How many US jobs sit directly in the path of AI automation, and how much total salary is at stake? Andrej Karpathy, the former Tesla AI director and OpenAI researcher, just published an interactive visualization that attempts to answer both questions with real employment data.
The project pulls occupation data from the Bureau of Labor Statistics Occupational Outlook Handbook, covering 342 distinct job categories. Each occupation gets an AI exposure score from 0 to 10, generated using Google's Gemini Flash model to evaluate how much of each job's core tasks could be performed or significantly assisted by current AI systems.
How the Scoring Works
Occupations fall into five tiers: minimal (0-1), low (2-3), moderate (4-5), high (6-7), and very high (8-10). The visualization displays jobs as a treemap where the size of each block represents total employment and the color indicates exposure level, from cool tones at the low end to warm tones for heavily exposed roles.
The interactive tooltips show median annual wages, employment counts, growth outlook, and a written rationale for each occupation's score. You can filter by salary band and education level, which reveals some patterns that aren't obvious at first glance.
Following the Money
The most striking metric is "wages exposed," which calculates total annual wages across all occupations scoring 7 or above. This number lands in the trillions of dollars, representing the aggregate salary pool in roles where AI could substantially change how work gets done.
Karpathy also breaks exposure down by pay bands and education requirements through horizontal bar charts. This lets you see, for example, whether highly paid knowledge workers face more AI exposure than lower-wage service roles, or whether a graduate degree offers any insulation from automation risk.
A Useful Tool With Clear Limitations
Using an AI model (Gemini Flash) to score AI exposure is an inherently circular exercise, and Karpathy doesn't hide this. The scores reflect what a language model thinks about its own capabilities and those of similar systems, which introduces obvious bias toward overestimating exposure in language-heavy roles and potentially underestimating it in fields where AI progress is happening through specialized systems rather than large language models.
That said, the BLS employment and wage data underneath is solid, and the visualization itself is genuinely well built. As a conversation starter about which parts of the labor market face the most near-term pressure from AI tools, it's more grounded than the vague "AI will change everything" claims that dominate most discussions. The full project is open source on GitHub for anyone who wants to dig into the methodology or run it with different scoring criteria.