Back to blog
24 min read

AI and Jobs: Who Benefits from the Narrative?

Are we being manipulated? Does the race with China justify everything? What are the numbers hiding? Is this a bubble about to burst? And what are the real risks? A European reading of the grand AI-and-jobs narrative.

  • ai
  • future-of-work
  • economic-impact
  • europe
  • agi
  • jobs
AI and Jobs: Who Benefits from the Narrative?

AI and Jobs: Who Benefits from the Narrative?

Manipulation, the race with China, numbers that lie by omission, a financial bubble, real risks — a European reading, without the hype

Difficulty: Accessible — no code, no prerequisites. Just critical thinking and a few numbers.

In twelve months, the conversation around AI and jobs did a complete U-turn.

May 2025: Dario Amodei, CEO of Anthropic, announced that AI could eliminate half of entry-level office jobs within five years and push unemployment to 10–20%. He even accused executives and governments of “sugarcoating it.”

May 2026: the same Amodei, now on stage alongside the CEO of JPMorgan, explained that AI will mostly transform and multiply jobs. And Sam Altman, at OpenAI, declared himself “glad to have been wrong.”

Same technology. Opposite message. One year apart.

And then, on June 1, 2026 — two days ago, as I write this — Anthropic filed confidentially for an IPO. Keep that timeline in mind; we’ll come back to it.

This 180-degree turn is, for me, the most instructive detail in the entire debate. Not because it proves these people are lying — I don’t know that, and nobody can prove it. But because it forces us to ask four questions we too often avoid. Are we being manipulated? What are the numbers hiding? Is this a bubble? And at bottom, what are the real risks?

I’ll try to answer honestly, from where I sit: Europe.

Personal note. I’m not an economist. I’m a software engineer, and for some time now I’ve been helping SMEs — many in Belgium — with AI and automation. My job is precisely to separate what a tool can do in a demo from what it actually changes once it’s in production. No crystal ball here. A ground-level reading, and a healthy dose of skepticism.

1. Are we being manipulated?

Let’s start with the word that stings. I don’t believe in a coordinated conspiracy in a Silicon Valley basement. But I do believe in convergence of interests, and it’s troubling. The narrative you receive serves a handful of actors remarkably well — OpenAI, Anthropic, Nvidia, and the small circle orbiting them. Let’s look at the logic in two acts.

Act one: “AI is going to replace your jobs.” That’s the fear narrative, and it serves its purpose well. A technology capable of replacing millions of white-collar workers is worth hundreds of billions. Fear justifies dizzying valuations, and therefore the colossal fundraising rounds needed to pay for the race. It also pushes companies to adopt the tool urgently — “adopt or disappear.” And it positions these companies as the responsible adults, the ones who “tell the truth” and should therefore be consulted when the rules are written. Fear sells power, and power attracts capital.

Act two: “AI is going to create jobs.” That narrative arrives later, and the timing is no accident. Let’s recap the chronology:

  • May 2026: Amodei and Altman soften their tone and bring back Jevons’s paradox (“if you automate 90% of a job, everyone does the 10% that’s left”).
  • Late May 2026: Anthropic closes a massive funding round.
  • June 1, 2026: Anthropic files confidentially for an IPO, and OpenAI is reportedly preparing a similar move.

Now, for an IPO, a jobs-apocalypse narrative is a nightmare: retail investors and regulators buy a “job-creating growth” story far more readily than a promise of mass unemployment. The tone turns reassuring right when the markets need to be wooed. Coincidence, perhaps. But the timing is suspiciously convenient.

The orders of magnitude, to put the stakes in perspective — and this is where my own numbers recently blew up. Anthropic closed late May with a $65 billion raise at a $965 billion valuation — nearly a trillion, more than double its February value. The company has overtaken OpenAI; its annualized revenue crossed $47 billion according to Reuters and Bloomberg — but that’s not an audited figure published in a prospectus.

OpenAI, for its part, is valued at $852 billion (after a record $122 billion round in March) and generates roughly $2 billion in revenue per month; according to press-reported projections — unaudited — it would still be looking at roughly $14 billion in losses for 2026 alone. Massive losses on one side, a race for capital on the other: maintaining a future-growth narrative isn’t a luxury for these companies, it’s a necessity.

