Yves here. Please welcome Goran Lazarevski, who sent us an article of his on LinkedIn which he then revamped and expanded for publication here. Lazarevski, an AI industry participant, takes issue with the cheery views of economists Lukas Althoff and Hugo Reichardt on the labor market effects of AI. As Lazarevski summarizes and you can read in full in Task-Specific Technical Change and Comparative Advantage. The authors come to the astonishing conclusion that AI will both substantially increase wages and reduce inequality. Lazarevski unpacks the flaws in their model and presents an alternative view, of likely labor market effects using a more perspective of how they operate.
Note that both the Althoff/Reichardt article and Lazarevski’s forecast assume that AI will be widely employed by businesses and will take over many tasks performed now. Some business commentators suggest that many of the announced corporate headcount cuts that cite AI adoption as a major driver are Wall-Street-pleasing exaggerations, that most of these reductions are to roll back Covid-era overhiring or otherwise rationalize their workforces. More and more reports of AI not resulting in cost reductions and unduly high and perhaps rising levels of AI errors may also dent the current inevitable seeming trajectory to pervasive use. Nevertheless, Lazarevski’s forecast is consistent with what our technology kingpins are trying to achieve.
By Goran Lazarevski, an economist currently working in Paris as AI Solution Design Lead at Pfizer
Over the past two centuries, technological revolutions have continuously reshaped the labor market, redefining the division of labor between humans and machines. Each wave of innovation, from industrial mechanization to digital automation, has altered not just productivity but the social fabric of work itself, creating new winners and losers and shifting the balance between capital and labor. Today, generative AI represents a new inflection point in this long evolution: it touches the cognitive and creative domains once thought uniquely human, threatening to fracture the labor market into sharply diverging strata.
A new economic paper by Althoff and Reichardt has been making waves in the press this week. It estimates that GenAI implementation in the economy will result in average wage increases of 21% and will substantially narrow wage inequality—the opposite of what most people fear.
So should we believe economists, given their long history of confidently modeling golden futures that reality politely ignores? As Joan Robinson said, the purpose of studying economists is to not be fooled be economists. In what follows, I explain how the authors’ conclusions rest on fragile assumptions within the neoclassical framework and offer an alternative more realistic way of thinking about AI’s effect on the labor market.
The model elegantly formalizes the dynamics from a microeconomic perspective. It defines occupations as bundles of tasks requiring different skills to be completed. Workers are modeled as forward‑looking, optimizing agents endowed with innate skills who choose occupations and can accumulate further skills while on the job. AI can affect production in three ways:
- Automation: AI replaces human labor entirely because it’s cheaper.
- Augmentation: AI boosts worker productivity.
- Simplification: AI lowers the skill requirement for a task.
The simplification channel is the authors’ novel contribution to the literature, and it fully drives the anticipated reduction in wage inequality as it empowers lower-skilled workers to compete for more jobs.
The authors then calibrate the model using historical data in order to estimate the effect of a hypothetical full implementation of 2024 frontier LLM capabilities in the US economy. They get remarkably optimistic conclusions, implying that workers can now look forward to their future under the AI overlords.
But hidden beneath the model’s impressive technical architecture lie standard neoclassical assumptions, which if relaxed could completely reverse the model’s results. Labor markets are treated as perfectly competitive, so wages for each occupation and skill are set to that worker’s marginal product contribution. As a consequence, an AI-induced productivity boost automatically raises higher wages, even in those occupations where full AI automation of routine tasks displaces large number of workers who are then seamlessly reassigned to other tasks where they’d be more productive thanks to AI augmentation and task simplification. Remember, in the neoclassical utopia that mainstream economists inhabit nobody gets fired because perfectly competitive markets assume away unemployment – contrary to recent evidence from AI-exposed occupations experiencing declines in employment. In this neoclassical world, supply creates its own demand (Say’s law), as profits automatically turn into investment and output expands to the exact level needed to accommodate these displaced workers.
