Habr AI Explained Why the 'Write Like a Human' Prompt Doesn't Save AI Texts from Clichés

Habr AI released an analysis explaining why the request "write like a human" rarely makes AI-generated text feel authentic. The core insight is simple: the problem isn't isolated markers like lists, participial phrases, or bureaucratic language, but rather that models often don't understand what type of speech and authorial voice the task actually requires.

Not in the Words

The author challenges the popular practice of "cleaning" text from everything that exposes the neural network at first glance. If an editor mechanically forbids lists, constructions, introductory phrases, and academic tone, they treat symptoms rather than the cause. Such an approach sometimes makes paragraphs shorter and punchier, but doesn't add thought, internal logic, or the sense that someone real stands behind the text. In the end, the form changes, but the feeling of emptiness remains.

According to Habr AI, the command "write like a human" sets the model a false goal from the start. It doesn't explain what exactly needs to be written: an instruction, news, column, analysis, or controversial comment. Instead, the algorithm receives an abstract requirement to imitate subjectivity. Hence the standard result: averaged, safe, and overly smooth text that resembles everything and nothing specific. It replaces concrete editorial briefs with a vague mask of humanity.

Styles and Tasks

At the heart of the argument lies an old philological idea: different tasks require different functional styles and text types. There is no single universal "human writing" that works equally well in news, instructions, and journalism. When an author or editor first defines the task, genre, and expected role of the text, language requirements become much more precise, and work with the model becomes notably more practical. This is where philology proves more useful than formulaic prompt engineering.

• Instructions rely on imperative mood, sequence of steps, and clarity of formulations.
• Short news is built around facts, context, and minimal interpretation.
• Analytical text requires conflict, paradox, and the author's intellectual move.
• Journalism works when the text carries a position, not an impersonal summary.

From this follows a simple conclusion: you cannot edit all AI-generated materials with the same set of prohibitions. Where dry structure is appropriate, lists only help. Where argument and tone are needed, the problem isn't bureaucratese per se, but the absence of a viewpoint. The author compares a model's standard response to a meaningless universal "42" answer to any question: the form may look decent, but substantively it misses the mark.

Why Text is Dead

A separate emphasis is placed on the difference between someone who simply types and a professional editor. A copywriter or philologist typically has years of training behind them: reading, analyzing literary and journalistic texts, sense of rhythm, understanding of subtext, attention to syntax. That's why a person notices not just clichés, but a subtler problem—the absence of internal intellectual movement that makes the text feel lifeless even without obvious "AI markers."

"Write like a human" is a command for the algorithm to imitate subjectivity.

That's precisely why fighting the "AI smell" through cosmetic editing looks like a dead end. A model can be taught to avoid certain words and constructions, but that doesn't equal the emergence of intention, position, or intellectual risk. If the task requires genuine reasoning, one prompt isn't enough: you need either a strong human editor or a much more precise task formulation, where genre, conflict, audience, and quality criteria are defined in advance. Without this, the model almost inevitably drifts into safe middle ground.

What This Means

For editors and content teams, the conclusion is uncomfortable but useful: good AI-generated text doesn't start with prohibiting "neural network" constructions, but with the right choice of genre and authorial role. The more precisely the task is described, the less magic in prompts and the greater the chances of getting text that works by design, not just masquerades as human-written. This changes both the editing approach and the task-setting process itself.