Did AI kill jobs? I checked the numbers, from developers to other professions

I started from a hunch shared by just about every developer I meet: AI sent the job market into a nosedive, and it is only now starting to recover because employers have worked out that a model does not replace a developer. I added two personal convictions on top. That the United States was two or three years ahead of us, so its curve told us what was coming. And that a developer equipped with AI produces ten times more than before, because I see it in my own work.

Before writing about it, I opened the data series instead of recycling impressions: Indeed Hiring Lab, the US Federal Reserve, Stanford, INSEE, APEC, the Bureau of Labor Statistics, the controlled trials on productivity. I did not stop at developers, because the same question applies to translators, accountants, customer service agents and radiologists. On all three of my convictions, the data contradicted me. Here is what it says, and what it suggests for what comes next.

The gist in 30 seconds

• The developer job market fell because of the 2022 rate hikes, not AI: 96.5% of the fall had happened before the first coding agents.
• Coding agents (Cursor, Claude Code, Codex) do coincide with the recovery, which mostly benefits experienced profiles.
• The “10×” is real for writing code, not for what a team ships: studies measure 5 to 15%.
• In almost every exposed occupation, AI first closes the door to beginners. It rarely lays off people already in the job.
• By 2034, US projections expect modest declines (customer service, secretarial work) and growth in data, security and software development.

The wrong suspect: for developers, ChatGPT is not the thing to look at

When people talk about AI and jobs, everyone takes ChatGPT as the starting point. For a developer, that is the wrong reference. A chatbot open in a browser tab does not replace an engineering position. The tools that could are the ones that write code directly in the editor, then modify it, test it and fix it on their own: GitHub Copilot, Cursor, Windsurf, Claude Code, Codex. Their release dates are what you need to compare with the employment curve.

Two very different eras follow one another. The first is the era of autocomplete. Copilot arrives in preview in June 2021 and becomes a paid product in June 2022: it suggests the next line, the developer accepts it or not. Cursor comes out in 2023 on the same principle, with a chat on top. The second is the era of agents, which read an entire project, modify several files, run the tests and try again until they pass. It starts in November 2024 with Windsurf, then Cursor 0.43, the first Cursor able to use the terminal. Claude Code follows in February 2025, Codex in May 2025.

The verdict is clear. Throughout the big fall of 2022 and 2023, the only coding tool on the market was Copilot's autocomplete, which had about one million paying subscribers in October 2023 according to Microsoft, out of some 27 million developers worldwide. The agents arrive between November 2024 and May 2025, when 96.5 to 99.7% of the fall has already played out. They cannot have caused a collapse that was almost over by the time they shipped.

While I am at it, I am correcting a mistake from the first version of this article. I wrote that 35% of the fall had happened when ChatGPT came out. The real figure is 53%: the 35% measured the drop of the index relative to its level at the time, which I had confused with the share of the total fall. And the peak is not December 2021, it is 233.84 on February 28, 2022. The argument comes out stronger, but the number was wrong, and I would rather say so.

The agents do, however, coincide with the recovery

The timeline gets interesting on the other side of the trough. The US recovery starts in May 2025, exactly when coding agents are spreading, and it speeds up while they become mainstream, from late 2025 into the first half of 2026. Its shape matches what you would expect from tools that multiply the output of experienced profiles: 71% of the rise in postings since May 2025 is for senior roles, and entry-level roles make up only 4.5% of postings in the first quarter of 2026. That is also how Indeed Hiring Lab reads it, attributing the lead of English-speaking countries to earlier adoption of agentic tools.

I am adding a caveat, because it matters. An ordinary recovery after a rate-hike cycle would also favor experienced profiles, and the shift toward senior roles was underway before 2025: mid-level postings had already lost 7.7 points of share between 2019 and 2025. No study yet links the use of a specific tool to hiring. The coincidence is striking, it is not proof.

Cursor, or adoption speed as the vendors tell it

Cursor alone illustrates how fast the shift happened, with one important caveat: Anysphere, the company behind it, does not publish audited accounts, and the figures below come from the company itself or from press leaks. Its annual recurring revenue reportedly went from about $100 million in January 2025 to $500 million in June, $1 billion in November, $2 billion in February 2026 and more than $4 billion in June 2026. SpaceX completed its acquisition on August 14, 2026. Claude Code reportedly reached $1 billion in annualized revenue in December 2025, and Codex 7 to 8 million weekly users in July 2026.