As for Nvidia, which sells the chips, it has every incentive to keep the race going — to the point of investing in its own customers (more on that later).

Are we being manipulated? Let’s put it cleanly: you can’t prove intent, but you can name the incentive structure. And the hygiene rule that follows is simple. When someone tells you what AI is going to do to your job, always ask yourself: who benefits if I believe this? That question applies to AI vendors. It also applies to those crying bubble to bet against the market. And — I want to say this clearly — it applies to the companies that build AI while describing its effects to us. Nobody in this debate is a neutral observer.

2. “But what about China?” — the argument that ends every debate

Ask any uncomfortable question about AI — “why spend so much?”, “should we regulate?”, “is this a bubble?” — and you’ll almost always get the same three-word answer: “but China.” National security is the argument that shuts down the discussion. Nobody wants to be the one who hit the brakes while Beijing accelerated.

Except that argument, too, deserves a close look.

First, the race is real. In April 2026, China’s DeepSeek released its V4 model as open source — downloadable by anyone, unlike closed American models. And according to the Stanford AI Index 2026, Chinese labs have “nearly closed” the performance gap (other analyses, like CFR’s, estimate American models still hold a lead of a few months).

But the Western bet — blocking access to cutting-edge chips to strangle China — has largely backfired on whoever made it. Beijing ordered its giants to cancel Nvidia orders and pushed its own champions: Huawei released a rival system (CloudMatrix 384), Cambricon and Alibaba are ramping up at full speed. Better still: when Trump traveled to Beijing in May 2026 — bringing the Nvidia CEO at the last minute — and proposed selling chips to China, Beijing refused. “They choose not to buy, because they want to develop their own,” Trump summarized. By Jensen Huang’s own admission, Nvidia’s market share in China went “from 95% to zero,” and he now calls export controls a “failure” that “gave China the energy and state support to accelerate.”

That refusal isn’t a whim — it’s doctrine. Xi Jinping set an explicit target: make China the world leader in AI by 2030 and achieve technological self-sufficiency, now inscribed at the heart of the 2026–2030 five-year plan. Depending on Nvidia is precisely what that logic commands them to avoid.

Let’s be precise nonetheless, because the reverse narrative — “China doesn’t need anyone anymore” — is just as false. Chip for chip, Huawei remains far behind: its best accelerator tops out around 60% of a previous-generation Nvidia chip, and it takes five times as many, at nearly four times the electricity, to match — or even exceed — an Nvidia rack in raw compute. Above all, China still can’t produce enough: its fabs struggle with low yields and depend on imported memory. Its companies must therefore ration compute — while continuing, in parallel, to smuggle Nvidia chips in via Singapore or Malaysia. The brake leaks both ways.

Which leaves the question that interests us: who benefits from fear of China? It does for spending what fear of unemployment did for adoption — it removes the brakes.

Hard to push back against hundreds of billions in investment when it’s framed as an arms race. Larry Page, Google co-founder, reportedly said he’d rather “go bankrupt than lose this race.” And the voices calling loudest for export controls are often those of the labs that would benefit most — Dario Amodei, CEO of Anthropic, chief among them. Once again: nobody is neutral.

3. Are we being told the truth? What the numbers leave out

To answer, you have to look at the data. And the best data comes from the United States — not because it’s the center of the world, but because it’s the advanced laboratory: flexible labor market, rapid adoption, concentrated actors. What happens there often reaches us with a lag.

Here are three things the headlines forget.

The headline number is the least informative.

The US unemployment rate is stable around 4.3% (April 2026, latest available figure). Reassuring? Not really.

In 2025, after a massive revision of the figures — job creation for the year was revised from +584,000 to +181,000, and the March 2025 employment level was lowered by nearly 900,000 jobs (898,000 in seasonally adjusted data) — the economy created on average only about 15,000 jobs per month, virtually a standstill for a country that size.

February 2026 was even negative. The market isn’t collapsing, but it’s frozen: almost nobody is hiring, and almost nobody is laying off either. The quits rate fell to 1.9% in April — its lowest since summer 2020 — because nobody dares to move; and hiring has dropped back to among its lowest levels since 2020.

