I would love to read that study, as going off of your comment I could easily see it being a case of “more than 10% of jobs are bullshit jobs à la David Graeber so having an « AI » do them wouldn’t meaningfully change things” rather than “more than 10% of what can’t be done by previous automation now can be”.
The study you are referring to was released in late November 2025. It is titled “The Iceberg Index: Measuring Workforce Exposure in the AI Economy.”
It was conducted by researchers from MIT and Oak Ridge National Laboratory (ORNL).
Here are the key details from the study regarding that “more than ten percent” figure:
The Statistic: The study found that existing AI systems (as of late 2025) already have the technical capability to perform the tasks of approximately 11.7% of the U.S. workforce.
Economic Impact: This 11.7% equates to roughly $1.2 trillion in annual wages and affects about 17.7 million jobs.
The “Iceberg” Metaphor: The study is named “The Iceberg Index” because the researchers argue that visible AI adoption in tech roles (like coding) is just the “tip of the iceberg” (about 2.2%). The larger, hidden mass of the iceberg (the other ~9.5%) consists of routine cognitive and administrative work in other sectors that is already technically automated but not yet fully visible in layout stats.
Sectors Affected: Unlike previous waves of automation that hit blue-collar work, this study highlights that the jobs most exposed are in finance, healthcare, and professional services. It specifically notes that entry-level pathways in these fields are collapsing as AI takes over the “junior” tasks (like drafting documents or basic data analysis) that used to train new employees.
Why it is different from previous studies:
Earlier MIT studies (like one from early 2024) focused on economic feasibility (i.e., it might be possible to use AI, but it’s too expensive). This new 2025 study focuses on technical capacity—meaning the AI can do the work right now, and for many of these roles, it is already cost-competitive.
I’ll be honest, that “Iceberg Index” study doesn’t convince me just yet. It’s entirely built off of using LLMs to simulate human beings and the studies they cite to back up the effectiveness of such an approach are in paid journals that I can’t access. I also can’t figure out how exactly they mapped which jobs could be taken over by LLMs other than looking at 13k available “tools” (from MCPs to Zapier to OpenTools) and deciding which of the Bureau of Labor’s 923 listed skills they were capable of covering. Technically, they asked an LLM to look at the tool and decide the skills it covers, but they claim they manually reviewed this LLM’s output so I guess that counts.
Project Iceberg addresses this gap using Large Population Models to simulate the human–AI labor
market, representing 151 million workers as autonomous agents executing over 32,000 skills
across 3,000 counties and interacting with thousands of AI tools
user_prompt_template = "Your age is {age} {gender},{unemployment_rate} the number of COVID cases is {covid_cases}."# Using Langchain to build LLM Agents
agent_profile = "You are a person living in NYC. Given some info about you and your surroundings, decide your willingness to work. Give answer as a single number between 0 and 1, only."
The whole thing perfectly straddles the line between bleeding-edge research and junk science for someone who hasn’t been near academia in 7 years like myself. Most of the procedure looks like they know what they’re doing, but if the entire thing is built on a faulty premise then there’s no guaranteeing any of their results.
In any case, none of the authors for the recent study are listed in that article on the previous study, so this isn’t necessarily a case of MIT as a whole changing it’s tune.
(The recent article also feels like a DOGE-style ploy to curry favor with the current administration and/or AI corporate circuit, but that is a purely vibes-based assessment I have of the tone and language, not a meaningful critique)
I would love to read that study, as going off of your comment I could easily see it being a case of “more than 10% of jobs are bullshit jobs à la David Graeber so having an « AI » do them wouldn’t meaningfully change things” rather than “more than 10% of what can’t be done by previous automation now can be”.
Summarized by Gemini
The study you are referring to was released in late November 2025. It is titled “The Iceberg Index: Measuring Workforce Exposure in the AI Economy.” It was conducted by researchers from MIT and Oak Ridge National Laboratory (ORNL). Here are the key details from the study regarding that “more than ten percent” figure:
https://www.csail.mit.edu/news/rethinking-ais-impact-mit-csail-study-reveals-economic-limits-job-automation?hl=en-US#%3A~%3Atext=This+important+result+commands+a%2Cthe+barriers+are+too+high.”
I’ll be honest, that “Iceberg Index” study doesn’t convince me just yet. It’s entirely built off of using LLMs to simulate human beings and the studies they cite to back up the effectiveness of such an approach are in paid journals that I can’t access. I also can’t figure out how exactly they mapped which jobs could be taken over by LLMs other than looking at 13k available “tools” (from MCPs to Zapier to OpenTools) and deciding which of the Bureau of Labor’s 923 listed skills they were capable of covering. Technically, they asked an LLM to look at the tool and decide the skills it covers, but they claim they manually reviewed this LLM’s output so I guess that counts.
from https://iceberg.mit.edu/report.pdf
Large Population Models is https://arxiv.org/abs/2507.09901 which mostly references https://github.com/AgentTorch/AgentTorch, which gives as an example of use the following:
user_prompt_template = "Your age is {age} {gender},{unemployment_rate} the number of COVID cases is {covid_cases}." # Using Langchain to build LLM Agents agent_profile = "You are a person living in NYC. Given some info about you and your surroundings, decide your willingness to work. Give answer as a single number between 0 and 1, only."The whole thing perfectly straddles the line between bleeding-edge research and junk science for someone who hasn’t been near academia in 7 years like myself. Most of the procedure looks like they know what they’re doing, but if the entire thing is built on a faulty premise then there’s no guaranteeing any of their results.
In any case, none of the authors for the recent study are listed in that article on the previous study, so this isn’t necessarily a case of MIT as a whole changing it’s tune.
(The recent article also feels like a DOGE-style ploy to curry favor with the current administration and/or AI corporate circuit, but that is a purely vibes-based assessment I have of the tone and language, not a meaningful critique)