Independent surveys tell a slower story. In Stack Overflow's survey, the share of developers using AI goes from 44% in 2023 to 62% in 2024. In 2025, 84% use it or plan to, but only 31% use agents. Agents became mainstream between late 2025 and the first half of 2026, and for now only the vendors' own figures show it.

On the employer side, the big decisions all come at the bottom of the trough or later. Shopify's internal memo, in April 2025, asks teams to demonstrate that AI cannot do the job before requesting a new hire. Salesforce made no net engineering hires in its fiscal year 2026, with about 15,000 engineers working with Claude, Codex and Cursor, while still hiring around a thousand new graduates. These decisions shape the recovery, not the fall that came before it.

What actually brought the curve down

If it was not the tools, what was it? The cause economists point to is monetary. The Federal Reserve starts raising rates on March 16, 2022, two weeks after the peak. The correlation between its policy rate and postings in AI-exposed occupations is described as “extremely strong”, with the decline starting exactly when the central bank tightens. There is nothing mysterious about the mechanism: these occupations are concentrated in tech, finance and services, the sectors most sensitive to the cost of money.

Stanford, reviewing the literature, concludes that hiring declines began “after the Fed's monetary policy shift but before ChatGPT's debut”. Add to that the correction of post-Covid overhiring, remote work pushing employers toward experienced profiles, and a US tax factor little known in France: Section 174, which since 2022 forces companies to spread the deduction of their R&D costs over five years instead of deducting them in the same year. Engineering salaries became heavier from a tax standpoint at exactly the moment venture capital was drying up.

Since 2022, unemployment among workers highly exposed to AI has risen by 0.77 points, and among the least exposed by 0.85 points. The least exposed group did slightly worse. If AI were the main driver, that sign would be reversed.

That does not mean AI has no effect: the opposite position is just as wrong. The reading that holds up comes in two stages. From 2022 to 2024, a macroeconomic shock. From 2024 on, the tailwinds fade while substitution by AI becomes measurable, specifically at the start of careers. The AI effect is real. It simply arrived later than the story claims, on a market that was already down.

The recovery is real, but it does not look like a recovery

US postings have climbed back by nearly 15% since February 2025, while the overall job market fell by 7%. Tech is outperforming. Except the makeup of that rise is unambiguous: 71% comes from senior roles, and 37% from roles that mention AI in their title. In the first quarter of 2026, senior roles account for 69.3% of software development postings, entry-level roles for 4.5%.

Meanwhile, layoffs are breaking records: 128,536 people as of September 10, 2026, already more than in all of 2025. Postings going up and layoffs going up at the same time is not a recovery, it is a reallocation. The same company cuts and hires, but not the same profiles.

One last number to calibrate the optimism: at 76.62 on September 11, 2026, the US index is still about 23% below its February 2020 level, before the bubble. The recovery is a climb back toward a level that is still depressed.

The American lead: 4 months on the fall, 16 months on the recovery

I believed in a gap of two or three years. I calculated the peaks and troughs country by country. On the fall, there is almost no lead: the United States peaks in February 2022, France in July 2022. Four and a half months, nothing that looks like a strategic lead.

The gap is real on the recovery, and it is huge. The United States hit bottom in May 2025. France hit its low in July 2026 and is only 6.2% above it, within the noise of a flat bottom. Germany has not hit its low yet. Sixteen months apart at the very least, and the story is not over.

But the country that breaks my theory is European. The United Kingdom hit its low six weeks before the United States, and it has bounced back 16.4%. So the fault line is not the Atlantic.

This is where Indeed Hiring Lab's hypothesis makes sense: if English-speaking countries are recovering because they adopted agentic tools earlier, the French lag is not the delayed echo of the 2022 rate shock, but an adoption lag that turns into a hiring lag. The data does not settle which of the two theories is right, and I would rather say so than pick the one that suits me.

The French case, no sugarcoating

The French numbers are not good, and I would rather give them as they are. Hiring of IT managers and professionals is down 21% compared with 2023 according to APEC. Member companies of Numeum hired 83,000 people in 2023, 51,000 in 2025. Turnover is at a historic low of 15.6%, which reflects less a success at retaining people than a market where nobody moves anymore. INSEE confirms it from another angle: at the end of 2025, salaried employment among 15-29 year olds was down 7.4% year on year in IT services, against 0.7% for the private sector as a whole.

