Как удалить невидимые ИИ-метки. И не попасть в бан Claude и GPT

Summary

Maxim Popov explains four kinds of “AI traces” that may make text look machine-generated:

1. Visible formatting: Frequent em dashes, emojis, and similar habits.
2. Invisible characters: Non-breaking spaces or zero-width characters, sometimes introduced when copying text from a chat.
3. Stylistic patterns: Repeated sentence structures and stock phrases, such as “It’s not just X, it’s Y.”
4. Word-choice watermarks: Statistical patterns based on a model’s choices among near-synonyms, rather than hidden characters in the text.

A central point is that removing invisible characters may clean up formatting, but, according to the speaker, it does not remove a word-choice watermark. He also warns that public AI-detection sites generally cannot verify a company’s watermark directly; instead, they estimate whether text is AI-written based on style and formatting. Citing cases where detectors disagree and human writing is incorrectly flagged, he argues that a single detector result should not be treated as conclusive.

Methods and Tools Discussed

• Reduce obvious style markers: Set writing preferences in ChatGPT’s settings, Claude project instructions, or an agent’s rules. These might include avoiding certain punctuation, emojis, and stock phrases.
• Review for clichés: The speaker mentions humanizer.ru, a tool that highlights phrases it considers AI-like and provides a frequency score. He also points to a Russian Wikipedia list of generated-text characteristics.
• Clean invisible characters: Free web tools can reveal and remove such characters, but the speaker cautions that they do not remove word-choice watermarks. He advises against uploading sensitive documents to third-party sites.
• Avoid copying chat formatting: Ask the model to save its output as a file instead of copying directly from the chat window.
• Rewrite or translate: The video says substantial rewriting, or translating text with a separate service and then back again, may disrupt a watermark. DeepL and LibreTranslate are given as examples. The speaker cautions that translating with the same model may preserve its pattern.
• Use a local model: LM Studio can be used to download and run open models locally. The video says models with around 7–8 billion parameters may be roughly 5 GB in compressed form and run on a laptop with 16 GB of RAM. Local processing can keep business documents off external services, but the output should be checked for errors or changes in meaning.

The speaker recommends rereading automatically revised text to restore accuracy and the writer’s own voice. He also stresses that these techniques do not justify presenting AI-generated work as one’s own when rules require disclosure or prohibit AI use.

Account Restrictions and Responsible Use

Popov warns against asking Claude or ChatGPT directly to remove their own watermark. He says this could be interpreted as an attempt to bypass product protections, although he notes that Anthropic’s rules do not explicitly name watermark removal. He also says the rules prohibit misrepresenting AI output as human-written.

His practical warning is not to use AI to deceive a university or pass off someone else’s work as one’s own.

For business publishing, he says Google’s concern is content produced at scale merely to manipulate search rankings, rather than appropriate AI use itself. He emphasizes that published material should be useful and fact-checked.

Reviews, Guides, and Analyses Covered

• A guide to recognizing and reducing visible, invisible, stylistic, and watermark-related text traces.
• A review of the limitations of AI-text detectors, including reported error rates and a cited student case.
• A walkthrough of tools and approaches: humanizer.ru, invisible-character cleaners, translation services, and LM Studio with local models.
• An analysis of how the speaker says Claude, Google, and OpenAI watermark text, as well as possible account-policy risks.

Main Speaker and Sources

• Main speaker: Maxim Popov.
• Sources and organizations cited: Anthropic, OpenAI, Google/SynthID, Stanford researchers, Turnitin, and a U.S. court case discussed by the speaker.