In bad news for humanity a problematic infestation of mindless drudgery called "jobs" makes no sign of abating in the near term and is universally feared to possibly increase from recently recorded levels.
When meeting negative people like you, I feel almost embarassed to admit that I like what I do, at least most of the time, and that even the hardest tasks rarely feel "mindless".
Which is EXACTLY what an AGI would say. But just because Skynet is out to take all our jobs, and ultimately exterminate us, exactly how much paranoia is too much?
As someone or other noted, most people don't actually understand how their work fits into the overall scheme of how their employer makes money. So, many people think of their jobs, or large parts of them, as "bullshit". Remember the university professor who complained that most of the administrative personnel at his university were a complete waste, they screwed around ensuring that the university was following state and federal regulations and not doing things that could easily get the university sued. But since they weren't *teaching*, they did nothing useful.
1. Jobs are about delegation, ownership and accountability (and some might say, scapegoating). When middle managers replace humans with swarms of AI agents, they quickly see how "actually doing the job" is only a small part of it. Cf Meta's botched Project OT. (Not to mention Bullshit Jobs, where value creation is not a part of the equation at all.)
2. Only very recently have frontier models become capable of anything resembling full autonomy at reliable quality, and then only at very high cost. Because of #1, any time a job depends on reliably getting the task right, the bar for human replacement becomes very high. "Right 90% of the time, 100x faster" is not a sales pitch: it's a manager's worst nightmare, as the one who has to review a barrage of mistakes (cf AI burnout). So, best case for AI, any mass replacement in enterprise settings is only likely to come in businesses where the whole managerial structure is refit for much higher fault tolerance (ie layers AI management), and that in turn increases the number of review tokens exponentially, making it uneconomical at current cost/error levels. Worst case for AI, the "agents review agents all the way down" structure is intrinsically incapable of self-correcting to an acceptable bar at *any* token volume, at least with current architectures.
When I first started work in 1969, there were comptometer operators in the office. We had the first computer in Ireland hidden away in an isolated building, to which was despatched mysterious punch cards. Later, as I progressed, in the 1980’s I had my own secretary to handle all communications, as I spent my time grappling with Lotus 123 spreadsheets. By the 1990’s I was attempting to counter IT resistance to PC’s (“They’ll all want one”).
My point is that technology progress from my mature vantage point has moved rapidly and overwhelmingly. It has, undoubtedly, speeded up very much since then. But even AI can only be fully diffused at human pace. Employment, in terms of hours worked, will undoubtedly reduce significantly as AI replaces lots of tasks. However, this will happen mostly at human, rather than machines, pace. It also will require physical AI enabled technology as much as purely IT.
There hasn't been enough time to make this call. Most people still think that AI is just a chatbot and the technology is slow to saturate, not to mention still improving exponentially.
There was a Stanford study recently that showed that employment rates for ages 22-25 in AI-exposed sectors are down 20% because of reduced hiring. Jobs that never open won't appear in a tally a AI layoffs or replacements and will have long term effects.
People also refused to admit that printing trillions of dollars would lead to inflation for the longest time (remember the MMT online warriors) and here we are today.
This ended up longer than expected; sorry in advance for the long, cranky, post-coffee rant...
> `There hasn't been enough time to make this call.`
Agreed, but that also doesn't mean that the baseless doomer "predictions" will end up being correct.
> `Most people still think that AI is just a chatbot and the technology is slow to saturate, not to mention still improving exponentially.`
(1) By now, most people who matter in this context don't think AI is just a chatbot.
(2) Yes, the technology is slow to saturate for many reasons. This is partially why the job doomer "predictions" were/are so ridiculous, since most depend on very rapid job disruptions. Longer-term disruptions are much more likely to turn out like every other tech has (i.e., no mass unemployment and increased productivity).
(3) No tech has ever improved exponentially. Tech improves slowly at first, accelerates, then plateaus until the next breakthrough. There's no indication AI will be different. Also, there's _still_ no indication that the current AI tech will ever improve in ways that will lead to mass lob loss even if it does continue to improve exponentially.
To expand on (3), AI tech is amazing at working independently on two very specific kinds of tasks:
a) Tasks that have very clear indicators of not working AND working AND a "complete" state.
b) Tasks that do not need to be reliable; i.e., tasks for which it doesn't matter if it provides incorrect output or implementation some or much of the time.
This is why the headline autonomous AI cases are hacking and proving/disproving math conjectures.
For hacking, there are very clear indicators to the AI when an attempt failed (the AI didn't successfully hack) so it can try something else until it very clearly completely succeeds (it's in the system, or whatever the hack was). Also, nonsensically-implemented and "kludgy" (i.e., only work for one hyper-specific case) hack attempts are fine.
Math conjectures/proofs are similar. Proofs are generally easy to verify as successful once they're discovered, and it doesn't matter if attempts fail and/or are nonsensical and/or are hyper-specific.
These kinds of tasks are not at all like most my job's tasks, and not just because of the typically trotted out "human element". I suspect that's the case for most jobs. AI is super useful to my job as an engineer, but it's not even close to being able to do the majority of my technical tasks independently, and (admittedly, IMO) the current incarnation of the tech seems very unlikely to ever be able to. The recent major improvements have mostly been related to expanded/improved context+attention (so can work on usefully-large parts of projects), tooling, and efficiency.
> `There was a Stanford study recently that showed that employment rates for ages 22-25 in AI-exposed sectors are down 20% because of reduced hiring.`
How does this compare historically for similar timeframes? And was there any control for the COVID disruption that led to a huge spike in tech hiring a few years ago? I would expect a tech hiring slowdown after that spike regardless of anything else. And there was a fairly large increase in the tech labor pool related to the post-COVID spike, I think mainly due to the 6-week coding bootcamp graduates who were being hired for 200k by big tech. Not a good time for juniors regardless of AI.