The most useful detail for a freelancer is elsewhere: the share of nearshore and offshore work in the revenue of French IT services firms (ESN) went from 15.5% to 18% in twelve months. The consequence is direct. Revenue growth at these firms in 2026 does not imply a recovery in developer hiring in France, because the extra capacity is bought elsewhere. We will see revenue recover before job postings do.

While I am at it, a correction: the “−80% developer postings in France” figure circulating in the tech press matches nothing in the index. The real drop is 62.4% from the peak. That is already severe enough without needing to be exaggerated.

And Switzerland, which interests many French developers

Switzerland deserves its own section, first for a methodological reason: it has no Indeed index. The Hiring Lab only covers eleven countries, and Switzerland is not one of them. So I cannot add it as a fifth row to the previous table, and I would rather say so than cobble together a comparable figure.

In exchange, it has something France does not: an actual academic study of the effect of AI on its developers. The KOF at ETH Zurich cross-referenced SECO unemployment registers with nearly all Swiss online job postings, from 2010 to August 2025. The results are the clearest I found in this entire investigation.

The authors write in black and white that developer hiring meant to support AI adoption “was not enough to offset the decline in labor demand in this field”. It is the most direct wording I have read, and it comes from an academic source, not a consulting firm.

But the Swiss trajectory looks like that of the United States and the United Kingdom, not like ours. Vacancies have more than halved since mid-2022, dropping below their 2014-2015 level, with a trough at −31% year on year in the second quarter of 2025, then a return to +6% year on year in the second quarter of 2026. Switzerland is already out of its trough while France is still in it and Germany is heading down into it.

On pay, the Swiss median sits around CHF 130,200 in 2026, but growth is close to zero: +0.2% this year, after +0.4%, +1.7% and +1.8%. The polarization is clear, with a median of CHF 177,938 for Senior Expert profiles, up 2.8%. A detail that matters for a French speaker: the regional index puts Zurich at 107.2 and the Lake Geneva region around 100. Working in French-speaking Switzerland costs almost nothing in salary.

For cross-border work, there are two traps to know about. The permanent amendment between France and Switzerland came into force on January 1, 2026: 40% remote work with no tax consequence, with named reporting by the employer and automatic data exchange from 2027. First, this 40% framework does not cover Geneva, which falls under a separate cross-border regime, even though Geneva is the largest employer of cross-border workers. Second, the social security threshold is different: 49.9%, not to be confused with the tax threshold.

The practical conclusion is a bit disappointing, but it is clear. swissICT reports a median of 2.5 days of mandatory on-site presence per week, which matches the 40% tax ceiling almost exactly. Working 100% remotely from France for a Swiss employer is therefore close to impossible. The realistic model remains the hybrid cross-border worker.

The “10×” exists in no study, and yet I live it every day

Now for my strongest conviction, the one I thought was beyond dispute because I observe it in my own work. No controlled study, whether from a vendor or an independent researcher, has ever measured anything close to a 10× gain for developers. The highest credible figure is +55.8%, on a toy task in a greenfield setting. Even the vendors selling these tools publish between 20 and 55%.

The most unsettling result comes from METR, in July 2025: sixteen experienced open source developers, 246 real tasks on their own repositories. The measured result, stopwatch in hand: 19% slower with AI. They had predicted they would be 24% faster. After the experiment, and after being slowed down, they still believed they had gained 20%.

A forty-point gap between what was felt and what was measured, among professionals working on their own code. This result does not say that those who feel faster are wrong. It says something more uncomfortable: none of us can settle this question by observing ourselves. That is exactly why we run controlled trials.

Then there is the argument that depends on no study and settles the question with arithmetic. Writing code accounts for 20 to 30% of a developer's work, and nearly 80% of us now spend less than half of our week coding. Even a 10× accelerator applied to that portion alone caps the total gain at about 1.22×. That is Amdahl's law. Claiming an overall 10× means claiming that AI has also made scoping meetings, code review and debugging ten times faster.

Why my 10× and the studies' 1.2× are both true

On a task I know how to specify, in a codebase I know, when the job is to produce code I could have written myself (a CRUD, a migration script, a test suite, a back office), the factor of 10 compared with typing by hand is real. It is even consistent with the studies: that is exactly the category where +55.8% and 35 to 40% are measured, and with tools from 2023 to 2025 at that.

