Correct. The reduced labor share and the switch of pay from high to low skill work do not necessarily mean a reduction in wages… but if the labor share goes down and the population hasn’t changed (or if profits/GDP have not increased absolutely massively), it inherently will mean reduction in wages. More importantly, what Minnitti et al. are identifying (and not just from “AI” in the sense of LLMs but decades of various forms of automation – which makes citing them as a study for AI in the current sense actually even more useless) is, effectively, an increase in inequality and a decrease in skills, and at the scale of huge regions. This is a serious concern (we should be trying to reduce not increase inequality), and the magnitude of that shift from higher and medium skilled to lower skilled workers (obscured by the relatively small magnitude of the general labor share decline) is worth addressing. Now, again, we agree that the concern isn’t automation per se but its implementation and specifics, as well as the broader political climate.

I didn’t say labor share can’t be measured for a nation. I said it doesn’t measure wages or jobs. Obviously the same metric can be aggregated for all companies not just one company. But it’s still the same thing. The rest follows (including the laughable effect size etc.).

Hence there is no evidence here that AI is killing jobs or lowering wages.

No, labor share is quite clearly (and is in Minnitti et al. ) a national metric. To quote Wikipedia: “In economics, the wage share or labor share is the part of national income, or the income of a particular economic sector, allocated to wages (labor). It is related to the capital or profit share, the part of income going to capital,[1] which is also known as the K–Y ratio.[2] The labor share is a key indicator for the distribution of income”. Minnitti clearly discuss national inequality metrics, so that is clearly the context they’re using it in. In their Appendix, they state clearly and repeatedly, “The dependent variable is the labor share, defined as the ratio of employees’ compensation to regional gross value added (at current prices)”. The regional gross value is not a firm-based calculation, it is the economic value of a given region. Regional gross value is how you calculate regional GDP: “It is the aggregate of gross value added (GVA) of all resident producer units in the region, and analogous to national gross domestic product”.

They’re not discussing the internal share of tasks in a firm. And, on the scale of a national economy, a small effect can affect hundreds of thousands of lives. As their conclusion states, “Our findings indicate that AI innovation is associated with a significant decline in the labor share, potentially accounting for up to one third of a percentage point of the overall decrease observed since the early 2000s. This highlights the notable impact of AI in exacerbating income inequality in terms of functional income distribution, particularly in regions more engaged in developing AI-related technologies, underscoring AI’s role in driving regional disparities in labor income distribution”. They’re referring to national macro statistics and their share is since the early 2000s so this effect when taking into account how relatively recent the AI we’re discussing is could be quite serious. (Also, I think that Minnitti et al. are using “AI” to mean computerized automation writ large, which means it’s not very useful for our purposes here, but they don’t have a good operational definition section).

So obviously it’s true that companies may lie about why they fired some workers to exaggerate the effect, but they also may lie when they did in fact fire someone for AI when they didn’t, and they mean fire folks to cover up for losses due to AI. So while that method is not ideal (it’s always hard to get at the internal strategy of companies), I don’t see a reason why it inherently biases one direction rather than the other. And if they add jobs due to AI, great, but that doesn’t help the folks who were laid off. If the result of the system is large-scale localized disruptions (for, again, systems that are of limited value), and then they also expand some enterprises for jobs that are likely to collapse if they realize these systems suck, that’s pretty bad.

And July was actually a really big month for AI. China announced they were going deep into it, AWS and Google had a number of new services, etc. Also, companies could very well have had an internal time (which lines up six months into the year) to determine if they were making redundancies, or that could have been when reportage happened. And I don’t find the idea that they’d blame AI, something that the capital sector is deeply interested in pitching as a total good, over tariffs, something most business have indicated is an issue (even against Trump’s retaliation), all that compelling either. I agree their method can’t sort that out.

