Do you remember what it felt like not to know something — and to figure it out yourself?
Not googling. Not asking. But: thinking, searching, being wrong, continuing. The feeling when something comes together, slowly, from your own effort. It wasn’t long ago that this was normal. It feels like a different era anyway.
We’ve gotten used to receiving answers before we’ve finished forming the question. This doesn’t happen through coercion. It happens through usefulness — and that’s what makes it insidious. Nobody convinced us. We chose it ourselves, every morning anew, because it works. Because it’s faster. Because it’s good enough. And somewhere in that “good enough” lies the beginning of something hardly anyone thinks about out loud.
Artificial intelligence is in the process of writing itself into workflows, decision processes, everyday routines — quietly, almost invisibly, always with the promise of efficiency. Companies build processes on systems they don’t control. Individuals delegate thought processes they would previously have carried out themselves. This feels like progress. And perhaps it is — partially. But progress and dependency are not mutually exclusive. They often grow together.
The question isn’t whether AI is useful. The question is: at what price, and who sets it?
Anyone who looks even superficially at the mathematics behind this industry quickly encounters an unease that can’t be argued away. The major providers — Microsoft, Google, Amazon, Anthropic, Open AI — have invested hundreds of billions of dollars in infrastructure over recent years. Data centers, power plants, chips. A bet on the future of a scale that has few parallels in the history of technology.
These investments need to pay off. That’s not speculation — that’s arithmetic.
And now do the math: a large portion of people who use AI systems daily pay nothing. Those who pay, pay twenty euros a month. The infrastructure behind it costs many times that. The hardware in these data centers — specialized processors operating at the limits of what’s physically possible — is obsolete in two or three years and must be replaced. The investment cycle doesn’t stop. There’s no threshold after which everything has amortized and prices can fall.
This isn’t a business model. This is market construction.
Lock-in doesn’t work through locks. It works through habit, through embedded processes, through the creeping impossibility of going back. Whoever has built their company’s workflow on one of these systems — texts, analyses, customer correspondence, decision preparation — hasn’t just integrated a tool. They’ve created a dependency. And dependencies have prices.
The moment these prices become visible doesn’t come as an announcement. It comes as a gradual shift. What today is included in the free tier becomes premium tomorrow. What today costs twenty euros costs five times that in three years — and by then switching is more expensive than the price increase. We know this. From software licenses, from cloud services, from platforms that were free first and then indispensable.
AI is no different. It’s just bigger.
Particularly loud right now are those who don’t want to see this — or can’t see it, because their business model depends on others not seeing it either. The courses, the webinars, the LinkedIn posts: How to double your income with AI in three months. They’re selling tools whose foundation they’ve never questioned. They’re building on sand and calling it strategy.
The problem isn’t just the naivety. It’s the propagation. Whoever thinks this way and lets others think this way accelerates exactly the dependency they themselves don’t see.
But it’s about more than money.
There’s a subtler form of loss that’s talked about even less. When systems take over what we used to think for ourselves — not just researched, but really thought, structured, questioned — something atrophies. Not immediately. Not noticeably. But steadily.
Abilities that aren’t used erode. This applies to muscles, to languages, to thinking habits. Whoever stops formulating for themselves loses their feel for language. Whoever stops structuring for themselves loses their feel for argumentation. Whoever has every uncertainty immediately resolved loses tolerance for not-knowing — and with it one of the most important engines for real learning.
This isn’t cultural criticism. It’s a practical observation.
So what to do?
That would be the wrong question — at least as an expectation of this text. There’s no clean answer, no list of measures that solves the problem. Whoever expects that hasn’t yet understood the scope.
What there is, is an attitude. The decision to know the dependency before it arises. To know which parts of one’s own thinking, one’s own working, one’s own judgment one is prepared to hand over — and which not. To know local models as a real alternative, not as dogma. Token efficiency not as a cost-saving measure, but as a thinking principle: what do I actually need, what is noise?
And the willingness not to let uncomfortable questions be delegated.
Here are the numbers. Hundreds of billions invested. Hardware that’s obsolete in three years. Prices that don’t add up today. Dependencies growing deeper daily. Users who don’t pay, users who pay too little, infrastructure that keeps being built anyway.
At some point this calculation has to work out.
Who pays it, and on what terms — that’s being decided quietly right now. Not in parliaments. Not in public debates. In server farms, investor rounds, and product decisions far removed from the moment when you open your device in the morning and ask.
Do you remember what it felt like not to know something?
Hold onto that.