What my experience cannot see comes down to three mechanisms. The first is the denominator: the studies do not measure writing code, they measure delivery. DX counts merged pull requests across more than 400 companies. Between my generated code and a merged PR, there is review, acceptance testing, security, integration. I feel the speedup on the part I execute, the organization measures what comes out at the end of the pipeline.

The second is memory selection. The gain is highest on simple tasks and falls to 10% or less on old, complex code. And the spectacular moments, the ones we remember and talk about, are concentrated in the first category. The three hours lost debugging a subtly wrong generation in an old module produce no anecdote: they look like an ordinary workday.

The third is the perception gap measured by METR, which targets introspection itself and applies just as much to someone who believes they are ten times faster as to someone who believes they have been slowed down.

But lived experience deserves its best argument, and it is a serious one: there is no controlled trial of the 2026 coding agents, Cursor in agent mode, Claude Code or Codex. METR tested early 2025 assistants, and the reference study on Copilot dates from 2022. When a developer describes a 10× today, they are often describing tools that published research has not yet evaluated. It is the gap most exploited by marketing talk, and it is also the best reason to think the published figures underestimate the reality of 2026.

Two things still prevent us from concluding that the 10× has arrived in teams without anyone noticing. First, the DX series runs until February 2026, with AI usage up 65% over the period, and still finds 5 to 15%: if agents had multiplied throughput by ten, that series would have seen it. Second, downstream friction is getting worse instead of easing. AI-generated code waits 4.6 times longer for its first review. Duplicated code blocks have increased by 81% to reach a record. And 2024 was the first year copy-paste overtook refactoring.

The 10× is real at the level of the generation task. It does not carry over to the team level, because the bottleneck has moved from writing to verification. Both statements are true at the same time, and the gap between them shows where a developer's value now lies.

Three hours on a button hover, or on a well-architected feature?

This is the question that started this whole investigation, and it is the only one where my intuition was right. For a measured reason, not out of conviction.

Veracode tested 80 completion tasks with a known security weakness on more than a hundred models. The assistants now exceed 95% functional correctness: the code works. But their pass rate on security tests is stuck around 55%, practically the same as two years ago. The models have become much better at writing code that works, and no better at all at writing code that is secure.

So AI has inverted the value of the two halves of the job. Producing syntax costs almost nothing anymore, and therefore no longer sets anyone apart. Judging whether code is correct, secure and well structured has become the limiting constraint, and so the place where a developer's hours pay off the most. The three hours on the hover effect buy a skill that is becoming a commodity. The three hours on the architected feature buy the one the market values.

That is also where the 10× finds its best use. If generation really is ten times faster on the tasks suited to it, the time freed up has only one truly profitable destination: the step that has become the bottleneck. The same three-hour gain gives opposite results depending on whether it pays for ten generated features that nobody reviews, which clog up review and get paid for later in technical debt, or a single feature that is generated, then architected, tested and checked for security. The 10× is not a reason to ship ten times more code. It is a reason to finally afford the work nobody had time to do.

One caveat, because it matters: to date there is no salary or job posting data showing that architecture, debugging or security command a growing premium. The mechanism is established, the effect on salaries has not been measured yet. And learning still has its place: the ability to evaluate what an AI produces is built by having written things by hand. For a beginner, the hover effect has a teaching value it no longer has for a professional. The right answer depends on whether the three hours are for learning or for shipping.

Other professions: where AI really hit

Developers are just one case among others, and not the most exposed. When you look at other professions, the same pattern comes back almost everywhere: the effect shows up at the entry point, on the hiring side, much more than in layoffs of experienced people.

Stanford measures it using payroll data from the provider ADP. Employment of 22-25 year olds in AI-exposed occupations is about 19% behind its trend, and falls by 3.8% a year while employment in the least exposed occupations grows by 2%. Among 41-49 year olds, no divergence. Young developers are down about 20% since late 2022, and young customer service agents are also clearly down. The mechanism is not layoffs, it is hiring coming to a halt.

The most instructive case is radiologists. In 2016, Geoffrey Hinton, one of the fathers of deep learning, advised that we stop training radiologists, since machines would soon read images better than they could. Ten years later, the United States is opening a record 1,208 radiology residency positions, more than 4,000 positions remain unfilled, and the average salary reaches $571,000. AI analyzes images very well. But a radiologist does more than read images, and demand for scans has grown faster than productivity.

