I wonder if these researchers include their own jobs in the analysis. Because AI can very easily spit out random numbers and a lengthy explanation to make them seem believable.
From the abstract:
"The Index captures technical exposure, where AI can perform occupational tasks, not displacement outcomes or adoption timelines." (emphasis mine)
The 11.7% figures is the modeled reduction in "wage value", which appears to be marketplace value of (human) work.
This is the same group (Ayush Chopra & Ramesh Raskar) that previously published the highly circulated (clickbait) article saying that 95% of AI pilots were failing based on extremely weak study design and questions that didn't even support the takeaways.
Anything coming from Ayush and Ramesh should be highly scrutinized. Ramesh should stick to studying Camera Culture in the Media Lab.
I will believe a study from MIT when it comes out of CSAIL.
You should read the paper (or at least the abstract) before making personal attacks. It makes no claims about job disruption (quite the opposite actually).
The real advantage AI gives is cover to change current processes. There's a million tiny tasks that could be automated and in aggregate would reduce labor needs by making labor more productive.
AI isn't a feature. Spellcheck is a feature. Templates are a feature. Search is a feature. A database of every paywalled article is a feature. AI can't do anything but it gives cover for features that do.
The fact that these very-smart people did not include ranges is absurd.
They know that 11.7% is WAY too precise to report. The truth is it's probably somewhere between 5-15% over the next 20 years and nobody has any idea which side of that range is correct.
Similar precision appears in other exposure studies also. E.g. This one was trending from OpenAI and Wharton a short while back: arxiv.org/pdf/2303.10130
This is like unbelievably awful journalism. From the abstract:
>The Index captures technical exposure, where AI can perform occupational tasks, not displacement outcomes or adoption timelines. Analysis shows that visible AI adoption concentrated in computing and technology (2.2% of wage value, approx $211 billion) represents only the tip of the iceberg. Technical capability extends far below the surface through cognitive automation spanning administrative, financial, and professional services (11.7%, approx $1.2 trillion). [https://arxiv.org/abs/2510.25137]
Does the author not know what displacement outcomes are?
It's possible we got 2.2% better quality software by augmenting engineers.
I expect we'll see at least 11.7% <metrixX> improvements in admin, financial, and professional services.
There is likely also a depressive affect on the labor market - there is nuance here and it would be equally disingenuous to believe there will be zero displacement (although, there is a case for more labor participation is administrative bottlenecks / cost are solved, tbd).
Either way this is like a textbook example of zero-sum minded journalist grossly misrepresenting the world.
Products like v0.dev (and gemini-3 with nano banana in general) continue to get better at building website designs that don't look obviously vibe coded.
Interesting - that’s a 1T market just in the US alone. Probably another 1T in EU. It’s unclear how much there is in the rest of the world (China is basically inaccessible to US firms and after that it’ll depend on low wage local labor vs AI models).
There’s also models getting more capable (larger share of the GDP) and GDP growing more quickly due to automation of GDP activities. But even without that it’s at least a 2T/year opportunity (assuming the model is even a little accurate).
To me this validates the bull case that is being raised in private equity. The major risks are not if the market or valuations exist but whether it’ll be captured by a few major players or if open models and local inference eat away at centralization.
>Beneath the surface lies the total exposure, the $1.2 trillion in wages, and that includes routine functions in human resources, logistics, finance, and office administration. Those are areas sometimes overlooked in automation forecasts.
Those routine functions could have been automated before LLMs.
Usually when theyre not it's due to some sort of corporate dysfunction which is not something LLMs can solve.
Here's a realistic path for how AI "replaces"/"displaces" a large chunk of the workforce:
- Even without AI most corpos could shed probably 10% of their workforce - or maybe more - and still be about as productive as they are now. Bunch of reasons why that's true, but here are two I can easily think of: (1) after the layoffs work shifts to the people who remain, who then work harder; (2) underperformers are often not let go for a long time or ever because their managers don't want to do the legwork (and the layoffs are a good opportunity to force that to happen).
- It's hard for leadership to initiate layoffs, because doing so seems like it'll make the company look weak to investors, customers, etc. So if you really want to cut costs by shedding 10%+ of your workforce and making the remaining 90% work harder, then you have to have a good story to tell for why you are doing it.
- AI makes for a good story. It's a way to achieve what you would have wanted to achieve anyway, while making it seem like you're cutting edge.
Running a shop at maximum productivity is not sustainable of course. Quality and morale suffers, best workers leave, and turns out you really need slack for things to work well. ("You should overprovision your capacity" for the engineer mindset)
But they should also look at the other side of the story. How many new problems will be created by that requires new jobs and investment. Most likely it's migration of jobs from one kind of work to other kind of work.
