In depth
AI and jobs: what really changes
Neither job-stealing robots nor a miracle: tasks, occupations, studies quoted with their caveats, and concrete habits to get ready.
Reviewed on August 22, 2026
Tasks, not jobs: what AI automates today
In 2026, what generative AI does well is fairly specific: producing a first draft (email, meeting notes, job ad), summarizing and translating, sorting documents and extracting information from them, writing and fixing code, answering the repetitive questions of a customer service desk. What it does badly remains broad: anything that needs hands, a physical presence, someone accountable, or judgment in a situation nobody has seen before. That is why economists think in tasks rather than jobs: a position is a bundle of tasks, and AI takes some of them, rarely all. In real usage measured by Anthropic on its own assistant (Economic Index, November 2025 data), about half of conversations are 'augmentation' (the human stays in charge, the AI helps) and a bit under half are 'automation' (the AI carries out the task alone). So the right question is not 'will my job disappear?' but 'which tasks in my job change, and what do I do with the time freed up?'.
Read the discovery version
Picture a job as a big bag full of little marbles: each marble is one task. An AI is very good at grabbing some marbles (writing a first draft, summarizing a text, translating, sorting documents) and hopeless at others (comforting someone, fixing a leaking pipe, deciding when the situation is new and messy). It takes a few marbles out of the bag, but it does not take the bag. The job stays, it just changes shape.
Which occupations are most exposed, and why
The major studies agree on the 'who', less on the 'how much'. The IMF (January 2024) estimates that about 40% of jobs worldwide are exposed to AI, close to 60% in advanced economies, about half of them more likely complemented than threatened. The ILO and Poland's NASK institute (May 2025) count 1 in 4 jobs potentially exposed to generative AI (34% in high-income countries), but only 3.3% in the most exposed band: data entry, payroll, clerical work. Women are over-represented there (27.6% of their jobs exposed versus 21.1% for men), because they hold more of these administrative positions. The OECD (Employment Outlook 2023) placed 27% of jobs in occupations at high risk of automation, all technologies combined. For France, Aghion, Bergeaud, Bunel and Delbouve (July 2025) estimate 5% of jobs directly replaceable in the medium term, 13.4% 'optimizable' and 81.6% neither threatened nor directly affected. Least exposed: construction, care, hospitality, agriculture, transport.
Read the discovery version
Jobs where you mostly work with words and numbers on a computer (writing, filing, counting, filling in spreadsheets) are the ones AI can help the most. Jobs where you use your hands, your body or your heart (plumber, nurse, teacher, cook) are much less affected. Careful: 'exposed' does not mean 'replaced'. It means AI can help, or change the way the work is done, the way the calculator changed the accountant's job without making it vanish.
What appears: skills and new roles
The World Economic Forum (Future of Jobs, January 2025, survey of over 1,000 employers) projects 170 million jobs created and 92 million displaced by 2030, a net +78 million and a churn of 22% of positions. The fastest-growing roles are technical (AI and data specialists, cybersecurity, software development), but the biggest volumes come from elsewhere: care, delivery, teaching, farming, driven by demographics and the green transition. Around AI itself, in-between jobs appear: model trainers and evaluators, annotators, AI compliance officers, 'translators' between business and tech, integrators who plug a model into a company's tools. On skills, employers rank analytical thinking first (7 in 10 call it essential), then AI and data, cybersecurity and technology literacy, with creative thinking, resilience and curiosity rising. In France, the Aghion et al. study (2025) observes that the most exposed firms hired markedly more AI specialists as soon as ChatGPT arrived.
Read the discovery version
When cars arrived, blacksmiths had fewer horses to shoe, but we invented mechanics, truck drivers and driving instructors. With AI, new jobs are being born too: people who teach AIs, who check what they say, who show others how to use them without getting fooled. And most jobs stay the same, with a new tool in the box, like when the carpenter got their first electric drill.
Getting trained: the right reflexes, depending on who you are
For an employee: pick two repetitive tasks in your role and test them with an assistant for a month; learn to verify (sources, figures, legal points) rather than to trust; write down what works and share it with the team; ask for , in France via the personal training account or the company skills plan; nurture what AI does not do, relationships, accountability, craft. For a student: learn the fundamentals thoroughly, because AI hides gaps as much as it fills them; use it as a tutor that explains, not as a ghostwriter; build a portfolio of work you can stand behind; aim for the crossing of a field and technology rather than technology alone. For a manager: map the team's tasks, not its positions; experiment with volunteers, measure, then scale; train everyone, the AI Act has made it a duty since February 2025; involve staff representatives early; redesign roles rather than cut headcount at the first time saving.
Read the discovery version
The best way not to be afraid of a new tool is to try it, the way you learn to ride a bike: you start small, on flat ground, with someone watching. For a worker, a student or a boss it is the same: try AI on one real little task, check what it says, keep what you do that is special (your hands, your ideas, the people) and stay curious. Nobody asks you to know everything, just not to stay on the side of the road.
Rights and rules: what the law says at work
The EU AI Act (2024) classifies as 'high risk' the systems used to recruit, screen applications, evaluate, promote, dismiss, allocate tasks or monitor employees (Annex III, point 4): quality data, logs, human oversight, information to the people concerned. Two rules have applied since 2 February 2025: a ban on emotion recognition in the workplace, and a duty to train staff in AI. The rest of the 'high risk' regime was pushed back to 2 December 2027 by the Digital Omnibus (Regulation 2026/1744, in force since 27 July 2026). In France, existing law applies right now: the GDPR (Article 22) forbids a significant decision based solely on automated processing without safeguards (human intervention, the right to contest); the Labour Code requires informing the candidate of the recruitment methods before they are used (L1221-8) and bans any hidden data collection (L1221-9); the works council (CSE) must be consulted before a new technology is introduced (L2312-8). In September 2025, the Paris court suspended the roll-out of an AI platform, pilot phase included, for lack of prior consultation of the CSE.
Read the discovery version
In a company, a machine is not allowed to decide on its own to hire you, fire you or grade you: a human has to look, be able to say no, and explain. In Europe, special rules have been written for this, like the rules of a board game: an AI that sorts candidates or watches employees has to be careful, transparent and checked. And before installing a new AI at work, the company has to talk it over with the employees' representatives first.