A Venture Capital Fund Says the AI Job Apocalypse Is a Fantasy. I Wouldn't Be So Sure
On May 6, David George of a16z published on the fund's Substack an essay titled "The 'AI Job Apocalypse' Is a Complete Fantasy". a16z is Andreessen Horowitz, the venture capital fund that has invested tens of billions of dollars in artificial intelligence startups in recent years. It’s worth keeping this in mind because the piece discusses whether artificial intelligence will destroy human work, and the author has some not insignificant economic reasons to conclude that no, it will not.
I read it twice, reviewed the graphs, looked for the cited academic papers, and one thing must be said right away. In the central part of the reasoning, a16z is correct. However, the conclusion is debatable. I will attempt to explain in what sense.
The Argument of a16z, in Summary with Three Examples
The main argument is that those predicting a "permanent underclass" of workers displaced by AI are falling into what economists call the lump-of-labor fallacy. Brief explanation: it is the idea that there is a fixed amount of work in the world, a kind of pie to be shared. Thus, if AI takes a piece of the pie, humans are left with less; and if AI takes the whole pie, humans are left with nothing.
a16z responds: the pie is not fixed. Every time a technology has significantly increased productivity, jobs have disappeared on one side and been created on the other—resulting in a net positive. Three heavy examples are provided in the piece.
American agriculture has transitioned from about one-third of total employment at the beginning of the 1900s to 2% in 2017. The introduction of the tractor was supposed to permanently disrupt the job market. Instead, the population grew, former farmers ended up in factories, offices, hospitals, laboratories, and then in services and software.
The electrification of American factories went from 5% to almost 80% between 1900 and 1930. Labor productivity doubled and continued to double for decades. The demand for labor rose, not fell, because more production means more of everything that revolves around it: commerce, finance, maintenance, distribution.
American travel agents went from about 130,000 in 2000 to less than 60,000 by 2025 due to the internet. However, the aggregate employment of the U.S. population, adjusted for age, is the same as it was 25 years ago. And the travel agents that remain earn 99% of the private sector average today, compared to 87% in 2000. Technology has made them fewer in number but more productive and better paid.
From this comes the point underscored by David George: the labor market is not a fixed stock; it is a flow that is reconfigured. When a significant input becomes cheaper, in this case cognition, the same thing always happens: the Jevons Paradox; when the price of a resource drops, total consumption of that resource rises because it becomes accessible for uses that were previously not economically sensible. This has been observed with coal, electricity, and computing power. If Jevons' Paradox also holds true for cognition, then total cognitive labor will ultimately expand in directions we cannot currently imagine.
a16z also provides recent data to show that, so far, AI is not actually destroying jobs. It cites an NBER Working Paper from 2025 that concludes: "AI adoption has not yet led to meaningful changes in total employment." It then refers to a Working Paper from the Atlanta Fed of 2026, in which over 90% of surveyed companies report no impact on employment over the last three years. Finally, it references the Yale Budget Lab from April 2026: "the picture of AI's impact on the labor market that emerges from our data largely reflects stability." Additionally, it notes that the demand for software engineers and product managers is rising in 2025-2026, that new business formations are exploding, that new apps on the App Store are growing by 60% year-on-year, and that the robotics industry is finally taking off due to computational costs.
I conclude the summary. The argument is solid in its structure. Historically, markets have absorbed technological shocks by creating new work. Today, the data does not show a net job destruction, therefore the apocalyptic prediction is speculation.
Where the Argument Really Holds
I think the part about the lump-of-labor fallacy is correct. The labor market is not a stock, and those talking about a permanent underclass of excluded individuals are making a serious conceptual error. History shows instead that some categories have disappeared while others, simultaneously, have emerged. In this, a16z is right, and the data it cites about the three historical examples are honest, verifiable, and not cherry-picked. The story of travel agents is particularly beautiful because it encompasses the entire pattern within a single sector: fewer people employed, but those remaining are better paid, and the rest of the workforce has not dissolved into nothing but has gone elsewhere.
Also, the data on recent academic papers is real. It is true that the literature of 2025-2026 struggles to find significant aggregate effects of AI on employment. It is true that where something is moving, it’s in the allocation between tasks, not in net destruction. On this, a16z is not rigging the deck.
Where the Argument Falls Apart
The problem is that the entire argumentative structure of a16z rests on a very convenient removal: that of the temporal dimension. American agriculture took a hundred years to move from about one-third of employment to 2%. The electrification of American factories took thirty years to go from 5% to 80%. Even the fastest case, that of travel agents, unfolded over about twenty years, from 2000 to 2020. None of these examples are quick transitions; they are all transitions that had time to be generationally absorbed: American farmers did not become programmers; their grandchildren did.
The adoption of generative AI in business is happening on much shorter scales. ChatGPT reached 100 million users in two months from its launch in early 2023. Tools like Claude Code or Cursor now allow those who use them competently to complete in half a day tasks that required entire teams for a week just two years ago. We're talking about an order of magnitude faster than how American industry was electrified.