As https://fortune.com/2025/08/08/ai-layoffs-jobs-market-shrinks-entry-level/ points out (though it definitely is frustrating that it is only the Challenger study on this topic), “Layoffs are surging in the U.S., with companies announcing more than 806,000 job cuts so far in 2025, the highest figure for that period since 2020, according to Challenger, Gray, & Christmas. The tech sector has been hit the hardest, with over 89,000 layoffs in the industry alone. The firm found that more than 27,000 tech job losses since 2023 have been directly attributed to AI-driven redundancy, as companies streamline operations and restructure departments. At the same time, companies are becoming more selective about who and where they hire. Entry-level roles are feeling the worst of this impact as the technology is increasingly good at automating junior-level work. Many firms are seeing easy cost-cutting opportunities at the entry level. “A lot of entry-level work when you’re fresh out of college is knowledge-intensive jobs where you’re collecting data, transcribing data, and putting together basic visualizations, and learning the organization from the ground up,” Tristan L. Botelho, associate professor of organizational behavior at Yale School of Management, told Fortune. “AI can do that quite well, and I’ve heard many managers say things like: ‘We can reduce our entry-level headcount.’ … The biggest disruption is likely among these low-level employees, particularly where work is predictable, tech-savvy, or more general'”. So that’s corroboration from experts, including McKinsey saying they are actively using AI and using it to make employees redundant. And the fact that the annual job cuts are not only so large but consolidated in tech (which, even given component costs going up from tariffs, is internationally diversified enough and service-centered enough to not be really as affected by tariffs as much more capital-intensive companies are, and also got some exemptions: https://www.ainvest.com/news/political-cost-tariffs-impact-tech-sector-valuations-2508/) I think is indicative of a potential trend. (Of course, tech companies are famous for bullshit layoffs, so that could easily be them just preparing for a lean time and using AI as an excuse, but it could also be them getting sucked into the AI hype, and the number of people in this Fortune article testifying to the latter makes me dubious that it’s all purely dishonest).

That diverted investment could be paying someone if it doesn’t end up being locked into various financial instruments due to our crap monetary velocity. But if it does, it could be going to massive consulting firms, tech firms with huge internal inequality, etc. etc. It is wholly possible for companies to go from investing into something that created hundreds of thousands of jobs to invest into something that only creates thousands. That’s one of the ways you can get medium-term systemic unemployment.

I should also add:

That diverted investment (like every other) is paying someone. Every dollar diverted to investing in AI hardware, software, etc. is paying for jobs somewhere else (the people building the hardware, software, etc., and selling it, delivering it, maintaining it, and managing all those people, and all the support effects, e.g. for every dollar of this, some is going to pay for the guy who fills the coke machine in the AI development office, the janitors, the security guards, etc.).

Even the companies that are growing overhead share for AI acquisitions are also likely hiring or paying existing employees to use it. Which effect is also not being measured by “labor share.”

So “labor share” is a completely useless metric for our purposes here. It tells us nothing useful about the effect of AI on jobs or wages.

measure is labor share, which is compensation paid to workers

Incorrect.

Labor share means the amount of an investment that goes to labor.

For example, a steel factory has a low labor share because most of its overhead is buying steel and coke.

Their study simply says that companies that invest in AI divert capital to hardware, software, etc., which is simply just always the case (every year capital is redirected to new things, like new cash registers, new market development, building new stores, etc.); doesn’t have anything to do with taking money away from labor (where the capital is directed from, or whether new capital is raised for it, is not being measured here); and their effect size (the amount of capital they claim is diverted) is so small as to be literally LOL.

Why bother with all this R&D outside of key markets when you can just have poor people drive?

A concept Andor nailed: actually, prison labor is cheaper than droids.

Sorry, I do not understand the relevance of this remark to the article here or even what point you mean to be making.

But what is a “being”? Here we have the issue of the difference between being and existence. Did Anselm recognize that (in Latin)? Let’s say “maybe. If so, we must assume that God “is” (Descartes’ “sum”) and the world “exists” in these terms.

Interestingly, the sages in Kashmir around the same time as Anselm had a similar understanding. Aphorism 1 of the Shiva Sutras: Chaitonyam Atman. God is Consciousness.

Nothing exists that is not that Consciousness. Anselm would have agreed. Holographic theory would complete the view.

—Jack Hill

Agreed on the problem being the systems. I think that the criticism people are offering is precisely, “AI can’t do the things you guys are saying it is, so you’re going to use it as an excuse to just not do things“. Like with screenplays. They won’t actually use an AI screenplay for the foreseeable future. But what they will do, unless blocked (and I think the writer’s strike did get some assurances in this regard), is say “Here’s an AI script, we’re going for it, now you rewrite it for scale”. For the cost of nothing, you get a script concept that your boardroom literally likes (because they wrote the prompt) and then you save a ton on the screenplay. And I am already seeing clearly AI-generated ads, which is denying work to people who are trying to get writing and acting work in ads. Ads don’t need to be any good on YouTube, they can just be slop. Just like how Uber was an excuse to destroy and worsen taxis using an app. (And why self-driving is really taking a long time: Why bother with all this R&D outside of key markets when you can just have poor people drive?)