The real damage isn’t in mass layoffs. It’s the door to entry-level jobs quietly closing.

The clearest signal points to juniors — even if the cause is debated.

A Stanford Digital Economy Lab study, built on payroll records for 25 million workers, shows that in the occupations most exposed to AI, employment among 22–25-year-olds has fallen by about 16% since late 2022 — up to nearly 20% among developers. And in the same occupations, employment among those over 30 has actually increased. It’s not seniors taking the hit, it’s beginners. The question is why.

Stanford sees AI, which excels at reproducing the “textbook knowledge” of a young graduate. But a Federal Reserve Bank of New York study (June 2026) offers another explanation: remote work — and the difficulty of training beginners remotely — accounts for nearly two-thirds of the rise in unemployment among young graduates since the pandemic, with AI exposure, in its view, insufficient to explain the gap over 2022–2024. Stanford counters that its effect holds even when excluding remote-workable occupations.

The solid point, however, doesn’t move: the first rung of the labor market is weakening — the exact cause remains disputed.

Many layoffs “due to AI” have nothing to do with AI.

Here, let’s be frank. Forrester estimates that a significant share of layoffs attributed to AI are actually financially driven: the technology serves as justification (“AI washing”), because the tools capable of actually doing those jobs don’t exist yet. Better still: 55% of employers who laid people off “because of AI” regret it, and more than half of those positions could be quietly rehired — often elsewhere, at lower pay.

When Meta, Amazon, or Oracle (profitable companies) cut jobs, it’s often to finance their AI investments. As the head of a firm that tracks these layoffs puts it: whether the job is replaced by AI or not, the money that paid for it definitely is.

And the figure that should close the debate on “mass replacement already underway”: according to the MIT report “The GenAI Divide,” 95% of generative AI initiatives studied show no measurable impact on the bottom line. If AI were truly destroying jobs at scale, we’d see the corresponding productivity gains. For the most part, they’re not there.

That’s been my through-line for a long time: what a tool can do and what it will replace are two very different things. Between them lie organizational inertia, regulation, and trust that has to be built. That slows everything down.

That said — and this is a nuance that matters — that 95% doesn’t condemn the technology. The MIT report is explicit: failure comes not from model quality, but from the approach. It points to a “learning gap”: AI is grafted onto old processes without rethinking workflows or training teams.

The minority that succeeds does the opposite — one specific problem, well executed. Shopify, for example, made AI use a baseline expectation: before asking for more headcount, a team must demonstrate it can’t do the work with AI. And its CEO makes a fair point: “using AI well is a non-obvious skill,” and many give up after a single disappointing prompt.

In other words, missing ROI is often a deficit of training and method, not proof that the tool is hollow.

But watch out for the opposite mistake, because “you’re using it wrong” is also vendors’ favorite argument. Klarna, which replaced 700 agents with a bot, reversed course in 2025 when quality collapsed — adopting a hybrid model.

The lesson is therefore neither “AI is useless” nor “automate everything”: value comes from disciplined, well-trained use, with humans still in the loop. Which, once again, is an organizational issue — not a technological one.

4. And here at home? Europe isn’t America

This is where it gets interesting for us, and where caution is warranted, because Europe is playing a different tune.

First, the big picture. EU unemployment stands at 6.0% in April 2026 (6.3% in the euro area), broadly stable over the past year: no visible shock in the aggregate numbers. Youth unemployment stands at 15.1% in the EU (14.7% in the euro area), slightly down in recent months. On productivity, a 2026 study covering more than 12,000 European companies — conducted by economists at the Bank for International Settlements and the European Investment Bank — finds that AI adoption raises productivity by about 4%, without short-term job losses. And according to a European Commission survey of more than 70,000 workers, 30% of them already use AI at work, mostly conversational assistants.

I note a paradox nonetheless: at 6.0%, European unemployment is higher than in the United States (4.3%). If Europe is “protected,” it’s therefore not in the sense of a more dynamic market — it’s that the AI-specific signal (the collapse in junior hiring, layoffs attributed to AI) is, for now, much less clear. Why? Three structural reasons. Stronger employment protection than in the United States. Slower adoption — Europe lags on AI patents and investment, which, for once, acts as a cushion. And an aging population creating labor shortages (healthcare, logistics, engineering) that partly offset losses.

