Česky
Slovensky
EnglishFor a decade the warning was always the same: robots would come for truck drivers, assembly-line workers, and cashiers. Then late 2022 arrived, ChatGPT landed, and the ground shifted first under copywriters, illustrators, and junior programmers. The exact jobs that science fiction promised would stay safe the longest.
That reversal is instructive on its own. It shows how unreliable predictions about what technology will "take" really are, and it hints at why "will AI replace my job?" is almost always the wrong question to ask.
So what do the numbers actually tell us? From the famous Oxford study of 2013 to the payroll data of 2025, the picture has flipped more than once. And how it flipped points to some surprisingly practical advice for your own career.
The predictions that never came true
In 2013, economists Carl Benedikt Frey and Michael Osborne of Oxford published a study that traveled the world. Their conclusion: roughly 47% of US jobs sat at "high risk" of automation over the next ten to twenty years. The figure showed up in headlines, in political speeches, and in more than one essay about the end of work.
The method had cracks in it. The authors hand-labeled just 70 of more than 700 occupations as automatable or not, then trained a model on that small sample to fill in the rest. Makeup artists, school bus drivers, and hairdressers all ended up flagged as endangered.
The deeper problem was conceptual. Frey and Osborne reasoned about whole occupations rather than the parts they are built from. And the parts turn out to matter enormously.
Three years later, Melanie Arntz, Terry Gregory, and Ulrich Zierahn took a different route for the OECD. Instead of whole professions, they examined the individual tasks inside them. The result across 21 OECD countries: on average only about 9% of jobs were highly automatable (Arntz, Gregory, and Zierahn, 2016). The gap between 47% and 9% is not an arithmetic error. It is the difference between asking "will this profession vanish?" and "how many of its tasks can a machine handle?"
The specialists were skeptical too. When one survey asked AI and robotics experts to estimate exposure, they guessed roughly a fifth fewer endangered jobs than people outside the field did. Enthusiasm for apocalyptic numbers was always highest where nobody had seen the technology up close.
And what actually happened? The mass disappearance of work never came. Analysts at the think tank ITIF noted dryly in 2022 that the predicted loss of 47% of jobs had failed to materialize; US unemployment had just been sitting at historic lows. Automation was reshaping work, not deleting it.
The turn nobody expected
For a long time an unwritten rule held: machines take the work of the hands, not the head. Manual and routine tasks could be automated; creative and analytical work was safe. Generative AI turned that rule nearly upside down.
In 2023, Tyna Eloundou and colleagues (in a paper aptly titled "GPTs are GPTs") estimated that for about 80% of US workers, language models could touch at least 10% of their work tasks. For roughly 19% of workers, more than half the tasks were in play. The striking finding lay elsewhere, though: the most exposed jobs were educated, well-paid office roles, not the lowest-paid ones.
This is the counterintuitive flip. A copywriter, an analyst, a junior programmer, or a designer does things a language model learned to handle surprisingly well, because text and code are exactly what it trained on. A plumber, a caregiver, or a line cook does things that do not fit into text at all.
Picture a graphic designer who in 2022 still made a living on small jobs: banners, simple illustrations, photo touch-ups. A chunk of that work can now be done by an image generator for a few dollars. Meanwhile the electrician you call when half the outlets in your building die is booked out for a week if you are lucky. Ten years ago most people would have bet on the reverse.
A task is not the same as a job
Almost every job is a bundle of different tasks. An accountant does not just enter numbers; she also explains to a client why the tax came out the way it did, handles the exceptions, and stands behind her signature. AI can shave off the first part and struggles with the rest. That is why most serious estimates now talk about transformation rather than extinction.
In its updated index for 2025, the International Labour Organization reports that about one in four workers worldwide holds a job with some exposure to generative AI, yet only around 3.3% of employment falls into the highest category. One detail stands out: for women the figure is 4.7%, for men 2.4%. Clerical and administrative work, still disproportionately done by women, is more exposed.
Here is a question worth sitting with. How much of what you do in a day is really text, a spreadsheet, or code that could be dictated to a machine? And how much is judgment, trust, and being present for someone who needs you? That ratio tells you more about your future than the title printed on your business card.
The labor market as a whole is not shrinking because of AI, either. In its Future of Jobs 2025 report, the World Economic Forum estimates that by 2030 about 170 million new roles will appear and 92 million will disappear, a net gain of roughly 78 million jobs. It adds that nearly 40% of the skills employers ask for will shift within five years. Work does not vanish. It moves.