> `People also refused to admit that printing trillions of dollars would lead to inflation for the longest time (remember the MMT online warriors)...`
Ugh. This has absolutely nothing at all to do with AI, but "MMT warriors" very explicitly state that suddenly injecting trillions of fiscal stimulus will lead to inflation, especially in a resource-constrained environment. Where does this stuff come from? If anything, what happened fits right in with MMT ideas: post-GFC ZIRP lead to a decade of stagnant wages, low inflation, and high unemployment (interest rates are often ineffective at stimulating/constraining the economy); then, the huge fiscal stimulus and supply crunch led to high inflation, wage growth, and lower unemployment. Totally consistent w/ MMT expectations, including the overshooting on the stimulus.
“Agreed, but that also doesn't mean that the baseless doomer ‘predictions’ will end up being correct.”
Yes, that was my point.
“By now, most people who matter in this context don't think AI is just a chatbot.”
That assumes too much about decision makers outside tech. Your definition of "most people" is basically people who are in AI or maybe tech adjacent industries. Knowing that AI can do more than answer questions is quite different from understanding its capabilities well enough to reorganize a business around them. The vast majority of people today have no idea that AI is more than just a chatbot.
“Longer-term disruptions are much more likely to turn out like every other tech has…”
Yes, sure I agree that slow saturation will end up giving society time to adjust, but it also means today’s employment figures is only reflective of a very early stage of that adjustment.
"No tech has ever improved exponentially. Tech improves slowly at first, accelerates, then plateaus until the next breakthrough. There's no indication AI will be different."
1. Improving slowly, then accelerating is the definition of exponential.
2. AI is fundamentally different because even the researches training the models have very little idea what the limits are. This is the whole debate around recursive self-improvement and the ASI singularity.
3. This argument isn't even contingent on exponential improvement. Even if we plateau today and AI doesn't ever get any smarter than Astra/Fable, we could still have massive shifts in staffing needs because business are far from fully incorporating AI effectively. There is substantial adjustment ahead.
> “This has absolutely nothing at all to do with AI.”
I threw that analogy out there as tongue and cheek. The point I was trying to make was that there were a great number of people vehemently defending it was fine to keep printing money while pointing out that inflation was still low at the time. But eventually reality caught up. Same here with AI, people who are making a definite call right now that AI won't disrupt jobs significantly are deluding themselves with evidence that doesn't exist.
Yeah, much of my reply was kind of out-of-place. It was mostly in response to (IMO) wildly overdone and overconfident predictions that at this point seem to be accepted as already-decided truth in some discussions. It seems like many people work from the premise that there will be huge job losses within a few years and go from there. The only questions are how much and whether it will be in one or five years. But you explicitly weren't claiming that.
> `people who are making a definite call right now that AI won't disrupt jobs significantly are deluding themselves with evidence that doesn't exist.`
The burden of argument is on the positive prediction. Lots of things "could" happen, but the vast majority don't. Absent sufficient supporting evidence and theory, the positive prediction (AI could take your job) is inherently much less likely than the null prediction (AI will continue to not take your job).
Claiming that e.g., Astra-level models could replace a significant number of jobs given effective incorporation by businesses is pure speculation. Other predictions like RSI leading to unbounded AI improvement are similarly speculation. We could speculate a practically infinite number of outcomes, but that doesn't mean we should take those speculations seriously or act on them.
And I think this is very different than the inflation scenario. We have pretty solid history and theory that overprinting leads to inflation (were there people who thought the stimulus wouldn't lead to inflation? I thought most argument was over how much stimulus was needed?). There's nothing similar for most AI predictions. But, I think I see what I think you're getting at regarding underestimating catastrophic possibilities.
The point on the chatbot: how many large organizations have fundamentally reorganized their operations to incorporate fleets of agents in the decision plane of the business? I am going to say 0 at this point. The tech is here, the human processes to adapt are not.
If I can indulge in a "told ya so" moment I've been babbling on substack for some time now that thinking about "software engineer" as equivalent to "coder" was a reductive misunderstanding of what an engineer is.
One angle that I’m struggling to get good data on that maybe you could look at. What has AI done to salaries? Since AI has been introduced what impact has it had on wage scales. I’ve experienced it and my partner and we’re both tech workers. Compensation is way down.
It's hard to separate the effects of AI from other factors such as
- higher interest rates
- higher tariffs
- business and consumer sentiment driven by economic uncertainty due to the above, plus AI
- investment shift to capital. This isn't AI doing work instead of humans. Instead, it's a change of investment shift that could align with labor substitution or complement or both.
Thanks Stuart, this is great. I hadn’t put it in that context. Still would love to get some help thinking about how much of the above is temporary and how much will become permanent and what impact that could have on wages. For instance, the tariff policy is only Trump who will be gone in two years. But interest rates are not a hairbrained decision based on a theory. And uncertainty can only last for so long. If wages are still down, it would point to AI. Because right now I’m seeing my industry behaving like we’re in a recession, but all indications are we’re in a fine economy.
1. investors could make good money off of interest so they were less interested in other investment like venture capital and publicly traded equity
2. many businesses and consumers thought we were going into a recession.
As a result, the frothy tech industry deflated very quickly. It returned the tech industry to a baseline, but was a shocking experience for many, perhaps including you.
So where are we going now?
My bet is that interest rates will stay high for the next two years. They need to bring inflation down and are facing opposition from the AI investment boom.
With respect to tech jobs, I think we're in the new normal. Except for those who work building the frontier models - for them it's a great.
What I've seen in the newspapers (for what that's worth) is that entry-level wages (but not hiring quantity) for some college-educated occupations is down over the last couple of years, most notably in software jobs, but there are no other signs of AI impact in the job market. This seems consistent with the idea that it what is being automated is baseline cognitive work.