There's always a lot of bending over backwards in these comments to create explanations for why the invention whose purpose is to replace labor won't replace labor.
Read the project and its key paper before commenting:
arxiv.org/abs/2510.25137
The key takeaway buried between technical jargon is that these figures aren’t measuring workforce replacement, but task replacement. They aren’t saying AI can replace 12% of the workforce, rather that AI can replace 12% of the work performed, and its associated wage values, expected concentrations, and diverse impacts (across the lower 48). There does not seem to be a more user-friendly visual available to tinker with, at least that I could readily find on mobile.
They try to couch this conclusion at the end, stating that workforce displacement isn’t going to happen by AI so much as by decision-makers in government and enterprise. It’s entirely possible to use AI tools to amplify productivity and output and lead to smaller work weeks with better labor outcomes, but we have ample evidence that, barring appropriate carrots and sticks, enterprises will fire folks to keep the profit for themselves while governments will victim-blame the unemployed for “not being current on skills”. This creates a strong disincentive for labor to cooperate with AI, because it’s a lose-lose Prisoner’s Dilemma for them: cooperation will either result in a boost in productivity that hurts those around them through displacement and an increased workload on themselves, or cooperation results in their own replacement in the midst of a difficult job market and broader economy. Cooperation is presently the worst choice for labor, and the authors do a milquetoast job highlighting this reality - but do better than most of their predecessors, at least.
Really, it comes back to what I spoke about in 2023 when it comes to AI: the problem isn’t AI so much as a system that will hand its benefits to those of already immense wealth and means, and that is the problem that needs solving immediately.
If anyone is curious about automation and people's/worker's reaction to it, I recommend Blood in the Machine: The Origins of the Rebellion Against Big Tech by Brian Merchant:
> The most urgent story in modern tech begins not in Silicon Valley but two hundred years ago in rural England, when workers known as the Luddites rose up rather than starve at the hands of factory owners who were using automated machines to erase their livelihoods.
> The Luddites organized guerrilla raids to smash those machines—on punishment of death—and won the support of Lord Byron, enraged the Prince Regent, and inspired the birth of science fiction. This all-but-forgotten class struggle brought nineteenth-century England to its knees.
> Today, technology imperils millions of jobs, robots are crowding factory floors, and artificial intelligence will soon pervade every aspect of our economy. How will this change the way we live? And what can we do about it?
I think the real story isn’t that AI will replace 11.7% of workers. It is that we are about to discover that far more than 11.7% of the work we do was never actually work in the first place.
Workflows that were untouchable will now be overhauled and the productivity gains just raises the throughput ceiling.
The difficulty is in the implementation. Many jobs could already be mostly replaced with just a basic system of record (i.e. a database) but it hasn't happened. The world still runs on paper, email, or maybe a shared spreadsheet if they're sophisticated.
Organizations are glued together with interpersonal relationships and unwritten expertise so it's really hard to just drop in an AI solution - especially if it isn't reliable enough to entirely replace a person because then you need both which is more expensive.
there isn’t a govt on earth that can survive that large & sudden an increase in long-term unemployment; overthrown or bankrupted, they’re gone either way. the pitchfork mob will proceed to start burning data centers. the idea they’ll all quietly choose serfdom over revolution is wildly unrealistic. ai needs much stronger regulation to have a chance at survival.
26 comments
[ 2.9 ms ] story [ 43.0 ms ] threadFrom the abstract: "The Index captures technical exposure, where AI can perform occupational tasks, not displacement outcomes or adoption timelines." (emphasis mine)
The 11.7% figures is the modeled reduction in "wage value", which appears to be marketplace value of (human) work.
Anything coming from Ayush and Ramesh should be highly scrutinized. Ramesh should stick to studying Camera Culture in the Media Lab.
I will believe a study from MIT when it comes out of CSAIL.
The real advantage AI gives is cover to change current processes. There's a million tiny tasks that could be automated and in aggregate would reduce labor needs by making labor more productive.
AI isn't a feature. Spellcheck is a feature. Templates are a feature. Search is a feature. A database of every paywalled article is a feature. AI can't do anything but it gives cover for features that do.
They know that 11.7% is WAY too precise to report. The truth is it's probably somewhere between 5-15% over the next 20 years and nobody has any idea which side of that range is correct.
Similar precision appears in other exposure studies also. E.g. This one was trending from OpenAI and Wharton a short while back: arxiv.org/pdf/2303.10130
>The Index captures technical exposure, where AI can perform occupational tasks, not displacement outcomes or adoption timelines. Analysis shows that visible AI adoption concentrated in computing and technology (2.2% of wage value, approx $211 billion) represents only the tip of the iceberg. Technical capability extends far below the surface through cognitive automation spanning administrative, financial, and professional services (11.7%, approx $1.2 trillion). [https://arxiv.org/abs/2510.25137]
Does the author not know what displacement outcomes are?
