AI text watermark: what can actually be checked?

An AI text watermark can mean several different things. Clear definitions matter because no single detector can inspect every possible system.

Inspectable text artifacts

Some documents contain zero-width characters, unusual spaces, bidirectional controls, Unicode tags, or other formatting codepoints. These characters are part of the text itself. A deterministic scanner can identify their exact location and, when safe, remove or normalize them.

These artifacts are not automatically evidence of AI generation. They can come from websites, document exports, templates, accessibility tools, multilingual writing, collaborative editors, or ordinary copy and paste.

Statistical model-level watermarks

Some research watermark schemes influence token selection during generation. Detecting them can require the same secret key, tokenizer, sampling rules, and detector configuration used when the text was created. A generic AI text watermark detector cannot reliably infer every undisclosed scheme from arbitrary prose.

Paraphrasing or rewriting also changes wording and meaning. Scan AI Text does not present rewriting as a guaranteed watermark remover.

AI authorship is a different question

Authorship detectors estimate patterns. Their results can vary with language, genre, editing, length, and model changes. A score is not proof that a student used AI. Decisions with academic or professional consequences need context, process evidence, and human review.

Honest checklist

• Use a text watermark checker for visible or hidden characters you can inspect.
• Review exact codepoints and offsets before cleaning.
• Preserve characters that carry linguistic, emoji, or direction meaning.
• Treat writing-pattern signals as low-confidence observations.
• Do not claim universal model-watermark detection without the required detector configuration.

See how cleaning works or [scan text locally](/).