In reality, wages are determined by bargaining, not perfect market clearing. In those negotiations between employer and employee, it is most often the employer that has the upper hand due to his monopsony position and/or informational leverage. Employers can even leverage AI to obtain such leverage, also known as AI monopsony. Drawing from the work of Kalecki, the productivity gains from AI-augmented labor will be partially captured by the employer, at a rate that is greater for low-skill occupations than high-skill ones (due to skill scarcity translating into bargaining leverage). This channel can totally flip the predicted effect on wage inequality – turning wage convergence into widening inequality.
Moreover, when a task is fully automated or even simplified to become accessible to less skilled workers, there would now be a larger pool of less specialized workers with a more limited skill set, intensifying competition for jobs and further eroding their bargaining power. This happens because larger labor pools worsen outside options, reduce union threat points and heighten intra-worker competition, shrinking workers’ surplus share. Wages can stagnate or fall despite productivity rising. The resulting pressure on wages and layoffs lowers aggregate consumption, potentially triggering a recession even as profits skyrocket and investment remains strong. The most recent example was the 2001 dot-com recession with similar technological driving factors as today. Studies by Autor and Acemoglu show job polarization trends predating GenAI, but are likely to intensify under its diffusion.
In the post-Keynesian framework described above, multiple forces interact in opposing directions, leaving no straightforward conclusion about their overall effect on wages. Nonetheless, the system’s core dynamics remain driven by skill scarcity. AI primarily automates mid-level cognitive tasks, sharply reducing the demand for those underlying cognitive abilities once central to white-collar employment. What remains are roles demanding advanced interpersonal capacities, high-order reasoning, and practical or technical skills anchored in the physical world (at least until the robots come for those too).
This creates a structural challenge. Our education system is still designed to produce exactly the mid-level cognitive competencies that tools like ChatGPT now provide instantly and at almost no cost. While the school-to-college pipeline aspires to develop higher-order cognitive and creative skills, the question is how many graduates can truly reach that level, especially when the bar rapidly rises due to model advancement. If this trajectory holds, the labor market will fragment into four broad strata:
- Cognitive professionals capable of building and critically evaluating AI automation systems and their outputs;
- Skilled trade and technical workers, performing embodied, locally bound tasks;
- Creative workers talented enough to market their human authenticity;
- Everyone else, including a wide swath of displaced cognitive workers (with a potential differentiation for strong interpersonal communicators in sales, caregiving, social work etc.).
All of a sudden, former analysts, marketers, and technical writers will all be competing for the same type of menial jobs requiring only baseline cognitive skills. (This doesn’t mean these jobs will cease to exist, but there will be a lot fewer of them and they will be technically undemanding.) In this scenario, the bargaining power and wages of the workers saddled in this group will collapse, while top-tier cognitive professionals will see their compensation surge. Today’s weak unions, gig contracts, and corporate monopsony power tilt the scales even further towards inequality, letting firms pocket most AI gains while sidelined workers get scraps. This completely inverts the paper’s conclusion that AI will reduce wage inequality, and this, unfortunately, seems like the more plausible future. The question is not whether AI can make the economy bigger, but who will have the leverage to claim the new wealth it creates.
24 comments
Comments are closed.
I received a “task” in my position in my non business church related activity.
I was forwarded an e-mail to prepare and submit a report on a charitable activity our group had performed.
The e-mail(s) was abused by a gemini-AI task list.
It reminded me of how “suspenses” were given out over 50 years ago when the boss had a secretary.
Not only was the list sophomoric, the person who used to do that was eliminated as a job 50 years ago!
AI would take credit for “efficiencies” of the 1970’s.
In project management I do not see AI making PERT/CPM or what they call it today (earned value) any less “labor” intense than when it was hosted on Honeywell 6000 early mainframes!
Not considering the hallucinating and the boss not able to “talk” to the AI.
Hugely capital intense, ill conceived evolution of IT.
The obsession with microeconomic foundations for models was a mistake. The tripartite division also is a bit silly, Cory Doctorow’s “Centaur vs Reverse Centaur” seems better (does the machine serve you or do you serve the machine?)
For how much longer will the ghosts of Walras and Marshall make economics useless?