But let’s not fool ourselves either. For Europe, Carnegie describes a scenario of “hollowing out before elimination”: jobs don’t disappear all at once, they slowly lose their substance, creating insecurity that lasts. The German Institute for Employment Research points to 1.6 million jobs “reshaped or lost” over fifteen years in Germany alone — a long horizon, not a sudden shock.

And watch out for a typically European attribution trap: the weakness of German and French industry stems first from energy costs and slack global demand, not AI. Confusing the two distorts everything. To put a real shock in perspective, remember: at the worst of the eurozone crisis, in 2013, unemployment reached a little over 12%. We’re far from that.

One last point that strikes me, and should strike our decision-makers: the narrative roiling Europe is manufactured elsewhere. Fear, then reassurance, then the race with China: all of this is largely an American product. Our governments and companies react to a story written in San Francisco, while our reality is different. Worse: that same narrative now comes back to tell us we need to loosen our own rules so we don’t “fall behind” in a race we’re not running. The AI Act regulates certain AI uses in employment — recruitment, personnel management, classified as “high risk” — but it’s not a genuine transition framework for job displacement, and it’s already under pressure to be watered down. Double blind spot.

5. Is it a bubble? And will it burst?

Let’s get to the money, because that’s perhaps where the real issue lies.

The mechanism to understand is called circular financing, and it’s simpler than it sounds:

  • Nvidia invests in OpenAI.
  • OpenAI orders hundreds of billions in compute capacity from Oracle.
  • Oracle buys its chips… from Nvidia.

Money spins in a circle among a handful of actors, inflating demand that looks spontaneous when part of it is just the same sum passing through multiple times. Serious analyses put these cross-arrangements at several hundred billion dollars. A telling detail: Nvidia’s announced investment in OpenAI — up to $100 billion — is paid out as OpenAI deploys… Nvidia systems. In other words, a good chunk of that “investment” comes back to Nvidia as hardware purchases.

And it goes beyond transactions: circularity has contaminated income statements. In Q1 2026, Alphabet booked nearly $37 billion in gains on its investments — mostly unrealized gains, attributed by analysts to the revaluation of its stake in Anthropic. Amazon accounted for nearly $17 billion (Anthropic), Microsoft nearly $6 billion (OpenAI). In other words: the giants’ profits are partly juiced by the revaluation of the labs they themselves finance. You spin in circles, and everyone gets richer on paper.

All of this is eerily reminiscent of the run-up to the telecom bubble’s burst in 2000, when equipment makers lent to their own customers so they could buy their gear. And the comparison with the Internet bubble is no longer the province of a few pessimists: it’s now made by central banks. The Bank of England has repeatedly warned of the risk that a correction in AI-linked valuations could destabilize markets.

The amounts are dizzying. The four American giants are set to commit around $700 billion in investments in 2026 — cloud, data centers, and AI combined — nearly double 2025 (Alphabet alone is planning $180–190 billion). To finance them, they’re taking on more and more debt, sometimes through opaque structures — and their cash is suffering: Amazon’s available cash collapsed by nearly 95% over twelve months, to $1.2 billion, and could go negative according to some bank projections. Fortunes are being spent on a technology that, as we’ve seen, still returns almost nothing for most of its customers.

Will it burst? Honestly, I don’t know, and beware anyone who claims they do. Optimists have solid arguments: today’s players are rich companies, some highly profitable (Google still had over $60 billion in available cash over twelve months), not the revenue-less startups of 2000; some commitments are conditional on results; and revenue, for its part, is real and growing fast. Skeptics have their arguments too: debt is concentrated, circular financing flatters the numbers, and some market signals are starting to diverge. The lesson of 2000 fits in one sentence, and I keep it in mind: “the technology is real” and “the financing is sustainable” are two distinct claims. They can perfectly well diverge.

6. So, what are the real risks?

Let’s lay out the scenarios. I’m not putting precise probabilities on them — that would be dishonest — just directions.

Risk #1: the bubble deflates.

This is the paradox I find most underestimated: a burst would, in the short term, do more damage to employment than automation has caused so far. An abrupt halt to construction and all the jobs that depend on it, chain defaults on debt, a market crash concentrated on a handful of giants. And there, Europe is exposed even without having joined the party: our pension funds and savings hold those same American stocks. When Wall Street coughs, Brussels catches a cold.