And social problems don’t need to rise to the level of measurably impacting aggregate macroeconomics to be a serious issue. If they materially harm the wages of specific sectors while giving nothing of value back to society, that’s an issue. I agree that the apocalyptic concerns are silly (our economy is already so oriented at people doing tedious crap work that is already functionally so cheap that trying to replace it with AI is nonsense and they can’t even meaningfully do it any more than they have – e.g. customer phone lines are already as automated as they can be and the whole point of the phone banks is for when you need to talk to someone who either doesn’t understand automated prompts or has a problem that is outside of your first tier common issues), but the problems can be, again, at the scale of things like Uber and AirBNB… which are serious.

Their measure is labor share, which is compensation paid to workers. They’re detecting that that measure has gone down to high and medium-skilled workers and is offset by increases to low-skilled workers. That’s serious, especially if what is happening (and their data can’t support this but does suggest it as a serious possibility) is that high-skilled work is being transmuted by the systems into low-skilled work. That deskilling is very bad for people.

So I don’t think Minniti et al. have mentioned good enough controls, but at the same time they have a pretty clear measure of labor share and skill ratios, and they’ve demonstrated those have shifted. More importantly, it’s not terribly plausible that, say, tariffs would have the effect of shifting labor down-skill while also expanding the base. In particular, the fact that the effect tracks the doubling of AI investment means that no other variable makes a ton of sense: no other variable is going to be changing by multiples. I think they make a good argument that the only real culprit for what they’re surveying is AI. The effect they are surveying is small, precisely because their method does seem to be extracting technology effects as opposed to others.

(And, on the flipside, their positive, that the low-skill increase may be due to AI, is also only one hypothesis).

And for the Challenger Gray data: The fact that an effect hit in May, spreading outside of sectors which would be affected by DOGE and during a period where tariffs were of oscillating credibility thanks to TACO (and also would have hit different sectors), I think is indicative. I don’t think that’s smoke, not at least some fire. They’re obviously having to make inferences based off of a huge number of variables, but I think that plus the Minniti data is enough to reasonably conclude that some jobs are under threat because of deskilling.

https://futurism.com/ai-impossible-find-job makes a similar analysis from the jobs data for this year that I think is pretty plausible: companies are using AI as a threat (and also using it to justify layoffs they already would have done) and are also worried about money they’ve sunk in which is combining with tariff fears.

And while Challenger Gray obviously has a vested interest, their analysis includes when there have been positive spots or flat trends. They’re not just cherry-picking.

The effects have been around for too short of a period and have been paired with too much political instability to be sure, but again, the fact that it came up in the writer’s strike and is coming up in labor discussions tells me that it’s a reasonable concern. Of what magnitude? Hard to tell yet. I agree that doom and gloom (nor tech-optimism) is not reasonable to justify based on the data we have, but there are quite real concerns.

As for “The capitalists would do it anyways”: While that is broadly true, it is important to bear in mind that capitalists are not completely omnipotent in terms of using deceptive approaches. AI plausibly looks like something that could do what they’re saying, if you squint. In contrast, NFTs were so obviously a bubble even ahead of time that their impact on the economy really was marginal. (Crypto partially escaped containment but still mostly was just the obsession of drug dealers and gambling addicts). So I’d argue that LLMs are pretty perfect for these effects, and without LLMs existing (or, much more importantly, LLMs being deployed differently), other systems may recognize the problem much more effectively and not buy-in. And insofar as companies damage themselves or other companies (which will have knock-on effects – layoffs for example) from the hype, I think that can be blamed entirely on the AI.

Again, I really doubt we’re going to see even something like a 90s Internet bubble for AI bursting. Venture capitalists may lose their shirts in some of these applications, but they’re playing with money they can afford to lose. So the current scale of the problem is, I think, on the level of the gig economy: a worsening of already serious problems.

AI will have no significant direct effect on jobs.

Indeed it’s the change in political-economic regime between 1950-1970 and 2000-2025 that is causing the decline of capitalist society. It is eating itself, just like Marxist societies do. Both worldviews have a vision that sounds nice, but always fails in practice, as control of resources simply gets hoarded by a narrow elite who inevitably mismanage it all and the remaining population becomes increasingly impoverished and stressed until the system collapses altogether.

