I've felt afraid of AI too. Not movie-style fear, but something more everyday: watching a machine write, summarise, code, translate or answer in seconds, and thinking that maybe part of human work stops making sense if this keeps going.
That vertigo isn't an overreaction. It arrives alongside very fast adoption: according to Stanford's AI Index 2026, generative AI is already used in at least one business function at 70% of organisations, and most reports agree that employees are adopting it at a pace not seen with any previous technology, not even the internet.
And the emotional climate is ambivalent. Stanford records that global perception of AI's benefits improved slightly between 2024 and 2025, but at the same time more than half of people say AI makes them nervous. Pew found in 2025 that, among US workers, concern outweighs hope about its future impact.
This article is about putting order into that feeling. Not repeating that "everything will be fine", nor selling you an inevitable dystopia. It's about answering a more useful question: which part of that fear is well founded and which part comes from looking at AI as if it were a total replacement for people, when in practice it still works more like a technology that automates chunks, speeds up specific tasks and pushes us to redesign professions.
What I think about this
Being afraid of AI is normal. Thinking human work will disappear overnight isn't.
The available data paints something both more uncomfortable and more realistic. AI is already altering tasks, processes and entry routes into the labour market. But the best studies don't describe a mass extinction of employment, but an uneven transition: some jobs change, others shrink, others become more productive, and some profiles —especially the most junior in routine cognitive tasks— are indeed starting to feel the pressure.
The ILO is clear: a significant share of the world's workers are in occupations with some degree of exposure to generative AI, but most of those jobs will be transformed, not eliminated. The World Economic Forum projects that, by 2030, labour-market transformation could create around 170 million jobs and displace around 92 million, with a net positive balance. That doesn't guarantee the adjustment will be comfortable or fair for everyone. But the base scenario isn't "jobs disappear". It's "jobs change, and who's prepared to do them changes too".
Nor should we settle for cheap reassurance. In that same WEF survey, 41% of employers said they plan to reduce headcount as AI automates certain tasks, while 77% expect to bet on upskilling. AI doesn't arrive as an abstract wave of innovation. It arrives as concrete pressure to reorganise teams, redefine roles and make certain human and technical capabilities more valuable at once.
Fear of AI is logical, not irrational
Fear of AI is often described as ignorance or resistance to change. It isn't.
A Max Planck Institute study with more than 10,000 participants from 20 countries found that fear of AI replacing humans varies a lot by profession and by what people consider essential in each role. We're not equally unsettled by an AI that drafts a note and an AI that judges, cares for or leads. Fear appears mainly when we think a tool could invade spaces where we value deeply human qualities: honesty, fairness, empathy, responsibility.
And there's something less often said: direct contact with these tools doesn't remove the ambivalence. In a preregistered MIT experiment with 444 professionals using ChatGPT on writing tasks, contact with the tool increased both the sense of efficacy and the concern about automation. In other words, seeing that something really works doesn't just excite you; it also forces you to reconsider the value of what you do.
That's why much of the public conversation is poorly framed. The problem isn't that people "fear AI too much". The problem is that this fear gets channelled into too crude a question: will AI take my job?
It's almost never the right question. The useful question is usually another: which part of my work is repeatable, codifiable and easy to speed up, and which part depends on judgement, context, trust, deep creativity or responsibility? That's where the conversation stops being apocalyptic and starts being practical.
What the data is showing at the entry door
If we look at the numbers coldly, we're not seeing an implosion of employment. Anthropic, in a March 2026 analysis by Massenkoff and McCrory on the occupations most exposed to AI, found no systematic rise in unemployment among the most exposed workers since late 2022. The OECD agrees: for now there's little evidence of mass job losses caused directly by AI, and the observed effects run more through changes in work quality, organisation and tasks than through direct destruction of the role.
That doesn't mean nothing is happening. It means change appears first where almost all labour change appears: at the entry door.
Stanford's AI Index 2026 records a roughly 20% drop in employment of software developers aged 22 to 25 since 2024, while employment of developers aged 30 and over at the same companies keeps growing between 6% and 12%. The same pattern appears in other occupations with high AI exposure: customer service, accounting, marketing, administrative support. Brynjolfsson and co-authors document similar drops in the 22-to-25 group in high-AI-exposure roles, using real ADP payroll data, not surveys.
This is one of the most important nuances in the debate. We're not seeing, for now, AI emptying entire offices. What is starting to show is that the most codified tasks of some junior roles no longer justify as much entry-level hiring.
And that's serious for a reason rarely discussed. Junior work isn't only for producing. It's for learning, making mistakes, picking up context and becoming, over time, someone senior. If that ladder starts losing rungs, the immediate problem isn't youth unemployment. It's where the seniors will come from ten years from now.
A necessary note against sensationalism: even with very high corporate adoption, Stanford points out that the deployment of fully autonomous AI agents was still in single digits in 2025 across almost every business function. AI is coming in fast, yes, but the "everything is automated already" version runs well ahead of observable reality.
In Andorra and Spain we don't yet have aggregate data like Stanford's or Anthropic's. The business fabric is different: SMEs dominate, the tech sector is proportionally smaller and hiring young people into cognitive roles has never had the volume it does in the United States. That's why the effect here doesn't show up the same way in the statistics.
But in the conversations it does. Every week I talk with small business owners in the Principality and in Catalonia doing exactly what the American data describes: they used to hire the next administrative assistant or the next intern almost by default, and now they first ask whether a tool can hold up that part of the operation. That doesn't appear in any INE report or in the AI Index. But it's happening. I wrote about it in more detail in another blog post: many companies here won't lay anyone off because of AI. They'll simply stop hiring the next person.