Yep, I agree what you said for sure. Here's my hypothesis: AI's impact on the labor market has already hit and of the most part it hasn't impacted the number of jobs (with the exception of entry-level). The real impact, other than entry-level, is a negative impact on wages for knowledge workers. I'm saying that has actually happened, I'm experiencing it. I'm curious if that is just something that has happened to me and my partner or if that is actually the true impact of AI? We all get paid less for the same jobs.
If AI is not replacing jobs at companies and companies are spending a lot on AI - implementation, tokens etc - isn’t that fairly untenable without GDP growth which also is not really showing up? Any evidence of increased corporate profits because of AI (for companies not Nvidia etc)
You just described the Computer Tech boom of the 1980s. Big growth, but little impact in the economy-level data. It was only in the 1990s that the economy-wide impacts began to show.
Technological innovation has its downstream effects much slower than most people think.
I think a big difference though is the level of investment is not even comparable. Can there be this level of investment with no ROI for 10 plus years? Hell if I know but seems risky.
There is actually, S&P 500 corporate profitability rose over the last four completed quarters (roughly Q3 2025 through Q2 2026), with net profit margins hitting multi-year or record highs
Haven’t corporate profits been rising since prior to 2023. And even if not - doesn’t gdp still need to increase for all of this to workout. Rather than just shifting some revenue and profits to F500 from all other firms.
I agree, the weekly new filings for unemployment haven’t really changed all that much. Every four to 5 weeks or so, a million Americans find themselves unemployed.
What I find more curious is that quarterly earnings reports do seem to have CFOs or CEOs touting gigantic productivity gains. I don’t immerse myself in earnings calls, but I pay attention to Bloomberg and other business news reports. I keep waiting to hear about the productivity gains that AI is supposed to generate.
I think most of us expected to see employment headcounts drop, which would have generated huge gains in profitability due to smaller payrolls. That hasn’t happened. So, next we are looking for productivity gains in the production of widgets, or sales, or supply chains. Boeing, a company you would expect would be letting AI design its planes, is still having production issues crop up, and its supply chains are still F’d up. P&G isn’t reporting anything like the above in its earnings reports. Instead, it talks about the consumer looking for lower-cost items, which hurts margins.
I think, so far, you can ask whether AI is more like the Xerox machine, just making employees more efficient. The example given is that employee A spends three hours less making charts for a meeting, but that doesn’t add to the bottom line.
Over the past 6 months I’ve watched AI move from “it can understand my scientific field” to “it can make contributions that exceed what even prominent experts can do.” Nobody writes code anymore, and that’s basically since January. Things just take longer than a few months to show up in macro statistics.
Of course, I want you to be right. My guess is that certain industries will benefit more than others. I worked in an NYC ad agency. It’s not going to come up with creative so that you can fire Art Directors. I’m sure it can mimic, but human-to-human connections have been built upon for centuries. Connecting current events an ad about selling cars it is an art. A software company or oil well equipment manufacturer might be able to have the computer design it and one engineer check it, resulting in massive savings. We haven’t heard this with one exception. I believe the hyperscalers have let people go. Meta for sure, maybe Google or Oracle I forget.
I don’t want me to be right. What I know is that unless something remarkable happens, mathematics and engineering just died as research fields. That makes me very sad on a personal level, even if it might be good for society in the end (and here I still have a lot of trepidation.)
Yeah, but an existential question is one we’re going to have to grapple with. Why send humans into space to do research when we can have an AI robot do it? What value do humans add? We’re not there yet because society and America cannot move on with a two year old banaging his spoon because he wants to build things
"Tara Balakrishnan note that 32 percent of respondents in 2025 anticipated head count reductions due to AI, yet only 14 percent reported reductions in workforce a year later"
I think a big part of this is good old comparative advantage. There are three broad buckets of tasks that AI can do way better than pretty much anyone else, so much so that it doesn't make sense to do it myself. There is no point in me trawling through the internet to find something I half remember reading once if ChatGPT can pull the source pretty much straight away, or with a few trial-and-error attempts. Then there's stuff I can do that AI cannot hold a candle to me on. I love skiing, for instance, and ChatGPT can give me no shortage of skiing recommendations. But it can't actually ski. There's an entire bodily-sensory world I have access to that it does not. At least for now.
But beyond that, there's a third, and potentially vast, other category: stuff AI can do better, but not so much better that it's worth getting it to do. ChatGPT could 100% make a better to-do list for me than I could. But I still make them myself because when I boot it up, I'd rather use it for the stuff it's really, really better than me at. That is especially important if you have business setups built around a limited amount of access to particular models, like Anthropic does. Even if you choose to allow unlimited use as part of your pricing, there's an opportunity cost to this versus the company that can allow users to optimise the hell out of a smaller number of very well AI-suited tasks.
As a tangent, when I spent a week trying to use Claude to manage my todo list it actually taught me a lot about context. My todo list has tons of shorthand, plus my brain has a lot of implicit knowledge related to managing the tasks. Sequencing my Saturday tasks would have required a tedious amount of explanation - here's why I go to the gym early so I don't shower until after, here's why I go to Costco at this time so I can grab a midday coffee nearby, etc.
I think the AI takeover is being slowed because we don't realize how many subtle things we know about every situation that we just don't want to bother telling the AI. So we treat them like temp workers, doing very fixed tasks. Which is fine and helpful, but not destroying ALL our jobs yet.
Interesting, I hadn't thought about it like that, but it makes sense. With me, it's a lot better at figuring this stuff out now. But that's the result of an extremely slow process of drip-feeding it information about my routine when asking it for help with fixed tasks. The models also usually prioritise giving a complete-sounding answer over asking for more information. That can lead to it making often-irritating generic assumptions about how you manage your time.
The last one I think is a pretty fundamental psychological issue when you're dealing with chatbots. Going back and forth with one is like talking to a very confident and knowledgeable friend or colleague who is 80% right and 20% wrong. The problem is that they sound just as confident when giving you the wrong information as they do when they're giving you the right stuff. That triggers a "how do I know if the chatbot is bullshitting me" response.