Technically it’s not that people don’t get fired, it’s that once the marginal disutility of labor goes beyond the utility of working, people just choose to laze around instead of working. Even more realistic (sarc).
I think this piece is a bit more realistic about the matter. Translators, copy writers and the like are getting reduced to utter poverty and frankly perhaps starvation. Not quite the rosy outcome predicted.
All in all the original article seems dreadful, but Lazarevski’s critique is very welcome. I think Zitron has been on the money about the AI stuff; his most recent (premium, so I only read the introduction) post on the topic about it most likely being much worse than the Dotcom bubble seems more apt. If he is right, then economistic analyses about how workers might or might not benefit from AI seem besides the point.
I think of all the enshittification people have to overlook to say any of this is working.
Ocypode: Cory Doctorow’s “Centaur vs Reverse Centaur” seems better (does the machine serve you or do you serve the machine?)
Yes, Doctorow’s framing is better, because — though simplified — it still centers the hybrid nature of potential human-AI relations.
Ocypode: Translators, copy writers and the like are getting reduced to utter poverty and frankly perhaps starvation.
Yes. And see this in today’s FT. No archived link yet, but: –
The great graduate job drought: Economic uncertainty and the arrival of AI have brought a reduction in entry-level roles, with potentially disastrous consequences for young people
https://www.ft.com/content/c89496b1-bc8d-425e-b86b-ec89402410e4
‘When 22-year-old Emily Chong graduated from University College London last year, she thought the job hunt would be simple … After a recent rejection at one large company, she was told she was battling it out against 3,300 applicants for early-career roles….Executives increasingly argue that AI can absorb much of the workload once assigned to early-career staff or to office-based roles such as marketing and communications — where Chong is looking for a job — and customer service.’
‘…the classic corporate pyramid structure in the workplace is at risk of shifting into a diamond, workplace experts say, with fewer juniors at the base, a swollen middle of skilled operators and middle managers, and a smaller cohort of leaders at the top.’
Me: The smaller cohort of leaders will like that.
Ocypode: Zitron has been on the money about the AI stuff; his most recent (premium, so I only read the introduction) post on the topic about it most likely being much worse than the Dotcom bubble seems more apt.
No. Zitron is naive and simplistic. Forex, he writes:-
‘The AI bubble bursting will be worse, because the investments are larger, the contagion is wider, and the underlying asset — GPUs — are entirely different in their costs, utility and basic value than dark fiber. Further, the basic unit economics of AI — both in its infrastructure and the AI companies themselves — are magnitudes more horrifying than anything we saw in the dot com bubble.’
Yes. Sure, Altman and his OpenAI Ponzi — and it is a Ponzi — will crash and, yes, the crash of Altman’s OpenAI and some of its competitors may very well be ‘horrific’ in terms of the resulting economic carnage for the greater US population.
But from fine from the POV of, forex, Satya Nadella’s Microsoft, that is excellent, great. Because when that happens Microsoft will then own OpenAI’s weights, infrastructure, distribution, and its own datacenters, which will be model-agnostic running every type of AI system — not just LLMs, but graph neural networks (GNNs), diffusion models, neuromorphic models, neural radiance fields (NeRFs) & 3D scene understanding — which someone somewhere wants to make available through those datacenters.
In other words, all today’s ‘technofeudalist’ corporations — Amazon, Google, etc., as well as MS and Oracle, which already existed — achieved their current dominance as a result of the Dotcom crash, when they became the biggest things left standing and were able to take over because of it. All the people running those companies today were around then during that crash and learned that lesson.
So the Great AI Crash of 2026, or 2027, or 2028 — whenever it triggers — is in fact the business model. They’re counting on it.
I agree with the author that AI is likely to increase the premium for people with valuable skills, including the skill of cutting through AI bullshit. But I’m not sure I agree with this part:
It seems to me we are likely to see a kind of Jevon’s paradox for bullshit generation. As bullshit becomes cheaper to generate, more of it will be generated. Somebody needs to be on the receiving end, and I suspect all those extremely skilled elites will make sure it isn’t them.