What Brynjolfsson helps us understand about productivity
Before generative AI, Brynjolfsson, Rock and Syverson already explained something they called the Productivity J-Curve: general-purpose technologies produce macro gains late, because at first all the investment goes into reorganising processes, training teams and creating new ways of working. We see impressive technologies in demos while, meanwhile, the aggregate statistics take time to move. The OECD still uses this framework to explain why AI's macroeconomic leap hasn't clearly appeared yet.
After generative AI arrived, the picture changes at the micro level. In a study published in The Quarterly Journal of Economics in 2025, Brynjolfsson, Li and Raymond measured the effect of a generative assistant on 5,172 support agents. Average productivity rose 15%. Among the least experienced and lowest-performing profiles, the improvement reached 30%. Agents with two months' tenure and AI help performed like agents with six months and no AI.
The takeaway is powerful: AI doesn't only speed things up. It can also spread best practices and shorten learning curves.
The MIT experiment by Noy and Zhang points the same way, in a more bounded context. On professional writing tasks, participants with access to ChatGPT took 37% less time, improved quality by 0.45 standard deviations, and the tool benefited those starting from a lower level the most. It compressed performance inequality rather than widening it.
This runs counter to the fatalistic story that AI only benefits a technical elite. Well implemented, it also makes ordinary people more capable at fairly ordinary tasks.
But the other half of the message matters just as much. In Dell'Acqua and co-authors' study on the "jagged technological frontier", consultants performed better with GPT-4 on tasks inside the model's capability frontier. On tasks outside that frontier, people with AI spent less time, produced answers that looked more coherent, and got the substance wrong more often.
Stanford sums up the idea: the most consistent productivity gains appear in structured, measurable work; they're smaller when work demands deep reasoning; and over-reliance can penalise long-term learning.
AI doesn't replace judgement. Sometimes it makes it even more necessary.
Competitive advantage becomes human again
Here, in my view, is the most hopeful part.
For years, many professional advantages depended on access to scarce resources: more information, more production capacity, more hands, more budget, more tools. AI is making many of those capabilities cheaper at a brutal pace. Stanford noted in 2025 that the inference cost for a system at GPT-3.5's level fell more than 280-fold between November 2022 and October 2024. And in 2026 it adds that generative AI adoption reached 53% in barely three years, faster than the PC or the internet.
My reading is direct: if more and more people can access "intelligence on demand", the real differentiator shifts back towards what isn't bought so easily.
And what isn't bought so easily? Judgement. Context. The ability to decide what's worth doing and what isn't. Taste. Empathy. Leadership. Trust. Responsibility.
The WEF insists that, alongside technical skills, competencies such as analytical thinking, resilience, leadership and collaboration will keep mattering. And the Max Planck study reminds us that precisely the professions where we most fear replacement are the ones where those human traits weigh most. The paradox is a nice one: the more accessible artificial intelligence becomes, the more visible the specific value of human intelligence becomes.
This also opens real opportunities for individuals and small teams. If a tool lets you research faster, write a better first draft, summarise documents, help you code or prepare a presentation, the threshold to compete drops. You don't need a large company's structure to produce work that used to require more time, more budget or more staff.
I say "can" on purpose. It's a reasonable inference from cheaper, more widely spread capabilities, not an automatic promise. Without judgement and without review, AI doesn't make you stronger. It makes you faster. And being faster isn't always being better.
An uncomfortable note to close this point. The IMF warns that demand for new skills is growing, and that in advanced economies around one in ten vacancies already requires at least one new skill. Those skills are associated with higher pay. But the benefits aren't shared equally, and tensions can appear for young workers and for part of the exposed administrative work without complementarity. Learning to work with AI can improve your position. Arriving late or without support can make you more vulnerable.
How to adapt without living in fear
If this worries you, my advice isn't "relax". It's something more useful: move with judgement.
AI is no longer a curiosity, and ignoring it entirely doesn't protect you either. What protects you is building a healthy relationship with the tool. Four ideas.
Don't delegate your brain where it doesn't belong yet. Pew found that people who use chatbots at work see them as more useful for going faster than for improving quality. And the jagged frontier study showed that AI can deliver very convincing yet wrong answers. Use it to accelerate the first draft, the outline, the summary, the exploration. Not as an automatic replacement for your final judgement.
Learn out loud. The OECD observes that training and consulting workers are associated with better results when AI enters an organisation. When AI is rolled out as a secret shortcut or a silent replacement, it generates anxiety. When you work with training, participation and clear rules, it produces better experiences and more trust. That logic holds at the individual level too: if you learn to use a tool consciously, your relationship with it changes.
Reinforce exactly what AI forces us to value more. I wouldn't obsess over learning twenty new tools every month. I'd obsess over combining basic AI literacy with skills the machine doesn't sign off for you: writing in your own voice, explaining well, spotting errors, asking better, connecting ideas, understanding people, owning consequences. The WEF doesn't say technical skills don't matter. It says technical and human skills will increasingly go together.
Accept an uncomfortable but liberating truth. Maybe AI will indeed drain value from part of what we do. But that doesn't mean it drains us. What gets cheapest with AI is the predictable, the formalisable and the repeatable. What gains value is choosing well, reviewing well, deciding well, supporting well and answering for the result.
My thesis, in the end, is this: AI doesn't reduce the value of people. It reduces the value of doing by hand what was already easily standardisable.