The headline -- IT jobs have fallen by 73% -- is not the story. The story is that the entry-level economic model of software work is under pressure.
In May 2026, Swedish media reported that the number of open IT jobs in April was 73 percent below the peak from December 2021, based on Arbetsförmedlingen data reported by Aftonbladet and summarized by Omni. Arbetsförmedlingen publishes historical time series for open jobs by occupational area through April 2026, so the underlying category exists in the official statistics, even if the newspaper number itself should be read as a vacancy signal rather than a complete labor-market diagnosis (Arbetsförmedlingen).
Job ads are not employment. A fall from December 2021 is not a fall from normal; it is a fall from a pandemic-era peak shaped by cheap capital, accelerated digitalization, over-hiring, consultancy expansion, and speculative demand for technical capacity. The Swedish consulting market also came out of record years in 2021 and 2022 into a much weaker 2025, with growth, headcount, and profitability slowing sharply according to Cinode’s Konsultkollen.
So the claim should be narrow and hard: AI did not cause every missing IT vacancy. The excess-demand regime ended. What returns from that regime will not have the same shape as what disappeared.
That is the continuation of the coordination shift. When AI makes execution cheaper, the scarce work moves toward intent, judgment, verification, integration, operability, domain understanding, and coordination. The labor market then starts repricing capacity. Not all at once, not cleanly, and not without cyclical noise, but visibly enough that the signal now appears in Sweden as well.
The earlier US signal
The US saw the broad shape earlier. Indeed’s Hiring Lab described a prolonged tech hiring freeze, with US tech postings weak since mid-2023 and down 36 percent from February 2020 by July 2025, while also warning that there was no clean "smoking gun" proving AI as the sole cause because much of the decline began before ChatGPT became public (Indeed Hiring Lab).
The important detail was not that "tech jobs vanished." It was that the recovery was selective. Indeed found that most tech titles were below pre-pandemic levels in early 2025, software engineers were down 49 percent, while AI-related roles were among the few areas still above early-2020 levels (Indeed Hiring Lab).
The age gradient made the signal sharper. A Stanford Digital Economy Lab paper found substantial employment declines for early-career workers aged 22-25 in occupations most exposed to AI, including software developers and customer service representatives, while employment for more experienced workers in the same occupations remained stable or continued to grow (Stanford Digital Economy Lab). The same paper found that young workers in the most AI-exposed occupations experienced a 6 percent employment decline from late 2022 to September 2025, while older workers in those occupations grew by 6-9 percent (Stanford Digital Economy Lab).
That is not a generic anti-tech pattern. It is an entry-ramp pattern.
Stanford also separated automation from augmentation. Employment declined for young workers where AI use was primarily automating work, but not where AI use was primarily augmenting workers (Stanford Digital Economy Lab). Anthropic’s labor-market work makes the same analytical distinction by separating theoretical exposure from observed professional usage, and by weighting automated use differently from augmentative use (Anthropic).
The pattern is not that AI replaces software engineers. The pattern is that AI changes which parts of software work are worth hiring for.
The Swedish echo
Sweden now shows a similar age-gradient signal.
Örebro University reports that Swedish register data from 2019 to June 2025 shows employment among 22-25-year-olds in AI-exposed occupations declining by 5.5 percent within the same employers after the launch of ChatGPT, compared with less-exposed occupations, while employment for workers over 50 in the same exposed occupations increased by about one percent (Örebro University). Magnus Lodefalk’s interpretation is the key line: generative AI does not primarily reduce the number of jobs; it changes who gets them (Örebro University).
TechSverige points in the same direction from the demand side. Demand for AI skills in Swedish job advertisements has increased by 328 percent since 2016, AI is now mentioned in almost twice as many occupations as eight years earlier, and one in four companies uses AI compared with one in ten the year before (TechSverige). The same report summary says AI is raising barriers to entry for young people and recent graduates as work processes are automated and roles change (TechSverige). A later Akavia/TechSverige report makes the institutional point more directly: recent graduates face a weaker labor market, the first qualified job takes longer to reach, and AI is rapidly changing role content and skill requirements (TechSverige).
