Content provenance vs. content detection

An AI content detection tool guesses from patterns and gives a score. Provenance checks a signed label and gives a plain result. Use detection to screen other people's work. Use provenance to sign your own.

Start signing free Check a file

Provenance reads back a signed record: signed by Town Daily, made with AI: no, unchanged since signing. Detection runs unsigned text through an AI detector and returns a guess, such as 72 percent likely AI. — Provenance is a record anyone can check. Detection is a guess with a score.
Content provenance
AI detector
False positive
Digital signature

Side by side

AI content detection and signed provenance, side by side.
	AI content detection	Content provenance
How it works	A model looks for patterns	A check reads a signed label
What you get	A score, such as 73% likely AI	A plain result: who signed, and whether it changed
Wrong flags	Human work flagged as AI happens	No score to misread; the signature holds or fails
After editing	Paraphrasing can lower the score	Any change to signed content shows
Needs content signed first	No	Yes
Works on anything, with no prior setup	Yes, its main strength	No, only on signed content
Screening submissions	Useful as a first filter	Only if submitters sign
Proves content is not AI	No	No

When to use each

Use detection: To screen work you did not create and cannot ask people to sign, like essays or open submissions. Treat the score as a hint.
Use provenance: To sign your own work: your articles, your files, and your AI output, which the EU AI Act asks you to mark.

What one detector's maker found

In January 2023 OpenAI released its own AI text classifier. It flagged 26% of AI-written text as likely AI, and wrongly flagged 9% of human text. OpenAI withdrew it in July 2023 for low accuracy.

OpenAI's AI text classifier, January 2023: share of texts it labelled likely AI-written.

AI-written text, correctly flagged26%
Human text, wrongly flagged9%

OpenAI, withdrawn July 2023

Source: OpenAI

Why a wrong flag costs more than a missed one

A false positive calls human work AI. A 2023 Stanford study found detectors flagged essays by non-native English writers far more often. A signed label never accuses anyone: it speaks only for what was signed.

A signed text mark gives a yes or no anyone can check, valid or changed; a statistical pattern runs through a model and gives a probability, such as 72 percent likely AI. — A cryptographic mark gives a checkable answer. A statistical pattern gives a probability.
Cryptographic watermarking
Statistical watermarking
Digital signature

Source: Liang et al., Patterns, 2023

Questions

Does provenance replace detection for every job?

No. Provenance only speaks for content that was signed. Some work comes from people who did not sign it, such as student essays or open submissions. There a detector is the only option, and its score is a hint, not proof.

Can provenance prove something is not AI?

No. A missing label could mean human work that was never signed, AI work, or a copy whose label was stripped. Provenance proves what was signed, not what was not.

Can a signed label be wrong?

A label shows who signed and what they declared. If the signer declares something false, the check still shows who said it. That is why the signer's name matters.

For engineers

Show the technical detail

Detection: Text detectors score token probabilities, perplexity and burstiness against a trained classifier. Scores shift with model updates and with editing.
Provenance: A C2PA check validates a COSE signature and X.509 chain, then compares a hard-binding hash. The result is per manifest, not a probability.
Watermarks: Statistical watermarks (such as SynthID) sit between the two: detected with a confidence level, not a signature.

Create. Mark. Endure.

Sign your own work. Your first file is free.