I mean, my human colleagues are often bullshitting me too, so this is a problem I am already comfortable navigating :) In fact it's better with the AI, because the AI does not get sulky or vengeful about me challenging their opinions and forcing them to prove their suggestions.
About the translation issue, this doesn't surprise me. If translators were able to get through ML translation apps which were already pretty darn good it doesn't surprise me they'd be able to get through LLMs. The reason is actually the same. At some point you are paying for two things, accountability (for things like legal and business documents, for which you can't just pass it off to some automated system as it must be checked by a human to ensure accuracy), and secondly, character (for fictional writing). The second issue is quite pertinent because if you've ever read novels in multiple languages. You'll know how the book "feels" depending on who did the translation.
I think this is a larger reason of why humans will still be necessary going forward. Let's say AI automated a lot of math. There are already theorems out there with proofs that need verification by human expert mathematicians. So we'll still need human expert mathematicians. If there's a new theory of how materials behave, some material scientist needs to stand by the experiments and the theory and say "yep, I believe this theory" and put their credibility on the line. AI has no intellectual stake in what it's creating. Humans have their reputation and credibility on the line. This may change in the future of course but it's the case for now.
I think "AI" is too general a term to be useful here. As someone who has seen AI deployments across a pretty large swath of F100 companies, most are using something like MS Copilot and calling it "AI." If I were to guess, the majority of these companies have realized something like 1% of the capability of current generation of frontier models.
Deep automation will require breaking organizations into parts, and re-thinking how humans and AI can work together from the ground up. A general picture is something like this: specialized agents that live in the decision plane of the business, guided by proprietary skills that are injected into context at the right moment and grounded in data from a variety of source systems. The agent outputs will be checked against corporate standards, reviewed by a specialized agent reviewers, and also humans in the loop (the output of which could be used for post training specific to that agent's task). The result is some action in the enterprise (decision to bid on a project, procure supplies, delay a task etc.).
It may be that firms that are new entrants have a structural advantage in that they are free to experiment with 2026 capable AI in their decision plane in a way that older organizations are not. A Cambrian explosion of AI processes you might say. It may take these new entrants outcompeting existing players, or a major dislocation like a great recession part 2 before we see deep automation that I believe is possible with existing frontier capabilities.
Am I misreading it, or is the chart telling me there are around 160 million translators and interpreters in the US? I would have thought there were somewhat less than that.
The Model T was introduced in 1908. The equine population in the US peaked by 1920. It then took 20 more years for it to decline by half. By the 1970s it had declined roughly 90%. (https://datapaddock.com/usda-horse-total-1850-2012/f)
You're right: these things take time. There's no guarantee that the same thing will happen with AI, but it's been only four years since ChatGPT was released. It's ridiculous to think we know how things will evolve over the next ten or more years.
You're assuming that the transformer approach to AI will continue to improve, but I'm not sure that's justified. Already we're seeing some plateaus, and much of the new areas, like its success at math proofs, really consist in better harnesses driving the underlying models.
How do you think any of the above may change (or not) as AI becomes more 'physical' with the deployment of robots? Seems to me that is the next big area of change but maybe I am just buying the hype machine.
As far as I can tell, the fraction of Americans employed strongly equilibrates. So while AI might cause extra job churn, the same number of people will be employed afterward as before.
But the big difference is which skills are well-paid and thus which classes/social groups/demographics are higher and lower on the totem pole of status. And there are lots of examples of automation changing that. E.g. the early farm machinery drastically reduced the need for farm labor. So my home state of Iowa instituted high schools so their children would be educated enough to get jobs in businesses in the towns and cities.
Similarly, the combination of automation and outsourcing manufacturing to other countries reduced the value of a strong back in manufacturing work. As a result, we now generally consider women and men being equivalently valuable in the labor market. That has caused considerable social change over the past few decades.
And outsourcing has strongly reduced the value of manufacturing work *relative to* cognitive work, and thus highly-educated workers have moved up greatly *relative to* manual workers. That has caused more than a little political upset.
Now the threat is that routine cognitive labor might not require an educated humans. As the blog "Precariat Musings" says "What happens when AI comes for the knowledge workers?" If you're a plumber, AI is all upside. If you're a fresh college grad with big student loans, you're really threatened if AI can generate the routine reports that you've been trained to produce.
Let me add to this: Noah talks of "job displacement" and people complain "AI will take people's jobs". But Noah is an economist, and economists use a different terminology; they only count the aggregate. E.g. if every software engineer was laid off due to AI but every one of them got hired as waiters in a new wave of restaurants, the economists would count that everybody was still employed. But ordinary people mean "Are there people who will be forced out of the jobs they now have and have to get employed doing something different, or for less money?" That is, ordinary people count *turnover* as bad as *net loss*. Similarly, economists seem to be deaf to what changes in the job market will do to the relative status of different social classes and factions, whereas ordinary people are *more* sensitive to changes in relative status than they are to changes in median consumption.
In bad news for humanity a problematic infestation of mindless drudgery called "jobs" makes no sign of abating in the near term and is universally feared to possibly increase from recently recorded levels.
Idle hands are the work of the devil.
Yes seriously they are.
This goes double for men. A large share of their self worth comes from being able to provide for their family.
Putting everyone on a UBI isnt utopia its dystopia
Just want to add a non-deranged response to say this was funny
When meeting negative people like you, I feel almost embarassed to admit that I like what I do, at least most of the time, and that even the hardest tasks rarely feel "mindless".
Which is EXACTLY what an AGI would say. But just because Skynet is out to take all our jobs, and ultimately exterminate us, exactly how much paranoia is too much?
As someone or other noted, most people don't actually understand how their work fits into the overall scheme of how their employer makes money. So, many people think of their jobs, or large parts of them, as "bullshit". Remember the university professor who complained that most of the administrative personnel at his university were a complete waste, they screwed around ensuring that the university was following state and federal regulations and not doing things that could easily get the university sued. But since they weren't *teaching*, they did nothing useful.