The point is that a16z's economic theory may be right in the long run and simultaneously irrelevant in the short run. If the absorption of displaced workers happens on a thirty-year scale while the displacement occurs on a three-year scale, there’s a generation in between that pays the price. This generation votes and carries political consequences that can change the trajectory of technology itself, not just its adoption.
There is a comment on a16z's piece, written by a user named Scenarica, that expresses exactly this much more clearly than I could. I quote (my translation): "The AI doomers are wrong about the destination. The optimists are wrong about the journey. What actually happens depends on whether the reskilling infrastructure scales fast enough to prevent the transition from becoming a political crisis before new jobs arrive."
Scenarica also brings a very heavy statistic. The U.S. spends about 0.1% of GDP on active labor market policies for those who have lost their jobs. The OECD average is 0.6%. Six times more. And this is the starting point at a moment when the speed of the transition is accelerating by an order of magnitude. Quick explanation on active labor market policies (in English known as ALMP): these are public programs for retraining, conversion subsidies, and support for new work for those who have lost their old job. They are the concrete mechanism through which a16z's economic theory should materialize, and today in the U.S. they are scaled for the transition speed of the 1980s, not for what is coming.
To this, I add two numbers from the BLS Displaced Workers Survey of January 2024, which is a snapshot of how transitions are actually absorbed today. In the BLS definition, a displaced worker is a person of at least 20 years old who has lost their job because the company has closed, moved, or eliminated the role—that is, for structural reasons and not individual layoffs.
Among long-tenured workers (those with at least three years of service at the lost job), 65.7% find new employment, 16.1% remain unemployed, and 18.2% exit the workforce entirely. For those over 55 years old, the reemployment rate drops to 55.3%. Above 65, to 34.4%. These are baseline numbers, before the true effect of AI is felt, and already they describe an absorption far from the 100% that the a16z model implicitly assumes. On a scale of 100,000 displaced individuals, we're talking about 35,000 people who exit forever from the productive workforce. On a scale of a million, 350,000.
There is also a macroeconomic dimension that a16z's piece does not consider, and which I discussed extensively in my book published last week, Humans for Now. I call it Ghost GDP, and it is the share of GDP growth generated by automated systems that does not translate into income for workers, and therefore does not return to the real economy as demand. GDP measures the added value of production, not how that value is distributed. If a company replaces one hundred analysts with an AI system and productivity doubles, GDP grows; meanwhile, those one hundred analysts have stopped paying rent, shopping, dining out, and renewing health insurance.
The demand they expressed disappears from the economy, and GDP does not immediately register it because it looks from the production side. Historically, the mechanism had a physiological brake because machines that replaced workers still required other workers to be built, managed, and maintained, while a model of AI trained once can be scaled to millions of instances without hiring anyone more, and the compensatory mechanism that historically softened the replacement becomes much less effective.
The Framing Problem
I realize that the a16z blog is written by a venture capital fund with very clear economic incentives to support the thesis that AI will not destroy jobs. However, the argument, separated from the author’s motivations, holds in the long-term aggregate, and it is important to recognize it. The problem, which recognizing the argument does not help to solve, is that the long-term aggregate plan is not the one on which the working life of real people is played out.
Real people live in their careers that last forty years if they are lucky, in their geography which is often not chosen, in their age which cannot be reset.
A 52-year-old copywriter whose agency takes away 80% of their work because three agents using Claude do it for her in a quarter of the time doesn’t have twenty years ahead of them to wait for absorption. They have a maximum of ten years of useful career, realistically two or three for retraining before entering the age group of the BLS statistics where reemployment falls below 60%. Their personal trajectory is decoupled from the aggregate trajectory that the a16z piece describes, and the two do not communicate.
And here I return to the starting point. The argument of a16z is a useful remedy against the apocalyptic rhetoric of AI doomers, who do indeed tell a story equally unsupported by data. However, it is also a fairy tale in reverse because it transforms the validity of the argument at the aggregate level into an implicit reassuring conclusion about individual people. The two do not ever completely coincide, and in a rapid transition, they coincide even less.
The exit route, which Scenarica's comment points to and a16z’s piece avoids addressing, is political. It means recognizing that historical absorption only works if the transition infrastructure exists at a scale proportional to the speed of the transition itself, and that today in the U.S. it does not exist at that scale. Bringing U.S. spending on active labor market policies from 0.1% to 0.6% of GDP would simply be an alignment with the OECD average on a specific budget line, within a time frame compatible with the adoption speed of AI, assuming a decision is made to do so.
David George closes the article with "LFG!", which stands for "let's fucking go;" I'll spare you the literal translation. It is a conclusion consistent with one who looks at the aggregate where billions of dollars have been invested. What it does not say is who will take care of the details. Historically, the pathways between one configuration of work and the next have lasted three generations; now it is expected that they happen in five years. This part of the discourse is entirely missing from the blog.