The Viral AI Labor Market Experiment
In the rapidly evolving landscape of artificial intelligence, few figures command as much attention as Andrej Karpathy. A co-founder of OpenAI and the former director of AI at Tesla, Karpathy has long been a bellwether for the industry’s direction. Over the weekend of March 15, 2026, he set the tech community ablaze—and then just as quickly sought to dampen the flames—with a project he dubbed a "vibe-coded" AI labor market analysis.
The project, which was briefly hosted online before being taken down, was a data visualization experiment designed to map the exposure of 342 U.S. occupations to AI automation. Using data from the U.S. Bureau of Labor Statistics (BLS) Occupational Outlook Handbook, Karpathy fed job descriptions into a large language model to score each occupation on an "AI exposure" scale from zero to 10. The results, visualized in a colorful treemap, immediately went viral, largely due to a stark, counterintuitive finding: high-paying, white-collar professions are significantly more exposed to AI disruption than low-wage, manual labor roles.
Methodology: The "Vibe-Coded" Approach
The term "vibe-coding"—a phrase Karpathy helped popularize in previous months—refers to a development methodology where the programmer relies heavily on LLMs to generate code, often without manual review or deep debugging, prioritizing the "vibes" or general direction of the output over rigorous software engineering standards.
In this specific labor market analysis, Karpathy applied this high-level, AI-assisted approach to social science data. By scraping 342 distinct job categories, he tasked the AI with analyzing the core responsibilities of each role to determine how much of that work could theoretically be reshaped or performed by automated systems.
The methodology relied on a straightforward premise:
- Input Data: The comprehensive BLS Occupational Outlook Handbook.
- Scoring Mechanism: A large language model evaluated each job description against AI capabilities.
- The Metric: A zero-to-10 scale where 10 represented maximum exposure to AI-driven automation.
While critics in the data science community quickly pointed out that this methodology lacks the academic rigor of a peer-reviewed study, it succeeded in creating a "ground-truth" visualization that resonated with the public. It moved beyond theoretical debates and provided a concrete, albeit imperfect, look at how the AI transition might impact the modern workforce.
Key Findings: The Automation Exposure Gap
The headline result of Karpathy’s analysis challenged a common assumption that AI automation would first target entry-level or routine manual tasks. Instead, the data suggested that highly skilled, high-income professionals face the highest exposure scores.
The analysis revealed a consistent pattern: jobs that produce digital artifacts—code, text, reports, or administrative data—scored significantly higher on the exposure scale. Conversely, jobs requiring physical dexterity and real-world environmental interaction, such as roofing or maintenance, remained largely insulated.
The following table summarizes the exposure distribution across income levels based on the project’s data:
| Job Category Characteristics | Avg. Exposure Score (0-10) | Primary Risk Factor |
|---|---|---|
| High-Income Roles ($100k+) | 6.7 | Cognitive/Digital output Information processing Routine logic tasks |
| Mid-Income Roles ($60k-$99k) | 4.9 | Hybrid tasks Administrative support Data management |
| Low-Income Roles (<$35k) | 3.4 | Physical presence required Real-world dexterity Unpredictable environments |
This data paints a picture of a "white-collar crisis" in the making, where the premium placed on knowledge work is being eroded by the very tools designed to enhance productivity. While software developers, legal analysts, and administrative managers saw scores ranging from 8 to 9, roles requiring physical presence consistently landed in the low-exposure bracket.
Why the Project Was Retracted
Shortly after the interactive visualization gained viral traction, Karpathy removed the website and the associated GitHub repository. In follow-up commentary, he clarified that the project was a "quick experiment," not a formal, scientific study. He explained that the project was born out of a desire to explore a specific question using the tools he had on hand, rather than to provide a definitive policy framework for the U.S. government or labor organizations.
The retraction highlights a growing tension in the AI research community: the speed at which AI influencers can generate and distribute data-driven insights versus the time required for traditional verification. While the map was influential, the nuance was often lost in the rapid dissemination of the results on social platforms like X (formerly Twitter) and Reddit, where the findings were frequently cited as absolute fact rather than a "vibe-coded" exploration.
Contextualizing with Institutional Research
It is important to note that Karpathy’s experiment did not exist in a vacuum. It emerged during the same week that researchers at Anthropic published a major, peer-reviewed labor market study. The Anthropic study focused on "observed exposure"—what AI is actually doing in professional workflows, rather than what it theoretically could do.
| Study Focus | Approach | Main Finding |
|---|---|---|
| Karpathy's Analysis | Theoretical/Potential | High-paying jobs are most exposed due to digital task nature |
| Anthropic Study | Observed/Deployed | Large gap exists between AI capability and actual enterprise deployment |
The Anthropic study provided a much-needed cooling effect on the narrative of imminent, mass-scale displacement. By tracking actual Claude usage data, the researchers found that while AI is technically capable of performing many tasks, the actual deployment in enterprise settings remains limited. The "gap" is not a warning of immediate job loss, but rather a "to-do list" for companies still navigating the challenges of integration, change management, and cultural adaptation.
The Future of Work: A Shifting Landscape
Despite the retraction of Karpathy’s project, the conversation it ignited remains critical. The core takeaway from both the viral "vibe-coded" map and the more formal Anthropic research is consistent: the workforce is not facing an overnight total replacement by AI. Instead, we are entering a period of fundamental role evolution.
For professionals earning over $100,000 annually, the exposure scores serve as a wake-up call regarding the nature of their daily tasks. The shift is moving away from purely execution-based work toward curation, management, and strategic oversight. The AI tools that are capable of replacing entry-level tasks are simultaneously becoming the primary levers of productivity for the roles that remain.
As AI continues to mature, the distinction between "highly exposed" and "safe" will likely blur. The most successful workers in the coming decade will not be those who ignore these trends, but those who embrace AI as a co-pilot, treating their own skill sets as dynamic variables in an increasingly automated economy. Karpathy’s experiment, while fleeting, successfully highlighted the reality that the future of work is not just about the technology we build, but how we choose to integrate it into the fabric of the global labor market.