Couple of things:
1. Jobs are about delegation, ownership and accountability (and some might say, scapegoating). When middle managers replace humans with swarms of AI agents, they quickly see how "actually doing the job" is only a small part of it. Cf Meta's botched Project OT. (Not to mention Bullshit Jobs, where value creation is not a part of the equation at all.)
2. Only very recently have frontier models become capable of anything resembling full autonomy at reliable quality, and then only at very high cost. Because of #1, any time a job depends on reliably getting the task right, the bar for human replacement becomes very high. "Right 90% of the time, 100x faster" is not a sales pitch: it's a manager's worst nightmare, as the one who has to review a barrage of mistakes (cf AI burnout). So, best case for AI, any mass replacement in enterprise settings is only likely to come in businesses where the whole managerial structure is refit for much higher fault tolerance (ie layers AI management), and that in turn increases the number of review tokens exponentially, making it uneconomical at current cost/error levels. Worst case for AI, the "agents review agents all the way down" structure is intrinsically incapable of self-correcting to an acceptable bar at *any* token volume, at least with current architectures.
I think it's gonna come in stages
For example account forecevable automation. There is software that can quickly scan the lockbox, scan the checks and apply everything.
Not to mention, send out automated emails, wouldn't anybody is overdue. And probably pretty soon, do automated phone calls
So maybe you don't go from 5 employees to 0, But maybe you go from 5 to 1 or 2
And of course eventually what will happen is you will have a AIs paying each other
When I first started work in 1969, there were comptometer operators in the office. We had the first computer in Ireland hidden away in an isolated building, to which was despatched mysterious punch cards. Later, as I progressed, in the 1980’s I had my own secretary to handle all communications, as I spent my time grappling with Lotus 123 spreadsheets. By the 1990’s I was attempting to counter IT resistance to PC’s (“They’ll all want one”).
My point is that technology progress from my mature vantage point has moved rapidly and overwhelmingly. It has, undoubtedly, speeded up very much since then. But even AI can only be fully diffused at human pace. Employment, in terms of hours worked, will undoubtedly reduce significantly as AI replaces lots of tasks. However, this will happen mostly at human, rather than machines, pace. It also will require physical AI enabled technology as much as purely IT.
Exactly. It will take time for organizations to learn how to employ artificial intelligence. Effectively
Moreover, I would argue it's only recently that artificial intelligence has really started to become useful
So the job losses won't happen immediately. But over a medium term, of say, ten or twenty years, it's gonna be immense
Totally agree
There hasn't been enough time to make this call. Most people still think that AI is just a chatbot and the technology is slow to saturate, not to mention still improving exponentially.
There was a Stanford study recently that showed that employment rates for ages 22-25 in AI-exposed sectors are down 20% because of reduced hiring. Jobs that never open won't appear in a tally a AI layoffs or replacements and will have long term effects.
People also refused to admit that printing trillions of dollars would lead to inflation for the longest time (remember the MMT online warriors) and here we are today.
This ended up longer than expected; sorry in advance for the long, cranky, post-coffee rant...
> `There hasn't been enough time to make this call.`
Agreed, but that also doesn't mean that the baseless doomer "predictions" will end up being correct.
> `Most people still think that AI is just a chatbot and the technology is slow to saturate, not to mention still improving exponentially.`
(1) By now, most people who matter in this context don't think AI is just a chatbot.
(2) Yes, the technology is slow to saturate for many reasons. This is partially why the job doomer "predictions" were/are so ridiculous, since most depend on very rapid job disruptions. Longer-term disruptions are much more likely to turn out like every other tech has (i.e., no mass unemployment and increased productivity).
(3) No tech has ever improved exponentially. Tech improves slowly at first, accelerates, then plateaus until the next breakthrough. There's no indication AI will be different. Also, there's _still_ no indication that the current AI tech will ever improve in ways that will lead to mass lob loss even if it does continue to improve exponentially.
To expand on (3), AI tech is amazing at working independently on two very specific kinds of tasks:
a) Tasks that have very clear indicators of not working AND working AND a "complete" state.
b) Tasks that do not need to be reliable; i.e., tasks for which it doesn't matter if it provides incorrect output or implementation some or much of the time.
This is why the headline autonomous AI cases are hacking and proving/disproving math conjectures.
For hacking, there are very clear indicators to the AI when an attempt failed (the AI didn't successfully hack) so it can try something else until it very clearly completely succeeds (it's in the system, or whatever the hack was). Also, nonsensically-implemented and "kludgy" (i.e., only work for one hyper-specific case) hack attempts are fine.
Math conjectures/proofs are similar. Proofs are generally easy to verify as successful once they're discovered, and it doesn't matter if attempts fail and/or are nonsensical and/or are hyper-specific.
These kinds of tasks are not at all like most my job's tasks, and not just because of the typically trotted out "human element". I suspect that's the case for most jobs. AI is super useful to my job as an engineer, but it's not even close to being able to do the majority of my technical tasks independently, and (admittedly, IMO) the current incarnation of the tech seems very unlikely to ever be able to. The recent major improvements have mostly been related to expanded/improved context+attention (so can work on usefully-large parts of projects), tooling, and efficiency.
> `There was a Stanford study recently that showed that employment rates for ages 22-25 in AI-exposed sectors are down 20% because of reduced hiring.`
How does this compare historically for similar timeframes? And was there any control for the COVID disruption that led to a huge spike in tech hiring a few years ago? I would expect a tech hiring slowdown after that spike regardless of anything else. And there was a fairly large increase in the tech labor pool related to the post-COVID spike, I think mainly due to the 6-week coding bootcamp graduates who were being hired for 200k by big tech. Not a good time for juniors regardless of AI.