Risk #2: AI is genuinely adopted and becomes profitable.

This is the opposite scenario, and it has its own danger. If return on investment finally arrives, the automation wave that hasn’t happened yet kicks in for real — and it climbs from juniors toward mid-level roles. The problem, raised by several analysts, is that there is then no obvious safe sector: unlike past waves, where you migrated from one occupation to another, AI advances everywhere at once. And a dizzying second-order question arises: if every company cuts costs by automating, who is left to buy? An economy can show record productivity while household incomes erode. Success itself becomes a risk.

Risk #3, always forgotten: energy.

In the United States, data centers accounted for about 4.4% of electricity consumption in 2023 and could reach 6.7% to 12% by 2028 (LBNL/DOE). But in Texas, the grid operator received interconnection requests far exceeding the state’s peak consumption — over two and a half times that, just for the projects deemed realistic — before finding that actual consumption per site reached only half of what was requested, and revising its forecasts downward. Turbine shortages, interconnection delays, lack of water for cooling: physics sets its own pace. An almost reassuring consequence: you don’t replace humans at scale with AI you can’t plug in. But a worrying one too: those same underutilized data centers weaken profitability, and therefore feed risk #1.

And there’s an asymmetry we forget: this electricity wall is ours, not China’s. On its own, it already produces more electricity than the United States and the European Union combined, and in 2025 it added more than 430 GW of solar and wind — nearly eight times the new capacity connected in the United States. Where the West runs up against its grid, China runs up against chips. Each has its wall.

What I take away

Three ideas, which I want to keep together.

The narrative has outrun reality, and it serves interests. You can’t prove manipulation, but you can refuse naivety. When the tone shifts from alarm to reassurance right as IPOs approach, and when fear of China is used to remove all brakes, we have every right to read each statement asking who it benefits.

Today, the real effect is concentrated and quiet. Not a bloodbath: a hiring freeze and the door closing on newcomers. In Europe, it’s even more cushioned, and partly for the wrong reasons (we’re just slower to adopt). And where AI creates value, it’s rarely plug-and-play: it’s when you rethink methods and actually train teams.

The real short-term risk is probably financial, not technological. A deflating bubble would do more harm, faster, than automation that still struggles to prove its profitability. And if AI finally delivers on its promises, a different kind of risk — deeper — opens up.

What this means in practice

Because a finding without follow-through is useless, here’s what it means for you in practice.

If you’re starting your career: you’re on the front line, no point denying it. The answer isn’t to run from AI, it’s to become the one who steers it. Give it real tasks from your job, every day. The gap widens fast between those who use it seriously and everyone else.

If you run an SME (my case, and probably yours): beware both extremes. Don’t replace a team with an agent betting on AI that doesn’t exist yet — the vast majority of generative AI projects still show no measurable gain, and the regret rate for “AI” layoffs is massive. But don’t stick your head in the sand either. Target one specific problem, actually train your teams — that’s where a good chunk of the ROI is — then measure and decide. Results, not promises.

If you save or contribute to a pension: simply know that your exposure to AI isn’t just about your job. It’s also in your savings, via a handful of overvalued American stocks. That doesn’t mean sell. It means understand.

In closing

No figure in this article is a prophecy. The revisions of nearly 900,000 jobs remind us: even official statistics are fuzzy, and often revised long after the fact. And the best sources on AI often come from actors with every interest in shaping the narrative — I had to correct Anthropic’s valuation between two versions of this article, so fast do these numbers move.

But caution isn’t blindness. The real risk isn’t the power of the models — it’s whether the money and the electricity will follow.

Neither reassuring nor alarmist. Clear-eyed. And in times like these, clear-eyed is already an advantage.

See you soon for more articles — and always ask yourself: who benefits from what I’m being told?

Sources (figures verified as of June 3, 2026; analyst estimates and unaudited projections are flagged as such):

All articles
  • ai
  • future-of-work
  • economic-impact
  • europe

Ready to go further?

Let's discuss your case in 20 minutes. Free, concrete, no strings attached.

Book a slot