I read, over and over, that artificial intelligence brings the apocalypse. The apocalypse was also coming with Gutenberg's printing press in 1450, also coming with the steam engine in 1769, also coming with electricity in the home, with radio, with television, with the internet, with smartphones. Today we can barely picture the world without those things. The apocalypse never arrived. What did arrive — because that part always arrives — was a quiet transformation of the people living on the other side. It changed how they remembered. It changed how they decided. It changed how they related to each other. It changed how they worked. And almost nobody was looking there, because all the noise gathered where it always gathers: on the cinematic.
Today I want to see what is actually true in the current fear. As methodically as I can manage.
The box office and the long shot
Cinema and listicles have turned the fear of AI into a genre with two stable leads: Skynet and the killer robot. Skynet, the intelligence that wakes up and decides to end humanity. The killer robot, the armed body executing orders with no brake. They are powerful scenes. They work on screen. They work in a headline. And they work, above all, because they soak up every drop of available attention while leaving in shadow other things that are in fact happening.
Let's get the first thing straight. It isn't that those risks don't exist as hypotheses. The technical discussion about alignment — how to make sure a very capable system does what we want without jumping the fences — is real, serious people are doing it, it fills thousands of pages of academic literature and funds entire labs. Stuart Russell, Nick Bostrom, the safety teams at Anthropic and OpenAI, the authors of papers at NeurIPS and FAccT: all of them take that debate seriously. The problem isn't that the debate exists. The problem is that the debate, once translated for a mass audience, has been boiled down to a single image — Terminator — that has nothing to do with what an alignment risk would actually look like, and that drains attention away from another class of problem already happening, today, to millions of people.
While the camera frames the T-800 sprinting down a highway, the long shot shows something else. And that is what this article is about.
What is actually happening: concentration
Let's do the numbers. In 2026, the technical frontier of generative AI is held by between four and six companies: OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, DeepSeek. Some in the United States, one in China. Global investment in AI reached 581.7 billion dollars in 2025, according to Stanford's AI Index 2026 — double the year before. Investment in generative AI specifically grew by more than 200% year over year — more than tripling — to absorb almost half of all private investment in AI. Global adoption of generative AI has reached 53% of the world's population in under three years, a speed greater than the personal computer's and greater than the internet's. And that whole flow of investment and use passes through the infrastructure of fewer than ten companies.
To grasp the problem, let's translate. Twenty years ago, if you had a question, you opened a book, a newspaper, a search engine with thousands of indexed sources, you talked to a colleague, you went to a library. Today, a growing share of the questions users ask passes through a conversation with an assistant — Claude, ChatGPT, Gemini, Grok. That conversation is a single channel. When you ask an assistant, you get an answer synthesised by a specific infrastructure trained by a specific company under specific criteria. And that single channel is, today, controlled by fewer companies than the number of carmakers in any mid-sized European country.
We aren't talking about Skynet. We're talking about something boring and genuinely serious: the concentration of the cognitive input of hundreds of millions of people in the infrastructure of six corporations. This is not a distant hypothesis. This is what's happening as you write. It takes no imagination. It takes looking.
What is actually happening: delegation
The second long shot is less visible but more intimate. It's described by people who have worked in the right lab. In 2025, a team at the MIT Media Lab led by Nataliya Kosmyna published Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. It's the first serious empirical study of how the sustained use of an AI assistant affects brain activity during a demanding cognitive task such as writing.
The method is solid without being final. Fifty-four participants, split into three groups — writing with only their brain, writing with a search engine, writing with ChatGPT — an electroencephalogram during the task, automatic analysis of the texts produced, evaluation by blind judges. Three sessions under the same condition for everyone, plus a fourth session with a subgroup that swapped tools. The result, in one sentence: the group writing unaided showed wider, more sustained brain activity in regions tied to memory, planning and creativity. The ChatGPT group showed a different pattern, consistent with the hypothesis that the brain is following suggestions instead of generating them.
The authors coined a term for the phenomenon: cognitive debt. The idea is simple. Like technical debt in software, cognitive debt builds up when we systematically outsource demanding mental tasks and, without noticing, run down the muscle that made them possible. The consequence isn't immediate. It's medium-term. And, like all debt, there comes a moment when it falls due.
It's worth marking the limits of the study. Small sample. Short time, weeks, not years. Lab setting, not real life. EEG is a useful but coarse tool. No one serious extrapolates from here to "ChatGPT makes you dumb". What the study does prove, with numbers and method, is that there is a measurable difference between the brain that articulates without an assistant and the brain that reviews the output of one. The difference isn't trivial. And it is, now, a documented fact, not a hunch.