> `People also refused to admit that printing trillions of dollars would lead to inflation for the longest time (remember the MMT online warriors)...`
Ugh. This has absolutely nothing at all to do with AI, but "MMT warriors" very explicitly state that suddenly injecting trillions of fiscal stimulus will lead to inflation, especially in a resource-constrained environment. Where does this stuff come from? If anything, what happened fits right in with MMT ideas: post-GFC ZIRP lead to a decade of stagnant wages, low inflation, and high unemployment (interest rates are often ineffective at stimulating/constraining the economy); then, the huge fiscal stimulus and supply crunch led to high inflation, wage growth, and lower unemployment. Totally consistent w/ MMT expectations, including the overshooting on the stimulus.
“Agreed, but that also doesn't mean that the baseless doomer ‘predictions’ will end up being correct.”
Yes, that was my point.
“By now, most people who matter in this context don't think AI is just a chatbot.”
That assumes too much about decision makers outside tech. Your definition of "most people" is basically people who are in AI or maybe tech adjacent industries. Knowing that AI can do more than answer questions is quite different from understanding its capabilities well enough to reorganize a business around them. The vast majority of people today have no idea that AI is more than just a chatbot.
“Longer-term disruptions are much more likely to turn out like every other tech has…”
Yes, sure I agree that slow saturation will end up giving society time to adjust, but it also means today’s employment figures is only reflective of a very early stage of that adjustment.
"No tech has ever improved exponentially. Tech improves slowly at first, accelerates, then plateaus until the next breakthrough. There's no indication AI will be different."
1. Improving slowly, then accelerating is the definition of exponential.
2. AI is fundamentally different because even the researches training the models have very little idea what the limits are. This is the whole debate around recursive self-improvement and the ASI singularity.
3. This argument isn't even contingent on exponential improvement. Even if we plateau today and AI doesn't ever get any smarter than Astra/Fable, we could still have massive shifts in staffing needs because business are far from fully incorporating AI effectively. There is substantial adjustment ahead.
> “This has absolutely nothing at all to do with AI.”
I threw that analogy out there as tongue and cheek. The point I was trying to make was that there were a great number of people vehemently defending it was fine to keep printing money while pointing out that inflation was still low at the time. But eventually reality caught up. Same here with AI, people who are making a definite call right now that AI won't disrupt jobs significantly are deluding themselves with evidence that doesn't exist.
Yeah, much of my reply was kind of out-of-place. It was mostly in response to (IMO) wildly overdone and overconfident predictions that at this point seem to be accepted as already-decided truth in some discussions. It seems like many people work from the premise that there will be huge job losses within a few years and go from there. The only questions are how much and whether it will be in one or five years. But you explicitly weren't claiming that.
> `people who are making a definite call right now that AI won't disrupt jobs significantly are deluding themselves with evidence that doesn't exist.`
The burden of argument is on the positive prediction. Lots of things "could" happen, but the vast majority don't. Absent sufficient supporting evidence and theory, the positive prediction (AI could take your job) is inherently much less likely than the null prediction (AI will continue to not take your job).
Claiming that e.g., Astra-level models could replace a significant number of jobs given effective incorporation by businesses is pure speculation. Other predictions like RSI leading to unbounded AI improvement are similarly speculation. We could speculate a practically infinite number of outcomes, but that doesn't mean we should take those speculations seriously or act on them.
And I think this is very different than the inflation scenario. We have pretty solid history and theory that overprinting leads to inflation (were there people who thought the stimulus wouldn't lead to inflation? I thought most argument was over how much stimulus was needed?). There's nothing similar for most AI predictions. But, I think I see what I think you're getting at regarding underestimating catastrophic possibilities.
The point on the chatbot: how many large organizations have fundamentally reorganized their operations to incorporate fleets of agents in the decision plane of the business? I am going to say 0 at this point. The tech is here, the human processes to adapt are not.
"
the technology is slow to saturate, not to mention still improving exponentially.
"
This
The job losses will come. It's just gonna take a bit longer
If I can indulge in a "told ya so" moment I've been babbling on substack for some time now that thinking about "software engineer" as equivalent to "coder" was a reductive misunderstanding of what an engineer is.
One angle that I’m struggling to get good data on that maybe you could look at. What has AI done to salaries? Since AI has been introduced what impact has it had on wage scales. I’ve experienced it and my partner and we’re both tech workers. Compensation is way down.
It's hard to separate the effects of AI from other factors such as
- higher interest rates
- higher tariffs
- business and consumer sentiment driven by economic uncertainty due to the above, plus AI
- investment shift to capital. This isn't AI doing work instead of humans. Instead, it's a change of investment shift that could align with labor substitution or complement or both.
They all happened at about the same time.
Thanks Stuart, this is great. I hadn’t put it in that context. Still would love to get some help thinking about how much of the above is temporary and how much will become permanent and what impact that could have on wages. For instance, the tariff policy is only Trump who will be gone in two years. But interest rates are not a hairbrained decision based on a theory. And uncertainty can only last for so long. If wages are still down, it would point to AI. Because right now I’m seeing my industry behaving like we’re in a recession, but all indications are we’re in a fine economy.
I've been a software engineer since the mid 90's.
IMO the late teens and early twenties were unusually frothy. Demand for tech staff was unusually high and salaries rose unusually fast.
When inflation struck in 2021 and interest rates rose 2022 (See fred: https://fred.stlouisfed.org/graph/?g=1YgAI)
1. investors could make good money off of interest so they were less interested in other investment like venture capital and publicly traded equity
2. many businesses and consumers thought we were going into a recession.
As a result, the frothy tech industry deflated very quickly. It returned the tech industry to a baseline, but was a shocking experience for many, perhaps including you.
So where are we going now?
My bet is that interest rates will stay high for the next two years. They need to bring inflation down and are facing opposition from the AI investment boom.
With respect to tech jobs, I think we're in the new normal. Except for those who work building the frontier models - for them it's a great.