We have no idea what hyper-Dilbert is gonna look like.
albrt: …we are likely to see a kind of Jevon’s paradox for bullshit generation. As bullshit becomes cheaper to generate, more of it will be generated. Somebody needs to be on the receiving end, and I suspect all those extremely skilled elites will make sure it isn’t them.
Yes. Absolutely this, and weaponized to the nth degree.
What else, after all, were and are the likes of Cambridge Analytica?
“The advanced societies of the future will not be governed by reason. They will be driven by irrationality, by competing systems of psychopathology.”
― J.G. Ballard
“that tools like ChatGPT now provide instantly and at almost no cost.” This is nonsense. The cost of AI is what is going to sink it. Just because the per token cost might be coming down in no way means the cost of inference is actually declining. These LLM models require eye-watering amounts of debt to both buy the gpus necessary and the buildings with which to house them. Cracks are already starting to appear in the financing, as Blue Owl walked away from Oracle’s Michigan project. It’s also becoming more public how useless AI is with more companies announcing how they aren’t getting any return on their investments. Unfortunately the lazy media still reports AI replacing staff because when they see layoffs at a company that says it’s doing AI they automatically report that AI took the jobs when usually the layoffs were planned long in advance and had nothing to do with AI.
From what I gather the cost of inference is rising, because once engineers figured out training couldn’t be much improved upon (the entirety of the internet + all digitized books) they had to start making inference more intensive. So in fact not only are the fixed costs of training not meaningfully being reduced, but also the current costs of operation seem to be rising.
My personal take is that the only cost-effective realistic use case for “AI” is very small models for voice recognition in specific devices (i.e. with very limited purposes) which can be run on very simple chips. A guy did that a few months ago, but I can’t seem to find the link right now; it was impressive seeing a voice assistant in extremely simple hardware.
AI-enabled monopsony causes late-capitalism to eat itself. Middle class (‘white collar’) jobs are not the only victims. “All that is solid melts into air, all that is holy is profaned”. That’s capitalism/AI (delete as appropriate), that is!
Record corporate profits and GDP growth occurring alongside rising unemployment and job stagnation (as seen in Q3 2025). This “growth without jobs” signals that capitalism is already no longer dependent on human labor for prosperity. Goran Paverevski’s analysis shows AI as the Executioner of the labour market, but it was on life-support already.
AI will write my letter to the company, and the company’s AI will read it and write a reply.
“that’s not at all responsive…”
A widespread adoption of AI requires a similar breadth of trust in the systems that are being built out, yet it seems that trust only continues to erode within the general public. Where is the magic “AI” tool actually unlocking these benefits; certainly it isn’t the chatbots, right?
So AI will have the same effects and impacts as with the Great Depression?
Based on the link provided, it seems Althoff and Reichardt have not yet published their paper in a journal. Perhaps you could offer them your services as a referee, to help get the paper into a publishable form. That sounds like snark but really, why not offer?
As for jobs requiring high-order reasoning, those usually require years of experience doing jobs of increasing responsibility. But there might not be many of those people after a while, and no practical way to train some, a bit like finding young programmers who can handle old computer languages.
It’s not snark, I already posted my commentary on Althoff’s LinkedIn and we had a rather productive debate in the comments, albeit maybe too technical to be of interest for the general public. :)
However, i doubt they will revise the paper to incorporate post-Keynesian considerations of bargaining power and demand constraints. The reason the model is so tractable and easy to estimate, and the results are so optimistic, is precisely because it abstracts away from the messy reality.
Thank you. I was thinking of trying to make the same point but you did it better than I would have.
In some occupational areas where the “good enough” AI product is popular, I wonder if boutique shops staffed by, gasp, humans producing very high value products might evolve? They might even train new people!
I am also wondering what happens when an AI makes, say, a poor translation or hallucinates and the law suits start?
As someone working in the industry, I would say there is a large gap equivalent to a year lag between what the general public thinks AI is capable of and what frontier LLM models can actually do. And that frontier is continuously being pushed further at a pretty constant pace.