It's possible we got 2.2% better quality software by augmenting engineers.
I expect we'll see at least 11.7% <metrixX> improvements in admin, financial, and professional services.
There is likely also a depressive affect on the labor market - there is nuance here and it would be equally disingenuous to believe there will be zero displacement (although, there is a case for more labor participation is administrative bottlenecks / cost are solved, tbd).
Either way this is like a textbook example of zero-sum minded journalist grossly misrepresenting the world.
Products like v0.dev (and gemini-3 with nano banana in general) continue to get better at building website designs that don't look obviously vibe coded.
There’s also models getting more capable (larger share of the GDP) and GDP growing more quickly due to automation of GDP activities. But even without that it’s at least a 2T/year opportunity (assuming the model is even a little accurate).
To me this validates the bull case that is being raised in private equity. The major risks are not if the market or valuations exist but whether it’ll be captured by a few major players or if open models and local inference eat away at centralization.
Those routine functions could have been automated before LLMs.
Usually when theyre not it's due to some sort of corporate dysfunction which is not something LLMs can solve.
- Even without AI most corpos could shed probably 10% of their workforce - or maybe more - and still be about as productive as they are now. Bunch of reasons why that's true, but here are two I can easily think of: (1) after the layoffs work shifts to the people who remain, who then work harder; (2) underperformers are often not let go for a long time or ever because their managers don't want to do the legwork (and the layoffs are a good opportunity to force that to happen).
- It's hard for leadership to initiate layoffs, because doing so seems like it'll make the company look weak to investors, customers, etc. So if you really want to cut costs by shedding 10%+ of your workforce and making the remaining 90% work harder, then you have to have a good story to tell for why you are doing it.
- AI makes for a good story. It's a way to achieve what you would have wanted to achieve anyway, while making it seem like you're cutting edge.
arxiv.org/abs/2510.25137
The key takeaway buried between technical jargon is that these figures aren’t measuring workforce replacement, but task replacement. They aren’t saying AI can replace 12% of the workforce, rather that AI can replace 12% of the work performed, and its associated wage values, expected concentrations, and diverse impacts (across the lower 48). There does not seem to be a more user-friendly visual available to tinker with, at least that I could readily find on mobile.
They try to couch this conclusion at the end, stating that workforce displacement isn’t going to happen by AI so much as by decision-makers in government and enterprise. It’s entirely possible to use AI tools to amplify productivity and output and lead to smaller work weeks with better labor outcomes, but we have ample evidence that, barring appropriate carrots and sticks, enterprises will fire folks to keep the profit for themselves while governments will victim-blame the unemployed for “not being current on skills”. This creates a strong disincentive for labor to cooperate with AI, because it’s a lose-lose Prisoner’s Dilemma for them: cooperation will either result in a boost in productivity that hurts those around them through displacement and an increased workload on themselves, or cooperation results in their own replacement in the midst of a difficult job market and broader economy. Cooperation is presently the worst choice for labor, and the authors do a milquetoast job highlighting this reality - but do better than most of their predecessors, at least.
Really, it comes back to what I spoke about in 2023 when it comes to AI: the problem isn’t AI so much as a system that will hand its benefits to those of already immense wealth and means, and that is the problem that needs solving immediately.
only thing better than pulling numbers out of the air is being very very precise
(not)
> The most urgent story in modern tech begins not in Silicon Valley but two hundred years ago in rural England, when workers known as the Luddites rose up rather than starve at the hands of factory owners who were using automated machines to erase their livelihoods.
> The Luddites organized guerrilla raids to smash those machines—on punishment of death—and won the support of Lord Byron, enraged the Prince Regent, and inspired the birth of science fiction. This all-but-forgotten class struggle brought nineteenth-century England to its knees.
> Today, technology imperils millions of jobs, robots are crowding factory floors, and artificial intelligence will soon pervade every aspect of our economy. How will this change the way we live? And what can we do about it?
* https://www.hachettebookgroup.com/titles/brian-merchant/bloo...
* https://www.bloodinthemachine.com/p/introducing-blood-in-the...
* https://www.goodreads.com/book/show/59801798-blood-in-the-ma...
* https://read.dukeupress.edu/critical-ai/article/doi/10.1215/...
Workflows that were untouchable will now be overhauled and the productivity gains just raises the throughput ceiling.
Organizations are glued together with interpersonal relationships and unwritten expertise so it's really hard to just drop in an AI solution - especially if it isn't reliable enough to entirely replace a person because then you need both which is more expensive.
what we will probably see is AI used to build tools and automations that will optimize/remove these jobs