Where AI still hits a wall
There is an old observation from robotics that reads almost like prophecy today. In the 1980s, Hans Moravec noticed a paradox: teaching a computer logic, math, or chess is fairly easy, but giving it the sensory and motor skills of a one-year-old is fiendishly hard.
Roboticist Ken Goldberg called it the "100,000-year gap." A language model learns from a few terabytes of text and reads like an adult. No robot comes close to gathering as much data about the physical world as a human collects by age three, when a toddler can already step around scattered toys without tripping over the threshold.
From this follows a fairly clear map of resilience. The best-positioned jobs combine physical dexterity in unpredictable environments with dealing directly with people. The table below simplifies things; reality is of course messier.
| Type of work | AI exposure | Why |
|---|---|---|
| Routine writing, translation, simple code | High | Text and code are exactly what the models train on |
| Data analysis, reporting, legal research | Medium to high | AI drafts first; a person checks and decides |
| Managing people, negotiation, advising | Low to medium | Trust and reading context resist automation |
| Trades, maintenance, on-site installation | Low | Physical dexterity in unpredictable spaces (Moravec's paradox) |
| Direct care for people (health, children, elders) | Low | Presence and touch do not fit into text |
None of this means AI leaves these jobs untouched. A nurse can use it to transcribe documentation, a tradesperson to schedule jobs. The core of the work stays out of the model's reach, though: presence at a patient's bedside, or the hand that fixes a burst pipe in a cramped shaft.
The first warning signs
That this is more than theory shows up in the first hard data. Economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen of Stanford (2025) combed through the payroll records of millions of Americans supplied by ADP. Among the youngest workers, ages 22 to 25, in the most exposed occupations, employment fell by roughly 13% after late 2022. Among older people in the same occupations, it stayed stable or kept rising.
That pattern makes sense. A junior often does exactly what AI handles first: research, first drafts, simple code. A more experienced person adds judgment and accountability, which are harder to automate. The authors also note that the decline shows up mainly where AI replaces work rather than where it complements it.
Freelance platforms tell a similar story. After ChatGPT launched, analyses found a noticeable drop in demand for automatable writing and coding, while demand for manual and physical jobs held firm. This is not a blanket apocalypse. It is a migration that begins at the edges, among the most vulnerable tasks and the least experienced people.
Newer figures point the same way while adding nuance. Anthropic's Economic Index (March 2026 report) tracked how its AI is used across the economy and found that collaboration still edges out full automation, roughly 55% to 42%. The catch: the model tends to absorb the higher-skill, more analytical slices of a role, leaving the lower-skill remainder behind and quietly deskilling the job. Absorbing tasks is not the same as absorbing people.
Behind all of this sits rising use. US data show that by the end of 2024 about 30% of employees used language models at work, and a year later closer to 38%. The share of work time spent with generative AI rose only modestly, from around 4% to just under 6%. In other words, the tool is spreading fast, but most people still use it as a helper on a slice of a task, not as a stand-in for a whole day of work. That is precisely the difference between augmentation and replacement.
What this means for your own career
The practical lesson is uncomfortable and, in its way, freeing: stop thinking in job titles and start thinking in skills and tasks. "I am a copywriter" is a fragile claim. "I can grasp a stranger's business quickly and explain it clearly to other people" is far more resilient, whatever profession you happen to do it in.
The second point is complementarity. The safest place is not far from AI but right beside it. People who can brief the model well, check its output, and take responsibility for it will be in more demand than those who ignore it, and than those who trust it blindly. That is why the World Economic Forum lists working with AI and analytical thinking among the fastest-growing skills.
So what do you do about it? A few starting points:
- Break your own job down into tasks and mark how many are purely text, spreadsheet, or code.
- For the exposed tasks, decide whether you want to do them faster with AI or move the center of gravity of your work elsewhere.
- The work that pays off most is what the machine lacks: dealing with people, judgment in murky situations, physical skill.
- Try new tools before your employer or the market pushes you to.
And if you genuinely have no idea which of those more resilient areas you sit closest to, start with yourself. The RIASEC personality test maps six kinds of work environment, from the practical and technical to the social, and suggests where you might do well, regardless of what the position happens to be called this year.
The question, after all, was never whether AI would replace people. It was whether the people who can work with AI would replace the ones who cannot. And each of us answers that alone, through what we decide to do over the next twelve months.