This tool is honest about what it can and cannot prove. Here is the full list, written without marketing language.
We covered the three kinds of "AI watermark" in What Is an AI Watermark, Really?. This page explains what a character scanner can and cannot find among them.
What it can find
- Invisible formatting characters such as zero-width spaces, joiners, soft hyphens, and byte order marks.
- Bidirectional control characters that can alter visual reading order.
- Special whitespace that looks like a normal space but has a different code point.
- Words that mix writing systems in ways that can spoof identifiers or URLs.
- Characters that would change under standard Unicode normalization.
What it cannot find
- Statistical watermarks like SynthID Text that change model output probabilities without inserting characters.
- Image pixel watermarks or C2PA metadata embedded in photos or videos.
- Whether a human or an AI wrote the text.
- Plagiarism, paraphrasing, or rewriting.
Why the tool stops where it does
A character scanner can only report what is present in the bytes it receives. Invisible characters and bidirectional controls are concrete code points, so detection is deterministic: the character is either there or it is not. Statistical watermarks and style scores are not like that. They are properties of a generation process or a statistical model, not characters sitting in the text, so no amount of scanning the finished text can reveal them. This tool is scoped to the deterministic half on purpose, because that is the half it can answer honestly and explain.
What "no findings" actually means
An empty report is a statement about the scanner, not about the text. It means none of the patterns the tool knows about matched. It does not mean the text is human-written, unwatermarked, original, or safe. A watermarking method the tool does not implement, a character outside its rule set, or an encoding it does not examine would all pass a single scan without appearing anywhere in the results. Treat a clean report as "nothing matched this specific check," not as a clearance.
How to read the results
Each finding is a signal, not a verdict. A zero-width space might be a watermark, but it might also be an accidental copy from a PDF. A BiDi control in code is suspicious; the same control in Arabic prose is normal. Use the report to decide what to review, not what to believe.
Reading a finding without over-reading it
- A finding tells you a character is present. It does not tell you why it is there.
- The same character can be a deliberate watermark, an artifact of copy-paste, or normal usage in a language that needs it.
- Severity levels describe how disruptive a character tends to be, not how guilty anyone is.
- Position and context matter more than the raw count. A control character inside a comment is less alarming than one splitting a string literal.
How to use the scanner within its limits
- Scan the exact text you care about, in the form it will be used. That means the same string you would paste into a form, commit to a repository, or publish.
- Read findings as pointers, not conclusions. Each one marks a line worth a closer look rather than a settled answer.
- Confirm before you act. When a finding appears, inspect the raw text around it and make sure you understand what you are seeing.
- Do not treat a clean scan as evidence in a dispute about authorship or originality; the tool does not address those questions at all.
- For those questions, use a purpose-built detector and weigh its stated error rates when you read the result.
Questions the scanner cannot answer for you
- Did a specific model produce this text? Use the provider's own detector if they offer one.
- Is this text original? Use a plagiarism or similarity service instead.
- Does this file carry metadata? Check EXIF or C2PA tooling, which reads a different layer than this scanner does.
- Does the writing style look machine-generated? Use a style-detection service, and read its score as an estimate rather than a measurement.
Remember: "No hidden characters found" only means the scanner did not detect the signals it knows about. It does not prove the text is human-written or unmarked.