Why AI text irritates readers: a Habr author dissects the reaction to neural-network style

A write-up has appeared on Habr about why material generated or polished by neural networks so often irritates readers. The author tested this on himself: he published AI-generated texts, compared the reactions, and concluded that audiences usually catch not a substituted thought but a substituted human form.

An experiment on himself

The author describes an almost laboratory test. He deliberately posted neural-network-assembled material on Habr and tracked how people reacted to the delivery, rhythm and overall tone. The most telling moment came when he took his own old, well-rated article, lightly "combed it out" with AI and put it in front of the audience again. The ideas, the arguments and the author were unchanged, but the reaction got sharply worse. That gave him a clean comparison between reaction to content and reaction to packaging.

Instead of recognizing a familiar style, readers saw what is often called "generatedness": a too-even pace, clean transitions between blocks and safe phrasing without any living unevenness. The experiment matters precisely because it removes the main counterargument: the problem is not a weak topic or a new author. The wrapper changes, and trust changes with it. For the author, this became the main takeaway of the whole episode.

"Same author, same thoughts. Only the form changed."

What gives AI away

The text ties this irritation not only to a cultural fashion for criticizing AI but also to basic workings of the psyche. A reader constantly judges how much a piece of writing resembles a signal from a living interlocutor: where there is risk, where a personal choice is audible, where strange but meaningful turns are visible. When everything is too averaged and smooth, the brain starts treating the material as synthetic, even if a human wrote it or at least edited it heavily.

Suspicion is usually triggered by signs like these:

• perfectly even sentence rhythm with no natural jumps
• formulaic links between paragraphs and predictable conclusions
• generalizations instead of observations that can be checked
• repeating one idea in different words
• a sterile tone with no authorial risk or roughness

The author also stresses an important problem: this intuition is often wrong. People, too, can write dryly, by template, or after heavy editing. So the "AI detector" in the head works not as an analyzer of a text's origin but as an alarm: this is writing with little individuality and too much of the statistical middle. Hence the false positives and the aggression in comments. That is exactly why gut feeling alone is no longer enough for evaluation.

What is wrong with code

The same logic carries over to programming. Neural-network code often looks tidy: a clear structure, sensible variable names, familiar patterns and even decent comments. But this outward well-being can be a trap. The model reproduces the average image of a "good solution" well, but is weaker at holding a project's real constraints: architecture history, non-obvious invariants, fragile integration points and edge cases. In real development, these usually decide the fate of the result.

That leaves developers facing a particular kind of distrust. The error is not always visible right away, because the code looks convincing and reads easily. The problem surfaces later: on unusual data, in combination with neighboring modules, or when someone tries to maintain the solution further. In essence, text and code behave the same way here: a neural network often produces a plausible form faster than a deep understanding of context. And when a person notices this gap, the irritation only grows.

What this means

The main conclusion is not that AI content should be banned, but that it cannot be judged by its surface alone. For authors, it is a signal to polish text less into impersonal smoothness; for editors, to preserve voice and specifics; and for developers, to check neural-network code as a draft, not as a finished solution. In the coming years, the winners will not be those who simply plugged in AI, but those who learned not to lose human unevenness and context.