Even customer service, presented in 2023 as the first profession doomed, is holding up better than expected. Klarna announced in 2024 that its AI agent was doing the work of 700 people. In May 2025, its CEO admitted that cost had been prioritized too heavily over quality, and the company was rehiring humans for complex cases. In the Philippines, call centers added more than 60,000 employees in 2025 according to their industry association, reaching 1.68 million. AI has slowed the sector's growth, it has not reversed it.

Accounting shows the difference between replacing a profession and replacing a task. In the United States, bookkeepers are declining while accountants are growing by about 5 to 6%. Same field, opposite directions. But the entry point is narrowing: at the UK Big Four, KPMG cut its graduate hiring by 29%, Deloitte by 18%, and these postings fell 44% year on year. AI and offshoring are working together, and they cannot be separated.

The best criterion I have found for telling which side a profession falls on is how AI is used in it. Where it does the task instead of the person, employment declines. Where it assists the person, employment grows, including among young workers. Translators are on the first side, radiologists on the second.

In France, the effect already goes beyond IT

France now has its own measurements, and they point the same way. INSEE found in March 2026 that employment of 15-29 year olds was down year on year by 7.4% in IT services, 5.8% in publishing and 3.7% in consulting, against 0.7% for the private sector as a whole, while specifying that this decline cannot be attributed to generative AI alone.

A study by Andreescu and Maurin, reported by Le Monde on September 22, 2026, goes further. Using France Travail job postings, it measures a drop of about 25% in postings across 125 AI-exposed occupations outside IT, including accountants and human resources professionals, compared with low-exposure occupations. In total, this would mean about 5% fewer postings nationwide in early 2026. Two caveats: I was only able to read the press coverage, not the study itself, and this is a relative effect on postings, not on jobs. But it is the first French measurement showing an effect beyond developers.

To be honest about the other side: the doom predictions of 2023 have not come true at the scale of the economy. Goldman Sachs reports that the correlation between AI adoption and employment is “weak and statistically insignificant”. Gartner found that only one customer service leader in five had actually reduced headcount because of AI. The dominant narrative goes beyond the data, in both directions.

For a freelancer: the market is splitting in two

This is the part that concerns me most directly, and the data is clear. Upwork publishes a statistic that sums it all up: generic execution gigs related to generative AI grew 90% in volume, and their pay per gig fell 13%. Over the same period, AI-augmented professional services grew 72% in volume and 22% in revenue per gig. Freelancers positioned on complex work earn 45% more.

More work, worse paid on one side. More work, better paid on the other. Fiverr, the platform most focused on standardized services, went from 4.1 to 2.7 million buyers since 2023, with revenue down 15.5% in the second quarter of 2026.

The dominant mechanism is not job cuts, it is the falling price of execution with no particular added value. Three unrelated markets show the same signature: Indian IT services openly talk about “AI deflation”, with revenue expected to fall 3 to 5%, generic Upwork gigs lose 13% of their value, and the nearshore rate card model is declared obsolete. Volume goes up, unit price goes down.

A freelancer who sells hours of execution is exposed to this mechanism even if nobody ever lays them off. The answer that comes out of three independent sources is the same: move from selling execution time to selling an outcome you are accountable for.

Should you rush into AI? Not exactly

I expected to conclude that yes, obviously, you have to get on board. The data says something more nuanced. Between 84 and 90% of developers already use AI. A behavior shared by nine professionals out of ten is not a competitive advantage, it is a prerequisite. Using AI no longer sets anyone apart.

More interesting: no study measures a penalty for developers who do not adopt it. What exists is a premium associated with AI skills in pay data, which is not the same thing, and employer expectations that create a risk at hiring time. “Adopt or disappear” is a claim, not a measurement. I say this even though I would have preferred the opposite, because it sells better.

The detail that explains everything: AI skills appear in only about 2.5% of US postings, and in front-end roles they are concentrated in senior and staff positions (16 to 18%) against 0.7% for entry-level roles. So the premium rewards a rare, architectural skill, not simply knowing how to use a tool. It is not about knowing how to use Copilot, it is about knowing how to design a system that integrates AI.

What about tomorrow? What the projections to 2034 say

That leaves the question I get asked most: what comes next? Between 2023 and 2026, the major institutions barely changed their exposure figures, but they clearly changed what they conclude from them. The IMF repeats that 40% of jobs worldwide and 60% in advanced economies will be affected by AI, the same figures as in 2024. The International Labour Organization estimates that one job in four is exposed, but that only 3.3% fall into the most exposed category, and that transformation is far more likely than replacement.

