does not predict the job will disappear. Software developers score 9/10 because AI is transforming their work — but demand for software could easily grow as each developer becomes more productive.”

Jobs Most Likely to Be Replaced (or Transformed) by AI — Highest Exposure

These roles get 8 to 10 out of 10. They are mostly digital, work well from a distance, and make outputs that can be seen on screens. AI can already do a lot of the work.
The most exposed jobs, with scores and job openings when they are available, are:

Transcriptionists in the medical field: 10 out of 10 (about 53,000 jobs)
9 out of 10 bookkeeping, accounting, and auditing clerks (1.6 million jobs)
9 out of 10 accountants and auditors work (1.5 million jobs)
Financial analysts—9 out of 10 (~350,000 jobs)
Lawyers—9 out of 10 (about 800,000 jobs)
Management analysts: 9 out of 10 (about 1 million jobs)
8–9 out of 10 software developers (about 1.8 million jobs)
9 out of 10: computer programmers, database administrators, data scientists, and mathematicians
8–9 out of 10 for paralegals, writers, editors, copywriters, graphic designers, and market researchers
8 out of 10 customer service representatives (2.8 million jobs)

Data analysts, translators, journalists, and many administrative and office jobs are also examples of jobs with a lot of exposure.

Pattern: If you can do your job from a laptop and your entire output is a digital file (code, report, design, contract, analysis), your exposure is very high.

Jobs That Are Safe — Lowest AI Exposure

These get a score of 0–3 out of 10. They need people to be there in person, be able to move around in the real world, or have interactions with people that are hard to predict, which AI (and even robots) can’t do well.
Jobs with the least risk (safest):

Roofers, construction workers, painters, ironworkers, groundskeepers, and janitors all get a score of 0–1/10.
Landscapers and commercial divers—0 to 1 out of 10
Electricians, plumbers, and firefighters get 2–3 out of 10.
Dental hygienists: 2–3 out of 10

Moderate-safety examples (4–5 out of 10):

Registered nurses, police officers, veterinarians, retail workers, and doctors (a mix of physical and knowledge work)

Key insight: About 25 million jobs are in the “very high exposure” group (8–10), but most of them are physical and hands-on. The exposed economy is the knowledge economy.

What Does This Mean for the Future of Work?

Karpathy’s data shows that AI won’t cause a lot of people to lose their jobs, instead it will change the way people work.

Jobs with a lot of exposure get more done. A single software developer with AI tools could do the work of three people.
There could be more demand for many jobs (there is a lot of latent demand for software, content, and analysis).
Highly exposed jobs pay $3.7 trillion a year, which means big changes and chances in the economy.
42% of jobs get a score of 7 or higher (59.9 million workers are affected in some way).

History repeats itself: computers didn’t get rid of secretaries or accountants; instead, they made more work for them and started new businesses. AI is like the same story but bigger.

The future of work is people and AI working together. Automated tasks that are purely digital, but roles that need a physical presence, empathy, creativity in uncertain situations, or real-world judgement stay (and often do well).

Should We Be Worried?

Not really — but we should get prepared.

Karpathy’s own take is optimistic: high exposure often signals opportunity, not obsolescence. White-collar professionals earning high salaries face the biggest changes, but that also means they have the most upside from AI leverage.

The real risk isn’t AI replacing you — it’s someone using AI replacing you. Those who ignore the tools will lose ground; those who master them will pull ahead dramatically.

How to Become More Irreplaceable by AI — 6 Practical Steps

Master AI tools today — Learn prompting, use Cursor/Claude/ChatGPT for your specific job. Turn yourself into a “centaur” (human + AI hybrid) who is 5–10x more productive.
Add irreplaceable human elements — Focus on empathy, negotiation, physical dexterity, or real-time decision-making in unpredictable settings.
Develop hybrid skills — Combine domain expertise with AI fluency (e.g., AI-assisted legal research, AI-generated design + human creativity).
Target low-exposure career paths or layers — Trades, healthcare with hands-on care, trades requiring physical presence, or roles involving complex human relationships.
Build personal brand and relationships — AI can’t replicate trust, reputation, or networks.
Commit to lifelong learning — The half-life of skills is shrinking. Re-skill every 2–3 years.

Explore Karpathy’s Tool Yourself

Head to karpathy.ai/jobs/, toggle the AI Exposure layer, and click any tile for the full BLS breakdown. The entire pipeline is open-source on GitHub if you want to dive deeper or create your own custom risk scores.

Bottom line: AI isn’t coming for all jobs equally. Digital, screen-based work is being transformed fastest. Physical, human-centric, and unpredictable roles remain safest. The winners in the AI era won’t be those whose jobs are “safe” — they’ll be those who use AI as the ultimate multiplier.

Stay curious, start experimenting with AI tools this week, and you’ll not only survive the future of work — you’ll thrive in it.

Learn more about what will happen in the AI landscape with NVIDIA GTC 2026 and the Next Wave of AI