Как удалить невидимые ИИ-метки ч.2. Без бана Claude и GPT

Summary

The video examines whether AI-generated text can be reliably identified or “de-watermarked,” and whether em dashes are evidence that a text was written by AI. Its main conclusion is that an em dash is not a reliable indicator of AI authorship: Russian writers and editors use them naturally, and some classic authors used them more often than the AI models discussed.

AI Text Markers and Ways to Alter Them

• The video describes current text watermarks as being embedded in the words a model chooses, rather than as invisible characters that can simply be removed by copying text into Notepad.
• Translating text through a separate service may disrupt an existing watermark, especially after substantial rewriting. However, ordinary users generally cannot verify whether a text contains a marker. Translation by the same AI may preserve its marker, while another AI may apply its own.
• A 2024 research demonstration reportedly showed that a watermark could be inferred or forged using many model outputs. It used about 30,000 responses, not 100, and tested older watermarking methods. The video says the approach has not been tested on current Claude, GPT, or Gemini markers.
• The presenter advises against asking ChatGPT, Claude, or another AI to remove its own markers. The companies’ policies do not specifically discuss “watermarks,” but the video says they prohibit bypassing product security. OpenAI’s verification documentation also reportedly forbids repeated checks intended to detect or evade a mark.
• DeepSeek’s documentation describes a visible notice in its app. The presenter says he did not find documentation of hidden, word-level markers like those discussed for Claude and ChatGPT. Some open-source models can be run locally without a provider’s service mark, though the video cautions that not finding a marker in documentation does not prove that none exists.
• Markers vary by content type:
  • Code may contain fewer markers because it needs to be precise, while explanatory comments may contain them. Simple edits, such as renaming variables, can disrupt a code marker.
  • ChatGPT images may have both file metadata and a pixel-level marker. A screenshot can remove metadata, while the pixel marker is intended to survive screenshots or cropping.

Em Dashes, AI Detectors, and Authorship

• The video distinguishes four marks often confused as “dashes”: a hyphen, an em dash, an en dash, and a minus sign. Word and Apple software can automatically convert two hyphens into an em dash.
• It traces the dash’s popularity in Russian writing to Nikolai Karamzin. People were mocked for imitating his style centuries ago, much as writers are now accused of imitating ChatGPT.
• The presenter cites analyses suggesting that some newer AI models use em dashes less often than earlier models. In one comparison, Dostoevsky used them more frequently than the most dash-heavy model. Editorial punctuation can also change a text’s dash count without changing its wording.
• AI detectors estimate authorship from patterns such as predictable word choice; they do not provide definitive proof. The video cites a student whose self-written thesis was reportedly flagged as AI-generated and argues that polished or conventional academic prose can be misclassified.
• For authorship disputes, the presenter recommends keeping a writing history rather than relying on punctuation or detector scores.

Guides and Practical Tips

• Avoiding em dashes in ChatGPT: Use its personalization or custom-instructions settings to ask it not to use long dashes.
• Avoiding em dashes in Claude: Add a preference in Claude’s instructions, such as replacing long dashes with commas, periods, or colons.
• Documenting authorship: Google Docs version history can show edits; the Draftback browser extension can replay the writing process; and Grammarly’s Authorship feature can record what was typed, pasted, or generated. The presenter says these tools should be enabled in advance.
• The video also points viewers to guides and links in its description, including instructions for setting model preferences and checking supported image markers.

Main Speaker and Sources

Main speaker: Maxim Popov, the video’s presenter.

Sources cited: Anthropic, OpenAI, Google, DeepSeek, Microsoft, GPT-2 detector documentation, Antiplagiat, and Grammarly; a University of Zurich research team; analyses by The Washington Post, The Economist, Graphite, and Arena; and works by Russian writers including Karamzin, Gogol, Chekhov, Pushkin, and Dostoevsky.