Andrej Karpathy, the former Tesla AI chief and OpenAI co-founder, just released an open-source tool that scores every major U.S. occupation on its estimated exposure to current AI capabilities. The project pulls 2024 data from the Bureau of Labor Statistics' Occupational Outlook Handbook — covering 342 job types, roughly 143 million workers, and every sector of the American economy — then runs each occupation through an LLM to score it on a 0-to-10 scale of AI exposure. Karpathy is upfront that this is not a formal economic publication; it's a development tool, a structured way to see where AI's current capabilities overlap most with the work people actually do.
The average score across occupations is 5.3 out of 10, which suggests that a large share of U.S. jobs already overlap meaningfully with AI capabilities. But the real story isn't the average — it's the extremes.
The $3.7 Trillion in Exposed Wages
When you isolate the occupations scored 7 or higher — the "high exposure" tier — the numbers are striking. There are 130 occupations in this band, representing 49 million jobs. Using BLS employment counts and median annual wages, the high-exposure tier represents roughly $3.7 trillion in annual wages. That's more than a third of all jobs tracked by the BLS, concentrated in the parts of the economy that produce the most economic value per worker.
To be clear, this is a back-of-the-envelope aggregate — BLS employment multiplied by BLS median pay, summed across occupations scoring 7 or above — not a formal economic forecast. But the scale of overlap between AI capabilities and current employment is hard to ignore.
The Pattern: If Your Work Lives on a Screen, You're Exposed
The single strongest predictor of a high AI exposure score is deceptively simple: does the job's output live on a screen?If your work product is fundamentally digital — text, code, data, designs, financial models — then AI can, in principle, replicate or dramatically augment what you do. Karpathy's scoring rubric makes this explicit: jobs that can be done entirely from a home office on a computer have inherently high exposure.
This explains why the top of the exposure list reads like a roster of knowledge economy jobs, including many that were considered safe, prestigious, and well-compensated just a few years ago.
The Perfect 10: Medical Transcriptionists
Only one occupation scored the maximum 10 out of 10: medical transcriptionists (43,900 jobs, $37,550 median pay). The rationale is straightforward — this is a purely digital, routine information-processing task where AI speech recognition and large language models have already reached near-human accuracy. The BLS projects employment in this role to decline outright over the 2024–2034 period. It's the clearest case of an occupation where AI doesn't just assist — it can perform the core function.
The 9/10 Tier: Where the Volume Is
Thirty occupations scored 9 out of 10, and this is where the sheer scale of exposure becomes apparent.
Customer service representatives — 2.8 million jobs at $42,830 median pay. The core of the role — answering questions, processing orders, resolving complaints — maps directly onto what conversational AI already does at scale. BLS projects this category to decline.
General office clerks — 2.6 million jobs at $43,630. Data entry, document formatting, scheduling, and information processing are among the most automatable tasks in any economy. Also projected to decline.
Software developers, QA analysts, and testers — 1.9 million jobs at a median pay of $131,450. This might be the most counterintuitive entry on the list. The people building AI tools are themselves scored as highly exposed, because coding, debugging, and test automation are precisely the tasks where large language models demonstrate the most dramatic productivity gains. Notably, the BLS still projects faster-than-average growth for this category — suggesting the field may restructure around potentially fewer developers per unit of output, but with much higher individual productivity.
Bookkeeping, accounting, and auditing clerks — 1.6 million jobs at $49,210, already projected to decline over the 2024–2034 period.
Financial clerks — 1.2 million jobs at $48,650. Routine data entry, record updating, and basic financial calculations are textbook cases for AI-assisted workflows.
Other 9/10 entries include market research analysts (942,000 jobs), financial analysts (429,000), paralegals (376,000), graphic designers (266,000), data scientists (246,000), web developers (215,000), writers and authors (135,000), editors (116,000), interpreters and translators (75,300), and computer programmers (121,200 — already in projected decline).
The 8/10 Tier: White-Collar Mainstays
At 8 out of 10, the pattern extends into roles traditionally considered insulated by professional barriers: secretaries and administrative assistants (3.5 million jobs), accountants and auditors (1.6 million at $81,680), lawyers (865,000 at $151,160), computer systems analysts (521,000), and advertising and marketing managers (434,000 at $159,660).
Lawyers are a particularly interesting case. At $151,160 median pay and 865,000 jobs, the legal profession represents one of the highest-wage categories in the high-exposure tier. The core of legal work — research, document review, contract analysis, regulatory interpretation — is deeply textual and increasingly within reach of LLMs. While courtroom presence, client relationships, and professional judgment provide real insulation, the economics of how legal work gets done are already shifting.
What Exposure Does Not Mean
High exposure does not mean a job disappears. In many cases, it means AI can handle a growing share of the workflow, pushing humans toward supervision, exception handling, judgment calls, and relationship-heavy tasks. Some occupations may shrink in headcount. Others may grow in total employment while the work inside them changes dramatically — fewer hours spent on routine production, more on strategy, oversight, and the parts of the job that require human presence or trust.
This distinction matters. Karpathy's tool measures susceptibility to AI capabilities, not a timeline for displacement. The gap between "AI can do parts of this job" and "employers will restructure around that capability" involves regulation, institutional inertia, client expectations, and the uneven pace of adoption.
Why Some High-Exposure Jobs May Still Grow
One of the most revealing cross-references in the dataset is comparing AI exposure scores against BLS 2024–2034 employment projections. Of the 130 high-exposure occupations, 29 are already projected to decline — including bookkeeping clerks, computer programmers, general office clerks, bill collectors, and insurance underwriters.
But several high-exposure occupations show faster-than-average projected growth: software developers, data scientists, market research analysts, and operations research analysts among them. This isn't necessarily a contradiction. It likely reflects a restructuring dynamic: demand for the output of these roles keeps rising, but AI-augmented workers can produce more of it per person. Total employment may grow even as the ratio of workers to output shifts. The work expands; the headcount per unit of work may not.
The Other End of the Spectrum: What AI Can't Reach
For context, the occupations scoring 1 to 2 are overwhelmingly physical, hands-on roles: construction laborers, janitors, roofers, electricians, plumbers, carpenters, home health aides, childcare workers, barbers, and bartenders. These 45 lowest-scoring occupations represent about 53 million jobs.
The dividing line isn't skill or education level — it's physicality. A plumber earning $63,000 scores a 2. A software developer earning $131,000 scores a 9. The traditional economic hierarchy, where digital knowledge work commands a premium over manual labor, is being directly challenged by the same technology that helped create that premium in the first place.
What This Means
Karpathy's project is not a labor market forecast. It is a provocative, structured snapshot of where current AI capabilities overlap most with existing occupations. And the signal is clear: the knowledge economy is not insulated from AI exposure. In many cases, it is the primary target.
The jobs most exposed are not the factory roles people feared a decade ago. They are the screen-based professions that define modern white-collar work: analysts, developers, designers, accountants, legal staff, and clerks. The question is no longer whether AI can assist with this work. It is how much of it can be done with fewer people — and how quickly organizations act on that.
It turns out, a growing share of routine cognitive work is exactly what AI has gotten good at.
Data source: karpathy.ai/jobs — built from 2024 BLS Occupational Outlook Handbook data with 2024–2034 projections, scored using Gemini Flash via OpenRouter. Karpathy notes this is a development tool, not a formal economic publication.
Unlock the Future of Business with AI
Dive into our immersive workshops and equip your team with the tools and knowledge to lead in the AI era.