The aggregate labor market can still look strong while this happens. ManpowerGroup’s Q1 2026 survey gave Swedish IT and Tech a positive employment outlook of +18 percent, although lower than the total Swedish labor-market outlook of +30 percent and one of the more muted sector forecasts (ManpowerGroup). In Q2 2026, ManpowerGroup reported a strong total Swedish employment outlook of +39 percent and positive forecasts across all surveyed sectors, including IT and Tech at +42 percent (ManpowerGroup). Almega’s May 2026 review also cautions that Nordic employment effects from AI remain limited so far, while noting that recent Swedish evidence points to negative effects for young workers in highly exposed occupations and positive effects for older workers in the same occupations (Almega).
That combination is the point. A strong aggregate labor market can coexist with a broken entry ramp. Positive hiring sentiment can coexist with fewer generic IT vacancies. More demand for AI skills can coexist with fewer first jobs.
What is being repriced
The work compressed first is not necessarily the work with the highest title. It is the work with the weakest claim to judgment.
Basic implementation, first-draft code, routine support, simple analysis, generic testing, ticket throughput, configuration, documentation drafts, and "look it up and assemble something" tasks are all easier to route through AI. These tasks were never worthless. They were the paid learning substrate of the profession.
This is why the problem is not solved by telling graduates to "learn AI." The issue is not whether they can use a chatbot. The issue is whether the organization still has a paid path by which a novice can become someone capable of supervising machine output.
That path is already under pressure. Anthropic’s randomized study on coding skill formation found that developers using AI assistance to learn an unfamiliar Python library scored 17 percent lower on a later mastery quiz than those who coded by hand, with no statistically significant speed gain on average (Anthropic). The important detail is not that AI makes people worse. The important detail is that skill formation depended on how AI was used. Participants who used AI to build comprehension preserved more learning; participants who delegated the work lost the learning path (Anthropic).
The labor-market version is straightforward. If companies delete junior work and replace it with AI output reviewed by seniors, they may get short-term productivity. They also delete part of the mechanism that creates future seniors.
This is not sentimentality about junior roles. It is production-system accounting.
The institutional error
The error would be to conclude that young engineers are obsolete. They are not. But the default junior value proposition has weakened. "I can take tickets and produce code" is less compelling when a senior engineer with good tools can generate, test, and revise several implementation paths in the same window. "I can learn fast, understand systems, test assumptions, verify output, reason about failure, and operate inside a real domain" is stronger.
The error would be to conclude that companies can simply hire fewer juniors and buy more AI. They can do that for a while. It may even look rational in quarterly numbers. But if the organization depends on humans for context, accountability, debugging, user judgment, architectural taste, and operational responsibility, it still needs a way to produce those humans.
The error would be to conclude that universities should respond with generic AI literacy. The market does not need more people who know that AI exists. It needs people who can work in the new production system: domain framing, constraint handling, verification, observability, integration, security, and evidence-based iteration. TechSverige’s own framing is close to this: AI skill demand is rising, but the transition from education to first qualified work now needs stronger bridges, internships, and earlier practical exposure (TechSverige).
The entry ramp has to be redesigned, not defended as nostalgia.
The signal
The Swedish 73 percent headline is useful because it is crude. It marks the end of a regime. It does not prove that AI caused the whole fall. It does not prove that IT employment is collapsing. It does not prove that software is less important.
It shows that the old market for advertised IT capacity has been repriced.
The US saw the same structure earlier: fewer broad tech postings, fewer software-engineer openings, persistent weakness after the pandemic boom, and a sharper hit to young workers in AI-exposed occupations (Indeed Hiring Lab, Stanford Digital Economy Lab). Sweden now shows the same age-gradient signal in register data and the same demand shift toward AI-related competence (Örebro University, TechSverige).
The conclusion is not apocalypse. It is more operational and more uncomfortable.
Software work remains valuable. Generic software capacity is less valuable. Junior work remains necessary. The old economic container for junior work is weakening. AI increases output, but it also increases the premium on the humans who can decide what output should exist, whether it is correct, whether it fits, whether it can be operated, and whether it should be trusted.
Companies can consume the apprenticeship layer faster than they can recreate it.
That is the risk now visible beneath the vacancy chart.