Andrej Karpathy Releases AI Replacement Scores for 342 Jobs
Andrej Karpathy analyzed 342 U.S. occupations from BLS data and assigned each an AI exposure score on a 0-to-10 scale. The overall average is 5.3, with actual scores ranging from 1 to 9. Medical transcriptionists topped at 9, while athletes and construction laborers scored the lowest at 1. The data reveals a paradox: the higher-paying and more digital the job, the greater its AI exposure.
Andrej Karpathy has released yet another open-source project. This time it's not about AI itself, but about the jobs AI is poised to affect. He pulled 342 occupations from the U.S. Bureau of Labor Statistics (BLS) database and scored each on a 0-to-10 scale of AI exposure. The results are available as an interactive treemap at karpathy.ai/jobs.
The overall average is 5.3 on a 0-to-10 scale, meaning more than half of all occupations fall within AI's sphere of influence. Notably, actual scores ranged from 1 to 9, with no occupation hitting a perfect 0 or 10. Karpathy used Gemini Flash to analyze the task descriptions of each occupation, then quantified how automatable those tasks are with current AI technology.
A Top Score of 9 for Transcriptionists, Just 1 for Workers in the Field
On the 0-to-10 scale, the occupation that topped the chart at 9 is medical transcriptionist. The job of converting speech to text is something AI can already do faster and more accurately than humans. It's widely considered the occupation most immediately replaceable by AI.
Sharing that top score of 9 are familiar titles: software developers, paralegals, data analysts, and editors. All digital-first occupations dealing in text and code. At the opposite end, athletes and construction laborers sit at the lowest score of 1. Work that requires physical presence and manual labor in real-world spaces remains largely beyond AI's reach.
| Rank | Occupation | Score |
|---|---|---|
| 1 | Medical Transcriptionist | 9 |
| 2 | Software Developer | 9 |
| 3 | Data Scientist | 9 |
| 4 | Editor | 9 |
| 5 | Economist | 9 |
| 6 | Financial Analyst | 9 |
| 7 | Computer Programmer | 9 |
| 8 | Database Administrator | 9 |
| 9 | Bookkeeping/Accounting Clerk | 9 |
| 10 | Customer Service Representative | 9 |
| Rank | Occupation | Score |
|---|---|---|
| 1 | Athlete | 1 |
| 2 | Construction Laborer | 1 |
| 3 | Drywall Installer | 1 |
| 4 | Animal Care Worker | 2 |
| 5 | Auto Body Repairer | 2 |
| 6 | Baker | 2 |
| 7 | Barber/Hairstylist | 2 |
| 8 | Bartender | 2 |
| 9 | Butcher | 2 |
| 10 | Childcare Worker | 2 |
Nurses at 4, Software Developers at 9: An Unexpected Gap
What's striking is the wide variance even among highly skilled professions. Nurses and physicians hover around 4-5. Their physical tasks, hands-on patient interaction, and on-the-spot clinical judgment serve as a shield against automation. Software developers, however, score 9. Code writing, debugging, and review all happen entirely within digital environments.
The pattern is clear: the more digitally self-contained the work, the higher the AI exposure; the more physical interaction is required, the lower. And this pattern surfaces an uncomfortable truth: higher-paying white-collar jobs tend to have greater AI exposure.
From 'Feeling Behind' to Two Projects in a Week
This project carries extra weight because of its context. On March 6, Karpathy released autoresearch, an open-source project demonstrating the potential for automated AI research. Just before that, he posted on X (formerly Twitter) about 'feeling behind,' a post that garnered 14 million views. Even someone at the forefront of AI confessing to being overwhelmed by the pace struck a deep chord.
In less than a week, Karpathy followed autoresearch with the jobs project, channeling his sense of falling behind into raw productivity. The full source code and data are available on GitHub, so anyone can audit the scoring methodology or analyze their own occupation.
