Compared · What each approach can actually prove
Authorship marks compared
Detectors, self-declared badges, content credentials and cryptographic attestation, measured against one question: what does each one actually prove?
Page last reviewed 5 August 2026. This page is deliberately hard on our own product. The limits of HumanMark are stated before its capabilities.
Start with the honest limit, because everything else on this page depends on it. Nothing can prove a human wrote something. No instrument observes the act of thinking. Anyone selling proof of human authorship is selling something that does not exist. What can be proved is narrower, and the useful question is which narrower thing each approach actually establishes.
The four approaches
| Approach | What it actually establishes | Where it breaks |
|---|---|---|
| AI detectorsstatistical classifiers | A probability that text resembles model output, given a training distribution. | False positives on non-native English, formal registers and edited text. Defeated by paraphrase. Produces a number about a document and is then used to make an accusation about a person. |
| Self-declared badges“human authored” marks | That someone asserted human authorship, and that a body accepted the assertion. | The assertion is the whole product. Nothing binds the badge to the file, nothing prevents the same badge appearing on a different file, and nothing is checkable by the reader without trusting the issuer. |
| Content credentialsC2PA and similar | A signed manifest describing capture and edit history, bound to the file, checkable by anyone. | Strong on provenance of capture. Says little about the human contribution to text, and the manifest is commonly stripped by platforms that re-encode. |
| Cryptographic attestationwhat HumanMark does | That a named key, held on a specific device and unlocked by a specific person, signed a specific file at a specific time, together with the observed facts recorded at that moment. | Proves the signature and the observations. Does not prove authorship of any sentence. Tells you nothing about a file that was never marked. |
Why binary marks fail
A badge saying “human” or “AI” cannot describe how professional work is actually produced. At least five distinct situations exist, and they carry different consequences in a court, a client relationship and a regulatory file:
- 1Written without AI involvement of any kind.
- 2Written by a person, with AI used for research or retrieval only.
- 3AI produced a first draft; a person substantially rewrote and verified it.
- 4AI produced the text; a person edited and verified it.
- 5AI produced the text and it was filed substantially unchanged.
Nearly every sanctions decision on this record turns on the difference between band 3 and band 5, and on whether the person could show which one they were in. United States v. Farris is band 5 with six hours of supplementation on top; the Sixth Circuit removed counsel from the case. Gamez looked identical on the face of the brief and drew no sanction, because the explanation was immediate and credible. A binary mark cannot distinguish those two, and a detector would score them the same.
What HumanMark will not do, on principle
These are design constraints, not roadmap gaps.
- No detectionWe do not estimate whether text was AI-generated. Estimates get presented as findings, and findings about people should not rest on classifiers.
- No authorship proofWe attest cryptographic facts and observed facts. We never assert that a person composed a sentence, because that is not observable.
- Never a fake-evidence machineThe record attests what was observed. It is not a mechanism for manufacturing a stronger claim than the facts support, and the verifier is built to say plainly when a record proves less than a reader might assume.
- Content freeWhat is captured is the fact of the decision, the fingerprint of the material and the identity of the person. Not the material.
- Checkable without usVerification is a property of the record, done with public mathematics by anyone holding the file, free and permanently, with no account and no call to any service we run.
The honest comparison
Content credentials and cryptographic attestation are the two approaches that produce something a stranger can check. They answer different questions and they are complementary, not competing: one is strongest on where a file came from, the other on who stood behind it and what they did. Detectors and self-declared badges produce claims that cannot be independently verified at all, and a claim that cannot be checked is not evidence, whatever it is printed on.
If you take one line from this page: attestation is worth having precisely because it claims less. A record that says exactly what was observed, and no more, survives cross-examination. A badge that overclaims does not.
What to do about it
Seven days free, then $349 a month. Or run it across the estate.
Every sanctions decision on this record asked the same question: which band were you in, and what can you show. A record that attests only what was observed answers that without ever having to claim something it cannot prove. The same engine either way: a signed record made on your own machine, content-free, and checkable free by anyone, forever.
Start the free weekOr Regulayer for enterprise →
See also the attorney certification, the free verifier, and every jurisdiction we track.
Related
- Detectors compared
- HumanMark
- C2PA content credentials
- China’s labelling measures, the only law that specifies what goes in the file
- The open verifier
Reviewed 5 August 2026. This page makes no claim about any named competitor’s product; it compares approaches. Information, not legal advice.
Common questions
Can you prove a human wrote something?
Not in the way people mean. Nothing can prove that a specific human composed a specific sentence, because no instrument observes the act of thinking. What can be proved cryptographically is narrower and still useful: that a particular key, held by a particular person on a particular device, signed a particular file at a particular time, and that certain observed facts about how the work was produced were recorded at that moment.
Do AI detectors work?
Not reliably enough to base a decision on. Detectors are statistical classifiers that estimate whether text resembles model output. They produce false positives on non-native English writing, on formal and technical registers, and on edited text, and they can be defeated by paraphrasing. A detector output is a probability estimate, not evidence, and should never be presented as proof of anything about a person.
What is the difference between attesting and detecting?
Detecting looks at a finished file and guesses how it was made. Attesting records how it was made at the moment it was made, and signs that record. Detection is a claim about the past made by a stranger. Attestation is a claim made by a person about their own work, cryptographically bound so it cannot be altered afterwards.
Is human authorship a yes or no question?
No, and treating it as binary is the main error in this field. Almost all professional work now sits somewhere on a spectrum: unassisted drafting, AI used for research, AI used for a first draft that a person substantially rewrote, AI output lightly edited, and AI output passed through unchanged. Those are five materially different situations and a badge that says 'human' or 'AI' cannot distinguish them.
What does HumanMark not do?
It does not detect AI. It does not prove that a person composed any particular sentence. It cannot tell you anything about a file that was never marked. And it cannot stop someone signing a false claim: it makes the claim attributable and permanent, which is a deterrent, not a prevention.
Information, not legal advice, and not advice on the law of any jurisdiction. Every entry is verified against the issuing body’s own document; where a source is reporting rather than the document, we say so.