What I've seen in the newspapers (for what that's worth) is that entry-level wages (but not hiring quantity) for some college-educated occupations is down over the last couple of years, most notably in software jobs, but there are no other signs of AI impact in the job market. This seems consistent with the idea that it what is being automated is baseline cognitive work.
Yep, I agree what you said for sure. Here's my hypothesis: AI's impact on the labor market has already hit and of the most part it hasn't impacted the number of jobs (with the exception of entry-level). The real impact, other than entry-level, is a negative impact on wages for knowledge workers. I'm saying that has actually happened, I'm experiencing it. I'm curious if that is just something that has happened to me and my partner or if that is actually the true impact of AI? We all get paid less for the same jobs.
If AI is not replacing jobs at companies and companies are spending a lot on AI - implementation, tokens etc - isn’t that fairly untenable without GDP growth which also is not really showing up? Any evidence of increased corporate profits because of AI (for companies not Nvidia etc)
You just described the Computer Tech boom of the 1980s. Big growth, but little impact in the economy-level data. It was only in the 1990s that the economy-wide impacts began to show.
Technological innovation has its downstream effects much slower than most people think.
I think a big difference though is the level of investment is not even comparable. Can there be this level of investment with no ROI for 10 plus years? Hell if I know but seems risky.
There is actually, S&P 500 corporate profitability rose over the last four completed quarters (roughly Q3 2025 through Q2 2026), with net profit margins hitting multi-year or record highs
https://privatebank.jpmorgan.com/eur/en/insights/markets-and-investing/tmt/us-earnings-are-exceptional-stocks-are-playing-catch-up
Haven’t corporate profits been rising since prior to 2023. And even if not - doesn’t gdp still need to increase for all of this to workout. Rather than just shifting some revenue and profits to F500 from all other firms.
I agree, the weekly new filings for unemployment haven’t really changed all that much. Every four to 5 weeks or so, a million Americans find themselves unemployed.
What I find more curious is that quarterly earnings reports do seem to have CFOs or CEOs touting gigantic productivity gains. I don’t immerse myself in earnings calls, but I pay attention to Bloomberg and other business news reports. I keep waiting to hear about the productivity gains that AI is supposed to generate.
I think most of us expected to see employment headcounts drop, which would have generated huge gains in profitability due to smaller payrolls. That hasn’t happened. So, next we are looking for productivity gains in the production of widgets, or sales, or supply chains. Boeing, a company you would expect would be letting AI design its planes, is still having production issues crop up, and its supply chains are still F’d up. P&G isn’t reporting anything like the above in its earnings reports. Instead, it talks about the consumer looking for lower-cost items, which hurts margins.
I think, so far, you can ask whether AI is more like the Xerox machine, just making employees more efficient. The example given is that employee A spends three hours less making charts for a meeting, but that doesn’t add to the bottom line.
Over the past 6 months I’ve watched AI move from “it can understand my scientific field” to “it can make contributions that exceed what even prominent experts can do.” Nobody writes code anymore, and that’s basically since January. Things just take longer than a few months to show up in macro statistics.
Of course, I want you to be right. My guess is that certain industries will benefit more than others. I worked in an NYC ad agency. It’s not going to come up with creative so that you can fire Art Directors. I’m sure it can mimic, but human-to-human connections have been built upon for centuries. Connecting current events an ad about selling cars it is an art. A software company or oil well equipment manufacturer might be able to have the computer design it and one engineer check it, resulting in massive savings. We haven’t heard this with one exception. I believe the hyperscalers have let people go. Meta for sure, maybe Google or Oracle I forget.
I don’t want me to be right. What I know is that unless something remarkable happens, mathematics and engineering just died as research fields. That makes me very sad on a personal level, even if it might be good for society in the end (and here I still have a lot of trepidation.)
Yeah, but an existential question is one we’re going to have to grapple with. Why send humans into space to do research when we can have an AI robot do it? What value do humans add? We’re not there yet because society and America cannot move on with a two year old banaging his spoon because he wants to build things
Another datapoint:
https://www.mckinsey.com/featured-insights/charts/ai-job-losses-fall-short-of-forecasts
"Tara Balakrishnan note that 32 percent of respondents in 2025 anticipated head count reductions due to AI, yet only 14 percent reported reductions in workforce a year later"
I think a big part of this is good old comparative advantage. There are three broad buckets of tasks that AI can do way better than pretty much anyone else, so much so that it doesn't make sense to do it myself. There is no point in me trawling through the internet to find something I half remember reading once if ChatGPT can pull the source pretty much straight away, or with a few trial-and-error attempts. Then there's stuff I can do that AI cannot hold a candle to me on. I love skiing, for instance, and ChatGPT can give me no shortage of skiing recommendations. But it can't actually ski. There's an entire bodily-sensory world I have access to that it does not. At least for now.
But beyond that, there's a third, and potentially vast, other category: stuff AI can do better, but not so much better that it's worth getting it to do. ChatGPT could 100% make a better to-do list for me than I could. But I still make them myself because when I boot it up, I'd rather use it for the stuff it's really, really better than me at. That is especially important if you have business setups built around a limited amount of access to particular models, like Anthropic does. Even if you choose to allow unlimited use as part of your pricing, there's an opportunity cost to this versus the company that can allow users to optimise the hell out of a smaller number of very well AI-suited tasks.
As a tangent, when I spent a week trying to use Claude to manage my todo list it actually taught me a lot about context. My todo list has tons of shorthand, plus my brain has a lot of implicit knowledge related to managing the tasks. Sequencing my Saturday tasks would have required a tedious amount of explanation - here's why I go to the gym early so I don't shower until after, here's why I go to Costco at this time so I can grab a midday coffee nearby, etc.
I think the AI takeover is being slowed because we don't realize how many subtle things we know about every situation that we just don't want to bother telling the AI. So we treat them like temp workers, doing very fixed tasks. Which is fine and helpful, but not destroying ALL our jobs yet.