But AI will have no effect on this story. It is just another tool that will shuffle the board but change no rules or averaged outcomes, and will be used for good and ill as any tool, but in twenty years no one will have any of idea of AI being any more radical or revolutionary than, again, robotics in manufacturing in the 1960s or computers in every industry in the 1980s or smartphones in every hand in the 2000s.

These don’t affect fundamental measures. They do not significantly affect the unemployment rate, the minimum wage, real wages, cost of living, or average household wealth, because those things are all driven by corporate and state decisions. In principle robotic factories and computers and smartphones and AI could improve all those things, but only if corporations and states decide to leverage them that way. And they never do. So I don’t expect they ever will.

And insofar as anyone tries to blame those things for worsening those metrics, it will be a red herring, because it won’t have been those things that did it—remove them causally from the timeline and there would be no change in outcome. Because they’ll just use whatever tools exist anyway. So it will always simply be the decisions of capitalists and their state cronies that make things worse, not the technologies that just by chance happen to exist in any given decade.

So AI panic is a problem because it is overblown and misdirected, distracting people from the real problem—which isn’t AI.

-:-

For example, they say …

Look more closely. Look at what they measured (it isn’t wages or jobs; their chosen metric is irrelevant to those things) and its effect size (almost zero).

The reason its voiced panic is inapt is that they are making claims that don’t exist in the report. The report doesn’t even measure anything they are panicking about, and what it does measure is so insignificant in scale as to be laughable that it was even published.

This is what I mean.

Jobs have also already been cut

No they haven’t. Media is unreliable. So are abstracts. Read the actual report, its actual methods and data, and hence whether its conclusion is even merited by its own data.

It’s not: look at their lead graph—almost all unexpected job cuts are in a single now-past spike when Trump announced tariffs, then cuts returned to the same baseline rate as the last three whole years, and then ticked up only slightly precisely as the August 1 tariff deadline approached—which AI cannot explain, but tariffs 100% entirely explains.

Also look at the effect size: 10k is so few as to be below margin of error for any causal analysis—it’s thus useless data.

Also look at the defective method: the study asked companies to “say” whether they cut jobs due to AI—and self-report of self-interested and dishonest capitalist bosses is never a reliable source of this kind of information—but never asked them to say whether they added any jobs due to AI, so the report was biased at its very methodological instrument to fake net loss numbers for AI, which invalidates all its results (you cannot use this report to argue net losses for any reason, much less AI, because it didn’t measure net).

Both flaws become apparent in the actual report’s Table 4 where inexplicably there were only 75 supposedly-AI-related cuts for the entire year (Jan, Feb, Mar, Apr, May, and June), and 10,300 lost in July. Explain to me how all those jobs were lost just last month alone. AI cannot be the reason. But the final August 1 implementation of Trump’s threatened tariffs does. Which signals a false-reporting effect: it looks like corporations are hiding their real reasons for sudden July layoffs behind something already unpopular (and which the Trump admin won’t scold them for blaming), thus circularly using the panic narrative to bolster that narrative in turn, to their advantage. Supporting that is the fact that these same corporations credited only half as many cuts to tariffs, a wildly implausible result given all other economic reporting on what is causing job loss, which is orders of magnitude more than that.

And so once we catch all these bullshit flags, we should be inspired to ask—who the fuck is “Challenger, Gray & Christmas,” the for-profit corporation writing this report? Oh, right, a company selling jobs placement services. Hm. I wonder why a company like that might want to inflate fears of AI job capture? Why would we trust them? Is the media really this dumb? Yes. Yes they are.

This is what I am talking about.

Leaping to any panic claim anyone makes simply because it exists, rather than checking whether it’s bullshit or not first. There simply is no evidence AI has led to any net loss in jobs, just as self check-out did not. And we need to pay attention to this. We need to wait for real evidence. Not pretend it already exists when it doesn’t.

I wrote a whole article on the importance of not falling victim to these kinds of bad-study-driven false narratives: Three Models of Critical Thinking: Remote Work, Generational Wealth, and Election Polling. We need to remember we are supposed to be critical thinkers.

None of this means AI won’t do anything to be worried about. Rather, the panic narrative so far is not supported by any actual evidence. So we need to not be selling it, any more than we should be selling the converse AI messiah narrative the corporations selling it are, which is just as factless.

[And remember, “labor share” metrics tell us nothing useful about the effect of AI on jobs or wages. See my comment below.]