Last week, Massachusetts Institute of Technology (MIT) published a study claiming that AI is already capable of replacing 11.7% of existing U.S. labor. It’s certainly the kind of eye-popping study guaranteed to get a lot of eyes on researchers’ work at a time of shaky faith in AI, as stockholders might want some reassurance that their AI investments are going to pan out.
The report on this research is called “The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy,” but it also has its own dedicated page called “Project Iceberg” that lives on the MIT website. Compared to the research paper, the project page has a lot more emoji. Where the paper on the study comes across sort of like a warning about AI tech, the project page, which is headlined “Can AI Work with You?” feels more like an ad for AI, in part thanks to text like this:
“AI is transforming work. We have spent years making AIs smart—they can read, write, compose songs, shop for us. But what happens when they interact? When millions of smart AIs work together, intelligence emerges not from individual agents but from the protocols that coordinate them. Project Iceberg explores this new frontier: how AI agents coordinate with each other and humans at scale.”
The titular “Iceberg Index” comes from an AI simulation that uses what the paper called “Large Population Models” that apparently ran on processors housed at the federally funded Oak Ridge National Laboratory, which is affiliated with the Department of Energy.
Legislators and CEOS seem to be the target audience, and they’re meant to use Project Iceberg to “identify exposure hotspots, prioritize training and infrastructure investments, and test interventions before committing billions to implementation.”
The Large Population Model—should we start shortening this to LPM?—claims to be able to digitally track the behavior of 151 million human workers “as autonomous agents” with 32,000 trackable “skills,” along with other factors like geography.
The director of AI Programs at Oak Ridge explained the project to CNBC this way: “Basically, we are creating a digital twin for the U.S. labor market.”
The overall finding, the researchers claim, is that current AI adoption accounts for 2.2% of “labor market wage value,” but that 11.7% of labor is exposed—ostensibly replaceable based on the model’s understanding of what a human can currently do that an AI software widget can also do.
It should be noted that humans in actual jobs constantly work outside their job descriptions, handle exceptional and non-routine situations, and are—for now—uniquely capable of handling many of the social aspects of a given job. It’s not clear how the model accounts for this, although it does note that its findings are correlational not causal, and says “external factors—state investment, infrastructure, regulation—mediate how capability translates to impact.”
However, the paper says, “Policymakers cannot wait for causal evidence of disruption before preparing responses.” In other words, AI is too urgent to get hung up on the limitations of the study, according to the study.
English (US) ·