Andrej Karpathy’s AI job displacement scoring project covering 342 BLS occupations
Karpathy Just Scored Every Job in America. I’m Not Sure I Like What I See.
There’s a particular kind of clarity that lands like a gut punch. Andrej Karpathy, one of the founding architects of modern deep learning and a former Tesla AI director, just released an open-source project that scores all 342 Bureau of Labor Statistics occupations by AI displacement risk, on a scale of 0 to 10. The methodology is clean. The results are uncomfortable. And if you work in tech, you should probably look at it before someone else shows it to you at a bad moment.
The Methodology
Karpathy scraped every occupation from the BLS database, fed each one into an LLM with a structured scoring rubric, and visualized the output as an interactive treemap. Rectangle size represents the number of employed Americans in that role. Color represents exposure level. The whole pipeline is open source: the scraper, the scoring logic, the visualization. No black box.
The core heuristic driving the scores is blunt in a way that cuts through most of the discourse I see on this topic. If the work product is fundamentally digital and the role can be performed entirely from a home office, exposure is inherently high. That’s it. That’s the filter.
What the Numbers Say
The distribution is not subtle.
Medical transcriptionists score a 10. Full replacement territory. Software developers and data analysts land at 8 to 9. Paralegals, same range. Roofers and janitors score 0 to 1, not because the work is simple but because it is physical, contextual, and irreducibly present in the real world. Nurses and physicians sit in the 4 to 5 range, held down by the physical examination component and regulatory liability.
The national average across all 342 occupations is 5.3 out of 10.
That number matters. People tend to frame AI displacement as a problem for a narrow slice of the workforce. A 5.3 average says otherwise. We are past the point where this is someone else’s problem.
🔍 The Irony Is Not Lost on Me
I build AI systems for a living. My own category scores an 8 or 9 on this rubric. I am in the business of building the thing that the rubric says will eat my business.
I’ve thought about this a lot. My honest read is that the score is probably correct directionally but overstates the near-term timeline for roles that require architectural judgment, system-level reasoning, and the ability to make calls that have no clean answer. What LLMs are very good at right now is the middle layer of software work: the stuff that’s well-specified, syntactically bounded, and reviewable by output alone. The harder work, deciding what to build and why, is still genuinely difficult for current systems.
But “still difficult for current systems” is not the same as “safe.” The timeline is not fixed.
What This Actually Predicts
Karpathy’s rubric is not a prediction about which jobs will disappear tomorrow. It is a map of structural vulnerability. Roles scoring 8 to 9 are not guaranteed to vanish, but they are the ones where a 30% productivity gain from AI tools translates directly into fewer headcount decisions at the margin. You don’t need full automation for the job market to contract. You just need fewer new positions opened and slower backfill on attrition.
That dynamic is already visible in tech hiring data from 2024 and early 2025. The headcount didn’t crater overnight. It just stopped recovering the way it did after previous downturns.
Naval Ravikant put it flatly on X last week: “Software was eaten by AI.” Elon Musk replied with a single word in agreement. That’s a thin data point, but the sentiment is consistent with what I see in actual engineering org charts right now.
Where This Goes
The most useful thing about Karpathy’s project is that it gives people a concrete frame instead of an abstract anxiety. You can look at your occupation. You can see the score. You can then make an informed decision about whether to upskill, pivot, or bet that the timeline is longer than the model suggests.
That last option is not irrational. Timelines in this field slip constantly. But I’d want to be making that bet consciously, not by default.
The full project is worth exploring directly at https://karpathy.ai and the BLS occupation data it draws from lives at https://www.bls.gov/oes/current/oes_nat.htm.
The average score is 5.3. Most of us are not on the safe side of that number.
Sources
#AIDisplacement #FutureOfWork #MachineLearning #SoftwareEngineering #ArtificialIntelligence