Add to this the qualitative observations — university professors reporting that their students can't start an exam without a model sentence, professional journalists describing that after two years of heavy use it takes them three times as long to get a piece going without an assistant, translators noticing that their ability to hold a long argument in their head has shrunk — and what emerges is a consistent pattern. Not proof, but pattern. In science the pattern eventually translates, sooner or later, into a long study. And the long studies will come. When they do, they will probably confirm what everyday observation already suggests: that sustained delegation produces detraining, that the detraining shows, and that recovering the skill takes explicit retraining.
What is actually happening: homogenisation
There is a third phenomenon and it's worth naming. In 2024, Anil Doshi and Oliver Hauser published in Science Advances a study titled Generative AI enhances individual creativity but reduces the collective diversity of novel content. Empirical conclusion: when people write with AI assistance, individuals produce somewhat better texts in terms of creativity as scored by judges, but the aggregate body of texts resembles itself far more closely. Collective diversity drops. What was a constellation of voices turns, slowly, into a tuned choir.
This is not Skynet. This is something already happening in newsrooms, law offices, communication agencies, publishing houses. Texts signed by humans look more and more alike, not because humans resemble each other more, but because the stylistic bottleneck is the same assistant. When one model mediates the articulation of millions of people, articulation drifts toward the model's average. Diversity isn't announced as lost. It erodes in silence. And silence is exactly what the listicle needs in order to keep selling "10 dangers" in which that erosion never appears.
Why listicles look in the wrong place
When a listicle from any SaaS brand says "10 dangers of AI", the format demands discrete bullets, interchangeable, abstract and solvable. "Algorithmic bias", "job loss", "privacy", "deepfakes", "disinformation", "Skynet". Each bullet lands on something stated in a sentence and solved, according to the listicle, with "governance policies" — read, with the service the brand sells.
Corporate concentration doesn't fit that format. It's a structural, slow phenomenon that no course or SaaS tool resolves. Cognitive atrophy doesn't fit either: it requires understanding an MIT study, accepting an uncomfortable hypothesis about your own behaviour, and accepting that the solution is behavioural, not technical. Collective homogenisation fits even less: it means recognising that the problem isn't in the individual user but in the whole system, and that it affects global cultural diversity, not your next quarterly report.
Listicles, by construction, look where the format lets them look. And that place is almost always the cinematic: what fits in a sentence, what wears a villain's face, what can be sold as solvable. Skynet works. Concentration doesn't.
The operational difference
Let's compare the two classes of risk, side by side, without alarmism and without condescension.
Cinematic risk: Skynet. Short-term probability, by current technical consensus, low. Plausible mechanism: highly speculative. Documented cases: zero. People affected today: zero. Serious research that deserves to exist: yes, but as a bounded branch. Media coverage: maximal. SERP positioning: dominant in listicles.
Silent risk, first layer: corporate concentration of cognitive input. Probability: high, it's happening. Mechanism: documented in the AI Index 2026, in market movements, in the cross-investments Microsoft-OpenAI, Amazon/Google-Anthropic, in operational dependence on the hyperscalers. Documented cases: the entire sector. People affected today: more than 5 billion users with internet access. Media coverage: peripheral. SERP positioning: marginal.
Silent risk, second layer: cognitive debt from delegation. Probability: high and rising. Mechanism: documented in MIT studies in 2025, accumulated clinical observation, coherent qualitative description across professionals. Documented cases: 54 participants in the main MIT study, thousands in complementary studies. People affected today: any heavy user of an assistant, several hundred million people. Media coverage: occasional. SERP positioning: residual.
Silent risk, third layer: homogenisation of cultural production. Probability: high. Mechanism: documented in Science Advances 2024. Documented cases: a study published in a top-tier indexed journal. People affected today: the entire content-creation ecosystem. Media coverage: token. SERP positioning: practically nil.
The asymmetry is brutal. What is least likely to happen to you today takes up the whole stage. What is already happening to you has no name in Spanish, doesn't show up in bullets, isn't discussed over dinner. This is no accident, no editorial bad luck. This is what an information ecosystem produces when the broadcasters are both judge and party.
