The two families of AI watermarks
AI watermarks fall into two broad categories: character-level artefacts and statistical fingerprints. Character-level artefacts are discrete Unicode code points — zero-width spaces, non-breaking spaces, directional marks — inserted into the text so it looks normal to a human reader but carries a hidden payload. Statistical fingerprints are woven into the model's generation process itself: certain words are systematically preferred over synonyms, creating a detectable bias in token probabilities that survives paraphrasing.
Why AI companies embed watermarks
Three commercial pressures drive watermarking. First, legal liability: outputs identified as AI-generated grant companies a defence in copyright disputes. Second, misuse prevention: flagging AI text helps platforms moderate spam, fake reviews, and disinformation. Third, research accountability: published AI outputs can be traced to a specific model version, which is valuable for auditing safety incidents.
Invisible Unicode characters
The most common character-level watermarks are invisible Unicode code points. They have zero visual width, so they are undetectable in any text editor displaying a proportional font. They survive copy-paste from browser to document without any visual trace. Common examples include:
- U+200B — Zero-width space
- The most prevalent AI watermark character. Inserted between words or after punctuation — undetectable visually but picked up by AI-detectors and some word processors as an anomaly.
- U+200C — Zero-width non-joiner
- Legitimate in Arabic and Farsi text to suppress ligatures. In Latin-script documents it has no visual function and is a reliable watermark signal.
- U+00AD — Soft hyphen
- Rendered only at a line break. Placing it mid-word creates an invisible marker that survives most copy-paste operations.
- U+FEFF — Byte-order mark
- A legacy encoding marker that should appear only at the start of a file. Found mid-text, it is always an insertion artefact.
Statistical and semantic watermarks
More sophisticated watermarks operate at the model level and require no special characters. Google's SynthID for text uses a 'tournament' system: at each token position the model secretly ranks candidate tokens and biases toward those in a predetermined green list. The bias is too small to affect readability but large enough to be detected with statistical confidence across a paragraph. These watermarks are harder to remove because cleaning them requires paraphrasing the content, which changes meaning.
How watermarks affect your workflows
Even invisible watermarks cause real problems. Zero-width characters break regular-expression patterns in code, corrupt token counts in APIs billed per token, trigger false positives in plagiarism checkers, and cause word processors to mis-hyphenate words. For SEO, search engine crawlers may interpret hidden characters as keyword stuffing. For publishing, they can trip CMS content-length validators and cause encoding failures in PDF export.
Detecting and removing character-level watermarks
Character-level watermarks are reliably detected by scanning the Unicode code points of every character in the text and flagging those outside the expected range for the document language. Removal is lossless: the hidden character is deleted with no effect on the readable content. Our tool performs exactly this scan in the browser, on your device, without uploading your text to any server.