Daron Acemoglu, the Nobel-winning MIT economist, just sat down with MIT Tech Review and laid out why he’s still not buying the AI jobs apocalypse narrative. Two years after his cautious take got dismissed by half the industry, the data keeps backing him up. Studies still find AI isn’t moving the needle on employment rates or layoffs.
What’s significant here is that Acemoglu isn’t a blanket skeptic. He’s been researching AI’s economic impact since 2018, and he’s flagged three specific shifts worth tracking right now.
Agents are a “losing proposition” as worker replacements
The biggest technical leap since Acemoglu’s original paper is agentic AI, tools that operate independently to complete goals rather than just answering prompts. Companies are pitching them as one-to-many human replacements. Acemoglu calls that “a losing proposition.”
His reasoning is grounded in task analysis. An x-ray technician handles roughly 30 different tasks, from patient histories to organizing mammogram archives. A human switches between formats, databases, and styles without thinking. An agent would need a separate tool or protocol for each transition.
The real question isn’t whether agents can write code or summarize docs. It’s whether they can orchestrate the messy connective tissue between tasks. AI labs are racing to extend how long agents can run without errors, sometimes inflating the numbers. Acemoglu’s bet: until agents can fluidly switch contexts, most jobs stay intact.
AI companies are quietly building economics departments
This is the trend that stands out most in the MIT Tech Review piece. The big labs aren’t just poaching AI researchers anymore. They’re hiring economists.
- OpenAI brought in Ronnie Chatterji from Duke as chief economist, paired with former Obama advisor Jason Furman to study AI and jobs.
- Anthropic convened a panel of 10 leading economists for similar research.
- Google DeepMind just hired Alex Imas from the University of Chicago as “director of AGI economics.”
That last title says a lot. The frontier labs aren’t just preparing for AGI technically. They’re staffing up to model what it does to labor markets, productivity, and policy. Read between the lines: they expect regulators, journalists, and politicians to start asking harder questions, and they want their own data ready.
The political climate is moving faster than the data
Acemoglu’s measured take hasn’t caught on publicly. Bernie Sanders is talking AI jobs at rallies. A California gubernatorial candidate proposed taxing corporate AI use to compensate “victims of AI-driven layoffs.” Previously skeptical economists are softening their position.
The gap between what the data shows (no employment effect yet) and what the public believes (jobs are vanishing) is itself becoming a policy risk. Even if AI doesn’t trigger mass layoffs, tax policy built on the assumption that it will could reshape how companies deploy the tech.
What practitioners should do
A few practical takeaways from Acemoglu’s framing:
- Stop benchmarking agents on full job replacement. Benchmark them on the 5-10 specific tasks they actually do well inside a workflow.
- Watch the orchestration layer. The teams winning with agents are the ones solving context-switching, not the ones extending single-task runtime.
- Track the economist hires. When Anthropic and OpenAI publish labor-market research, it’ll shape regulation. Don’t get caught flat-footed.
- Expect AI taxes to enter serious policy debates within 18 months, regardless of whether the underlying jobs data supports them.
Acemoglu’s track record on this is worth respecting. His earlier work estimated AI’s GDP impact at well under what the labs were promising, a number that drew mockery from companies pitching 10x productivity gains. So far, his estimate is closer to reality than theirs. The bigger story isn’t whether he’s right about agents. It’s that the labs themselves are now hiring people to figure out exactly how wrong or right he might be. Full piece over at MIT Tech Review.