Inference costs per token have actually dropped between 100-1000x times since ChatGPT was first launched and they continue to decline.
None of this implies that investment in AI is not fueling a massive bubble, whether in AI infrastructure, data centers or AI startup valuations. Indeed, running the financials reveals ample evidence in support od this view.
But the fact that many, if not most AI end-user applications have not shown ROI has more to do with poor implementation, deficient business process optimization protocols, and use case misidentification, especially in large corporations, than with the capabilities of the underlying models. It might take a long time, decades even for large companies to figure out how to properly implement and adopt the technology and rationalize their workforce (or be driven out of business by smaller competitors), but the fact that the capability is already there means that the outlined trajectory of the labor market fracturing is largely inevitable.
I agree that AI excels at generating bullshit (among other things), but this implies a shift in value: professional bullshitters will lose worth, while true experts who can spot AI-generated slop will become indispensable. The colleague endlessly churning out valueless slides will grow harder to justify, forcing companies to rethink business processes and reallocate employee time more effectively.
I’d like you to quantify your remark that inference costs have come down, at all. OpenAI and Anthropic are burning more money than ever. Other than NVIDIA there is not a single AI company that is profitable, not one. Amazon, Microsoft and Google are profitable from their other services. The latest chips from Nvidia cost more, and take more power than the last set. There is no possible way inference costs less today than it did in 2022. B200’s are far more costly than H100’s, and the latest models burn far more gpu cycles than ever. In particular with OpenAI, the way they now handle their so-called “routing” their inference costs can only increase. OpenAI especially is on target to lose more money this year than ever. It is absolutely false to say the cost of inference that OpenAI pays to Microsoft and others is less this year than last.
Too slow to edit my post. In 2024 OpenAI spent $3.7 billion on inference. OpenAI spent $5 billion on inference, just with Microsoft by July 1 in 2025. By September they had spent $8.7 billion on inference. That is not in any sense a 100-1000x reduction in the cost of inference.
If anyone is interested, here is a link to a newsletter from Ed Zitron that details OpenAI ever increasing inference costs:
https://www.wheresyoured.at/oai_docs/
Those numbers you point out is total inference cost, which of course is increasing because overall use is increasing.
Per token inference cost indicates what it would cost you as a user to run the same query and obtain the same “amount of thinking” from the model, which is the relevant consideration. That has dropped in the 100-1000x range as I was saying, here’s a link: https://epoch.ai/data-insights/llm-inference-price-trends
We’ve already been told that by 2030 “you’ll own nothing. And you’ll be happy” — 8 Predictions for the World in 2030, WEF, 2016. https://sociable.co/web/wef-director-youll-own-nothing-be-happy-disinformation-campaign/
The WEF’s ‘technocratic takeover’ is proceeding to plan.
fly comment, sorry:
(super busy w preps for alberta coming to visit.)
what all this requires…assuming the tech even works…is a totally different outlook, re: labor/people on the part of the Boss Class.
if this stuff does as advertised, then yes…lotsa jobs will suddenly vanish.
and theres musk, et alia, yammering about how great it will be for us peons, when we dont hafta work.
but are they also talking about UBI, or some civilisational dividend?
not that i see.
musk often talks about Star Trek…and wears a start trek shirt a lot…but does he really understand how an economy works?
do any of these guys?
or the multitudes of lesser rich guys?
i somehow think not.
for that fully automated luxury communism of star trek, the agreed upon purpose of the vast numbr of non-rich humans must change utterly…and a new sort of Humanism must be ginned up out of nowhere.
i think these guys, and the hypercapitalist, ordinary boss-types that came before…are the last people to engineer all that.
ergo, they will hafta be eaten or composted at some point.
I guess it depends on what kind of documentation that you’re doing. It definitely won’t replace documentation where there’s a new feature or revision of existing functionality. LLMs have no there-there. So useless for that.
Maybe you can train them on all your internal documentation, which is itself written by developers, scattered all over the org, and forever out of date.
But I doubt it.
But yeah, some companies’ leadership will stupidly think no we don’t need any professional technical writers.
Good luck with that.