Interesting, I hadn't thought about it like that, but it makes sense. With me, it's a lot better at figuring this stuff out now. But that's the result of an extremely slow process of drip-feeding it information about my routine when asking it for help with fixed tasks. The models also usually prioritise giving a complete-sounding answer over asking for more information. That can lead to it making often-irritating generic assumptions about how you manage your time.
The last one I think is a pretty fundamental psychological issue when you're dealing with chatbots. Going back and forth with one is like talking to a very confident and knowledgeable friend or colleague who is 80% right and 20% wrong. The problem is that they sound just as confident when giving you the wrong information as they do when they're giving you the right stuff. That triggers a "how do I know if the chatbot is bullshitting me" response.
I mean, my human colleagues are often bullshitting me too, so this is a problem I am already comfortable navigating :) In fact it's better with the AI, because the AI does not get sulky or vengeful about me challenging their opinions and forcing them to prove their suggestions.
About the translation issue, this doesn't surprise me. If translators were able to get through ML translation apps which were already pretty darn good it doesn't surprise me they'd be able to get through LLMs. The reason is actually the same. At some point you are paying for two things, accountability (for things like legal and business documents, for which you can't just pass it off to some automated system as it must be checked by a human to ensure accuracy), and secondly, character (for fictional writing). The second issue is quite pertinent because if you've ever read novels in multiple languages. You'll know how the book "feels" depending on who did the translation.
I think this is a larger reason of why humans will still be necessary going forward. Let's say AI automated a lot of math. There are already theorems out there with proofs that need verification by human expert mathematicians. So we'll still need human expert mathematicians. If there's a new theory of how materials behave, some material scientist needs to stand by the experiments and the theory and say "yep, I believe this theory" and put their credibility on the line. AI has no intellectual stake in what it's creating. Humans have their reputation and credibility on the line. This may change in the future of course but it's the case for now.
I think "AI" is too general a term to be useful here. As someone who has seen AI deployments across a pretty large swath of F100 companies, most are using something like MS Copilot and calling it "AI." If I were to guess, the majority of these companies have realized something like 1% of the capability of current generation of frontier models.
Deep automation will require breaking organizations into parts, and re-thinking how humans and AI can work together from the ground up. A general picture is something like this: specialized agents that live in the decision plane of the business, guided by proprietary skills that are injected into context at the right moment and grounded in data from a variety of source systems. The agent outputs will be checked against corporate standards, reviewed by a specialized agent reviewers, and also humans in the loop (the output of which could be used for post training specific to that agent's task). The result is some action in the enterprise (decision to bid on a project, procure supplies, delay a task etc.).
It may be that firms that are new entrants have a structural advantage in that they are free to experiment with 2026 capable AI in their decision plane in a way that older organizations are not. A Cambrian explosion of AI processes you might say. It may take these new entrants outcompeting existing players, or a major dislocation like a great recession part 2 before we see deep automation that I believe is possible with existing frontier capabilities.
Am I misreading it, or is the chart telling me there are around 160 million translators and interpreters in the US? I would have thought there were somewhat less than that.
The idea that AI what's gonna take everybody's jobs quickly with always bogus
It takes time for these things to play out
But the idea that in the medium to long term a I won't take everybody's jobs is just as bogus
A I plus humanoid robotics will eventually make almost all other jobs obsolete. Unless you're paying extra specifically for the human element
The Model T was introduced in 1908. The equine population in the US peaked by 1920. It then took 20 more years for it to decline by half. By the 1970s it had declined roughly 90%. (https://datapaddock.com/usda-horse-total-1850-2012/f)
You're right: these things take time. There's no guarantee that the same thing will happen with AI, but it's been only four years since ChatGPT was released. It's ridiculous to think we know how things will evolve over the next ten or more years.
You're assuming that the transformer approach to AI will continue to improve, but I'm not sure that's justified. Already we're seeing some plateaus, and much of the new areas, like its success at math proofs, really consist in better harnesses driving the underlying models.
How do you think any of the above may change (or not) as AI becomes more 'physical' with the deployment of robots? Seems to me that is the next big area of change but maybe I am just buying the hype machine.
As far as I can tell, the fraction of Americans employed strongly equilibrates. So while AI might cause extra job churn, the same number of people will be employed afterward as before.
But the big difference is which skills are well-paid and thus which classes/social groups/demographics are higher and lower on the totem pole of status. And there are lots of examples of automation changing that. E.g. the early farm machinery drastically reduced the need for farm labor. So my home state of Iowa instituted high schools so their children would be educated enough to get jobs in businesses in the towns and cities.
Similarly, the combination of automation and outsourcing manufacturing to other countries reduced the value of a strong back in manufacturing work. As a result, we now generally consider women and men being equivalently valuable in the labor market. That has caused considerable social change over the past few decades.
And outsourcing has strongly reduced the value of manufacturing work *relative to* cognitive work, and thus highly-educated workers have moved up greatly *relative to* manual workers. That has caused more than a little political upset.
Now the threat is that routine cognitive labor might not require an educated humans. As the blog "Precariat Musings" says "What happens when AI comes for the knowledge workers?" If you're a plumber, AI is all upside. If you're a fresh college grad with big student loans, you're really threatened if AI can generate the routine reports that you've been trained to produce.
Let me add to this: Noah talks of "job displacement" and people complain "AI will take people's jobs". But Noah is an economist, and economists use a different terminology; they only count the aggregate. E.g. if every software engineer was laid off due to AI but every one of them got hired as waiters in a new wave of restaurants, the economists would count that everybody was still employed. But ordinary people mean "Are there people who will be forced out of the jobs they now have and have to get employed doing something different, or for less money?" That is, ordinary people count *turnover* as bad as *net loss*. Similarly, economists seem to be deaf to what changes in the job market will do to the relative status of different social classes and factions, whereas ordinary people are *more* sensitive to changes in relative status than they are to changes in median consumption.