What a listicle does on the inside: anatomy of a misdirection
Let me stop on how the misdirection works, because understanding the mechanism is how you defend yourself from it. A listicle of "10 dangers of AI" splits the reader's attention into ten small containers. The total attention the reader will spend reading that listicle is finite — say, four minutes. Split four minutes across ten bullets and you get twenty-four seconds per bullet. In twenty-four seconds nothing is thought. A label is recognised and a nod follows. The operation ends by leaving in the reader's head a sensation: "I've understood the dangers of AI". But that sensation doesn't correspond to having understood a single one. It's a pure sensation of activity, like when you scroll for ten minutes and end up with the vague idea of having been informing yourself.
The listicle sells exactly that. The sensation. To genuinely teach one danger, it would have to spend thirty minutes on the danger. A single danger. With data. With context. With uncertainty. And that is no longer a listicle: it's a long, demanding article that loses traffic and doesn't scale. That's why it isn't done. That's why what sits in the SERP is the cheap model. And that's why the Spanish-speaking reader, after six months of reading listicles, knows ten words — "bias", "privacy", "deepfake", "Skynet", "unemployment" — and doesn't know a single idea.
Are there objective reasons for fear?
Yes, there are. But they aren't the ones the listicle offers you.
There are objective reasons to fear that the current corporate concentration of the AI sector — four to six companies controlling the infrastructure that already carries 53% of the world's population — produces a power asymmetry hard to reverse. It takes no imagining of futures. It only takes looking at the figures of the AI Index 2026 and comparing them with the figures of any earlier strategic sector. The concentration is greater, faster and geographically narrower than any comparable process of the last century.
There are objective reasons to fear that sustained cognitive delegation, absent compensating habits, produces measurable detraining of intellectual skills that took us decades to build and that don't come back on their own. It takes no imagining of futures. It only takes reading the MIT paper and accepting the metaphor of cognitive debt as a reasonable working hypothesis.
There are objective reasons to fear that the homogenisation of language produced through assistants erodes, within a decade, part of the cultural diversity that makes someone worth listening to. It takes no imagining of futures. It only takes reading Doshi-Hauser 2024 and multiplying by the millions of texts written that way every day.
And there are objective reasons to fear that the editorial ecosystem that should help you understand these risks is exactly the same ecosystem with an economic interest in you confusing them with Skynet. That, perhaps, is the most relevant objective reason of all. Not the revolt of the machines. The capture of the very language you use to talk about the machines.
What's actually happening is the boring thing. What sells is the cinematic. The difference between the two will not be corrected by Google's algorithm or by a multinational's marketing plan. It has to be corrected by the reader. By changing, one by one, the questions typed into the box.
Quick definitions
- Cognitive debt: concept introduced by Kosmyna et al. (MIT Media Lab, 2025). Analogous to technical debt in software. The build-up of deferred cognitive cost when a demanding mental task is systematically outsourced to an assistant, producing detraining of the underlying skill.
- Concentration of cognitive input: a situation in which the channel through which a user poses questions and receives synthesised answers passes through the infrastructure of a very small number of companies. In 2026 that number runs between four and six globally.
- Cinematic risk: a class of risk whose narrative shape fits the genres of commercial cinema (machine revolt, killer robot, hostile conscious AI). Optimised for media attention, frequently disconnected from real probability.
- Collective homogenisation: the phenomenon by which the aggregate diversity of a body of texts decreases when a single assistant mediates them, even if the individual quality of each text holds or improves.
Referencias
- Kosmyna, N., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab / arXiv. https://www.media.mit.edu/publications/your-brain-on-chatgpt/
- Doshi, A. & Hauser, O. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). DOI: 10.1126/sciadv.adn5290. https://www.science.org/doi/10.1126/sciadv.adn5290
- Stanford HAI (2026). AI Index Report 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report — source of the figures for global corporate AI investment (581.7 billion dollars in 2025), the growth of generative-AI investment (more than 200% year over year, to absorb almost half of private AI investment) and the adoption of generative AI by 53% of the world's population in under three years. Summary of figures at https://hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report
- ITU (2025). Facts and Figures 2025. https://www.itu.int/itu-d/reports/statistics/facts-figures-2025/ — source of the figure for internet users worldwide (more than 5 billion; the ITU estimate for 2025 is around 6 billion, roughly three quarters of the world's population).
- Carr, N. (2010). The Shallows: What the Internet Is Doing to Our Brains. Norton.
- Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs.
- Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.
- Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
También te interesa
- Why the critiques of AI aren't made by real experts
- The mind as limited space. Knowledge does take up room
- How did the public conversation about AI break?
No comments yet
No comments yet. Be the first.
Leave a comment