No text-only score can reliably prove that a passage was written by AI. Human and model writing distributions overlap, ordinary editing changes stylistic signals, and a confident percentage is not authorship evidence. tryb does not assign an ‘AI probability’ or turn linguistic guesswork into an accusation.
What can be checked deterministically is the text itself. Some generated, processed, copied, or watermark-bearing content contains invisible Unicode characters: zero-width spaces, joiners, directional controls, variation selectors, or unusual formatting code points. This tool reports the exact characters and positions it can observe, entirely in your browser.
Those artifacts are evidence of bytes and code points, not evidence of an author. They may come from a generator, a detector, an editor, a publishing system, a messaging app, a language-specific input method, or deliberate fingerprinting. Finding one does not prove AI use; finding none does not prove human authorship.
Use the result to answer a defensible question—‘what unusual invisible characters are present in this pasted text?’—not the unsupported question ‘who or what wrote this?’ The distinction is especially important in education, hiring, publishing, and investigations where a false authorship claim can cause real harm.
Edge cases worth knowing
- AI-generated text with no invisible artifacts
- Many model outputs contain ordinary Unicode text only. A clean result cannot establish human authorship and should never be presented as an AI-detector pass.
- Human-written text containing a zero-width character
- A person can copy text through an editor, website, PDF, or messaging app that introduces invisible formatting. The artifact remains real, but it does not identify the writing process.
- Legitimate multilingual formatting
- Joiners and directional controls can be meaningful in many writing systems. Context determines whether they are expected; their presence is not an AI signature.
Common mistakes
- Treating an invisible Unicode artifact as a unique signature of a particular model or vendor.
- Interpreting the absence of artifacts as proof that a person wrote the text.
- Using a detector result as the sole basis for discipline, rejection, grading, or attribution.
- Reporting a probability score as scientific certainty without validated provenance evidence.