You have to learn to read these figures, because the press often distorts them. When McKinsey writes in November 2025 that 57% of hours worked in the United States could be automated with existing technologies, it specifies itself that this is “not a forecast of job losses” but a technical potential. The actual employment forecasts are much lower. Goldman Sachs expects about 9% of US workers to be displaced over ten years, against 6 to 7% in its previous version, an increase that comes from a change in calculation method rather than new data.

Who declines, who grows by 2034

The most useful projection is the one from the US Bureau of Labor Statistics, the first to explicitly factor in the effect of AI occupation by occupation, for the 2024-2034 period.

Two things stand out. The declines remain modest in percentage terms: customer service representatives fall by 5.5%, but since the occupation is huge, that represents 153,700 jobs, the largest absolute figure in the table. And job creation does not offset losses within the same occupations: the 82,500 additional data scientists will not re-employ the customer service representatives. Software developers, for their part, are still projected to grow by 15.8%, or 267,700 positions over ten years.

The real debate among economists

Among serious economists, the gap remains wide. On one side, Daron Acemoglu, winner of the 2024 Nobel Prize in economics, estimates that AI will add no more than 0.66% to productivity over ten years, because only about 5% of tasks can be profitably handed over to it during that period. On the other, a faster scenario if agents manage to take over entire processes, not just isolated tasks. For now, the aggregate data sides with the former: in May 2026, the New York Federal Reserve finds no AI-specific break in job postings, and the Yale Budget Lab concludes there is “stability, not major disruption”.

The only signal leaning toward the fast side concerns young people. That is also where the most spectacular prediction sat: in May 2025, Dario Amodei, the head of Anthropic, said that AI could wipe out half of entry-level white-collar jobs within one to five years and push unemployment to between 10 and 20%. In March 2026, US unemployment stood at 4.3%. The five-year window is not closed, so we cannot say the prediction is wrong, only that it has not come true to date. And Amodei himself now invokes the Jevons paradox: when you automate 90% of a job, the remaining 10% ends up taking all the time.

The signals to watch

Rather than bet on a scenario, I prefer to watch what would tip things from one to the other. Five signals would point to an acceleration:

• the youth employment gap spreading to 26-34 year olds, then to those over 35 (Stanford already sees faint traces of it up to age 34);
• rising unemployment, and not just slower hiring, in exposed occupations;
• a break in overall productivity above its trend, which Stanford's tracking does not see yet;
• automated uses overtaking assisted uses in more and more occupations, which Anthropic already observes for code in its API traffic;
• a clear break after 2022 in the job posting series tracked by the New York Fed.

One indicator is already making a lot of noise, and it should be treated with suspicion. According to the firm Challenger, AI was cited for 5% of announced US job cuts in 2025, and for about 24% through July 2026, the top reason given five months in a row. But unemployment did not move over the same period, and several analyses show that companies attribute to AI job cuts decided for other reasons. This figure measures what companies say, not what AI does.

Conversely, other signals would rather point to a plateau: enterprise adoption stalling, which Stanford is starting to observe, rehiring like at Klarna, and a youth gap that stabilizes. In the long run, the World Economic Forum's 2025 report expects 170 million jobs created against 92 million displaced by 2030, with this warning: those who lose their jobs will not necessarily be the ones who take the new ones.

What I take away

Will AI replace professions or transform them? On the data available in September 2026, it is transforming many, weakening a few through price rather than replacement, and in almost every exposed profession, it has first closed the entry door. Translators, copywriters and bookkeepers are genuinely declining. Radiologists, customer service agents and experienced developers are holding up much better than was predicted in 2023.

For developers, the timeline settles one question: it was neither ChatGPT nor Cursor that brought the market down, it was the price of money. Coding agents arrive after the fall, and they shape the recovery: more senior, more demanding, with value shifting toward specification, architecture, security and review, while writing code gets automated. That is the exact opposite of the idea that we would no longer need developers.

For a beginner, the truth has to be said: the entry door has closed, and the French recovery, as currently forecast, skips juniors. APEC expects +4% in managerial and professional hiring in 2026, but only +1% for young graduates in those roles.

Looking ahead, the signals I watch are not the headlines. They are the measured shift in coding usage, in API traffic, from an assisted mode to an automated mode, and the youth employment gap that could spread to people in their thirties. These are not job losses yet. They are the only two leading signals pointing in the direction historically associated with decline, and they are the ones to watch, not the announcements.