AI Has Won

Neural networks already control your world. But perhaps not in the way you think.

When people talk about the dangers of AI, they usually paint two pictures. The first is boring: the neural network will speed up work, replace some professions, reduce costs, and then we'll figure it out somehow. The second is cinematic: superintelligence, manipulation, catastrophe, the doomsday button.

Both pictures are convenient. Both place the problem somewhere in the future. But the unpleasant news is that AI's main victory is already here.

Not in the form of a robot with red eyes. Not in the form of instant unemployment. It looks much more mundane: like a neatly written paragraph of text.


Text — the most insidious format

With images, it's simpler. Six fingers, a bear cooking borscht — the brain gets a signal: this is a fake. With video, we're also learning to be more cautious.

But text is the perfect smuggler.

If it's coherent, confident, and written without obvious nonsense, most people won't bother to double-check that the acceleration due to gravity is not 11, that "eating a rock a day is beneficial" is not a medical recommendation, and that the list of sources actually exists. The machine transmits nonsense with the tone of a reference book.


A New Authority

We have an old bug in our heads: the computer is right because it's a computer.

There's an Excel formula — so there must be strict logic somewhere. If the answer looks smooth — it must be based on something. If AI says it confidently — it must "know something."

But an LLM is a probabilistic machine for assembling plausible phrases. The problem is that people often read this as a new type of authority.

And research also speaks to this. In a Microsoft Research study with 319 knowledge workers, higher confidence in GenAI was associated with less critical thinking, and working with AI often shifted from independent analysis to verifying a ready-made answer.

In a series of experiments, people generally perceived an AI source as less biased, more informative, and less interested in persuasion than a human. That is, the very "non-humanity" of the system increases trust for many.


We Begin to Think Inside Someone Else's Framework

What happens next — people relay model responses as their own conclusions. Not because they are stupid. But because it's cheaper in terms of effort.

Thus, a very strange form of dependence arises: we seem to still speak for ourselves, but increasingly — with words, logic, and frameworks that the machine has pre-proposed.

And here, even labeling "this was written by AI" helps little. In an experiment with 1601 participants, such labels hardly changed the persuasiveness of the message, the assessment of its accuracy, or the willingness to share it further. In other words, the mere fact of artificial origin of the text does not make people noticeably more cautious.


Self-Reinforcing Nonsense

Then the most dangerous mechanism kicks in — a feedback loop.

The model makes a mistake
→ A person copies the mistake into a post, note, comment, SEO text, abstract, presentation
→ This mass of text settles on the internet
→ The next model learns from a world where fiction statistically looks like a normal part of reality.

In a study by Shumailov, this process is described as model collapse: when recursively trained on generated data, models begin to lose the tails of distributions and reflect the original reality worse, while maintaining the appearance of normal operation. That is, degradation can be quiet and almost imperceptible.

AI does not have to become smarter than humans to start shaping reality. It just needs to become widespread, convenient, and convincing enough.

Not the end of the world. The end of the world would at least be noticeable. But here, everything happens in a comfortable mode. You are not lying under the ruins of civilization. You are simply living in a world where truth is increasingly determined not by verification, but by the persuasiveness of the phrasing. And where the depth of understanding is gradually replaced by access to a more expensive, faster, "smarter" model.


❗️❗️❗️❗️❗️❗️❗️❗️ / Not banned in the Russian Federation