ИИ-текст видно сразу. Признаки, которые выдают ИИ и как их убрать.

Summary

The presenter explains why AI-generated writing often feels recognizable: language models tend to repeat familiar linguistic patterns. He cautions that no single phrase proves a text was written by AI. Clues become more meaningful when several recur frequently or seem out of place.

Why AI writing develops a recognizable style

The presenter attributes common LLM habits to several influences:

• English-language training data: Models absorb English phrasing and sometimes reproduce it as literal or awkward-sounding Russian translations.
• Academic and corporate writing: These styles contribute formal vocabulary and abstract constructions.
• Preference for structured answers: In human-feedback training, evaluators often favor answers that are clear, detailed, and neatly organized. Those preferences can reinforce formulaic writing.
• Probabilistic generation: Models tend to choose broadly acceptable, low-risk phrasing that works in many situations, which can make their writing sound generic.

Common signals to watch for

These are tendencies, not definitive tests. The presenter emphasizes that frequency, context, and combinations of clues matter.

• “Not X, but Y” contrasts: For example, “It’s not just a tool; it’s a new way of thinking.” The construction is legitimate in Russian and English, but repeated use can make writing sound formulaic. It can also create a simplistic position only to knock it down—a “straw man”—or manufacture a sense of depth through contrast.
• Frequent groups of three: Phrases such as “quickly, conveniently, and securely” can make prose sound polished but over-decorated. If a simple claim is padded with three near-synonymous benefits, shorten it.
• Emphasis formulas: Repeated uses of words equivalent to “emphasize,” “highlight,” or “underscore” may reflect academic or English-language conventions. The words themselves are normal; repeated, unnecessary emphasis is the issue.
• Nominalizations and passive constructions: Phrases like “the implementation of the system allows for increased efficiency” can obscure a straightforward action. Prefer direct verbs and name who did what: “The company implemented the system.”
• English-derived or inflated wording: Terms and phrases such as “nuanced,” “holistic approach,” “ecosystem of solutions,” “seamless integration,” and “from X to Y” may sound unnatural or needlessly abstract in Russian. Replace them with specific, idiomatic language where appropriate.
• Chat-style transitions in standalone articles: Prompts and transitions such as “Let’s take a closer look,” “Let’s start with the basics,” or “Now let’s see what this means” can make an article feel copied from a chatbot exchange. Keep them only when they genuinely suit the format.
• Vague appeals to research or experts: Do not claim that “research shows” or that “experts say” without identifying the research or providing a citation. The presenter also jokes that he referred to studies in the video without naming them, illustrating the problem.
• Em dashes: A long dash is not a reliable sign of AI writing in Russian; it is a normal punctuation mark. The presenter says it may be more characteristic of some English-language writing, but warns against treating it as a Russian-language detection rule.

Ways to make AI-assisted writing sound more natural

The presenter offers several approaches:

1. Write the text yourself. This is the most dependable way to preserve your own voice, though it takes more time and effort.
2. Use AI for a draft, then rewrite it. Let the model help with research, facts, or structure, then revise the prose in your own language and add your own transitions and perspective.
3. Provide the structure and examples of your writing. Give the model an outline and samples—such as articles, posts, letters, or dictated speech—so it can try to follow your style.
4. Edit specifically for AI clichés. Give the model a finished draft and ask it to proofread and revise formulaic phrasing. The presenter notes that this may require developing a reusable editing “skill” for an AI agent.

Speakers and sources

• Speaker: Dmitry Sharov, the presenter.
• Sources and influences discussed: English-language and academic or corporate writing; human feedback used in model training; unnamed statistical studies on repeated stylistic patterns; and an unnamed British study said to track increased use of “underscore” in English-language academic texts after the arrival of large language models. The subtitles do not provide citations or identify the researchers.