An AI detector reads finished text and returns a statistical guess about its origin, and it needs your document to do it. A sealed certification is a record made at the moment of work by a named person: what AI was used, for what, and what they personally checked, signed and checkable. One estimates who wrote a document. The other records what a person did.
“Proves” has a precise meaning here: the seal proves the record is unchanged since signing, and the signature proves who stands behind the declaration.
Start with the detector's own numbers
The most-cited measurement of detector reliability is a 2023 Stanford study published in the journal Patterns. Researchers ran seven widely used GPT detectors over 91 TOEFL essays, all written by humans, all by non-native English writers:
Liang, Yuksekgonul, Mao, Wu, Zou, “GPT detectors are biased against non-native English writers,” Patterns (Cell Press), 2023. The paper (DOI). A second team, Weber-Wulff and colleagues (2023), tested fourteen detection tools and found none both accurate and reliable; both studies are reviewed in this 2026 assessment.
Legal prose sits in the detectors' weakest zone for a second reason: formal, conventional writing is statistically predictable, and predictability is exactly what detectors flag. The better the brief, the more machine-like it scores. Detectors have their uses as screening signals, and their vendors say as much. What a detector score is not, on the published record, is evidence a court can weigh about who wrote what.
The five options, side by side
| Approach | What it is | Needs your document? | When it judges | What a court can weigh | Failure mode |
|---|---|---|---|---|---|
 The Regulayer™ Receipt |
A signed record of declared AI use and attested human review, sealed on your machine | No, only its fingerprints | At the moment of work | A signed, dated, tamper-evident record, free to check | Regulayer™ supplies the evidence. Your auditor, regulator or court makes the determination. |
| AI detectors | A statistical guess about a text's origin | Yes, the full text | After the fact | Very little: a contested probability score | False positives concentrated on non-native and highly formal writing; beaten by paraphrase |
| The typed Word declaration | Your statement, typed and signed | No | When you write it | Your word | Undated, unverifiable, silently editable; one made after the fact reads exactly like one made at the time |
| Cloud sealed-evidence platforms | Sealed records created and held by a vendor | Usually, into the vendor's custody | At creation | A third party's log, if the vendor still operates | Custody: privileged work leaves your machine, and the evidence depends on the vendor's survival |
| Free hash sealers (including our Sealer) | Integrity proof for a file: it existed, unchanged, at a time | Only its fingerprint, if designed right | At sealing | That this exact file existed then | No attestation: says nothing about what any person did or checked |
The detector
A detector answers a question nobody asked at the time, using only the artifact. It cannot see the process: the prompts, the drafts, the checking, the person. So it infers origin from style, and style is where its documented errors live. If you are accused on the strength of a detector score, the literature above is the rebuttal. If you are tempted to use one on opposing counsel's filing, expect the same literature back. Either way, a guess is not a record.
The typed declaration, honestly
A typed declaration is what most practitioners file today. Standing orders ask for your statement, and a signed Word document delivers it. It costs nothing. Its weaknesses only matter when the statement is doubted: it cannot show when the review happened, it cannot show the statement is unaltered, and one reconstructed after a flag reads exactly like one made before filing. The day your word alone is questioned, those three absences are the whole case.
The cloud platform, and the custody question
Sealed-evidence platforms create real records, but the record and usually the work sit in a vendor's custody. Two problems follow. First, custody is now a legal question: in February 2026 the U.S. District Court for the Southern District of New York held that a defendant's documents produced through a consumer AI platform carried neither attorney-client privilege nor work-product protection, because the platform was a third party (United States v. Heppner, No. 25-cr-00503-JSR; courts split the same week, and the analysis continues on our page on the ruling). Second, continuity: a record that can only be checked on a vendor's servers depends on those servers still being there. Verification that requires the maker to still exist is a service rather than a property of the record.
The hash sealer, and what is missing
A hash sealer does one clean job: it proves a file existed, unchanged, at a time. Our own Sealer does it free, locally, with nothing uploaded. But integrity is not attestation. A sealed brief tells the court nothing about what AI was used or what anyone checked, because nobody declared anything. A seal answers “is this the same file.” A standing order asks “what did you use, and what did you verify.”
The receipt, and its limits
The Regulayer™ Receipt keeps the seal and adds the declaration: the tools, the role they played, the named reviewer, the checks carried out, the attestation verbatim, signed at the moment of work on your own machine; it can be checked free, independently, with no upload. Its honesty is on its face: the evidence level reads “declared,” the time says where it came from, and no one, including us, can upgrade what the record claims after the fact. It is evidence, not a verdict. That is exactly why a court can weigh it. The full anatomy
Questions people ask
Are AI detectors accurate enough to rely on in a legal dispute?
The peer-reviewed record says no for high-stakes use. The Stanford study above documents a 61.2 per cent false-positive rate on non-native writers across seven widely used detectors, and a second team found none of fourteen tools both accurate and reliable. A detector score is a screening signal, not evidence of authorship.
Is a typed declaration enough for a court's AI disclosure requirement?
The orders ask for the attorney's statement, and a signed declaration is what most practitioners file. What it cannot show is when the review happened or whether the statement was later altered. A sealed, signed record answers the same requirement and adds both properties. Whether any particular filing satisfies a particular order is for the court.
Can I run a detector on opposing counsel's filing?
You can, and it will tell you almost nothing a court can use. The documented error rates run in both directions: false positives on careful human prose, false negatives on edited machine text. The stronger move is to ask for the record: the certification the standing order already entitles the court to expect.
What can a court actually weigh when AI use is disputed?
A record: who signed it, when it was made, what it declares, whether it has been altered since. A detector's probability score is contested in the literature, and a bare denial or accusation is not evidence either way. Contemporaneous, signed records an outsider can verify are the category with evidentiary traction.
If you remember one line: an accusation is not evidence; a record is.
Make the record before you need it.
The Regulayer™ Receipt: which AI you used, what you checked, your attestation signed on your own machine; anyone you send it to can verify it free. See the attorney's receipt, try the live demos, or verify a real record free.
Study figures verified against the linked sources on 4 August 2026.
What the rules themselves ask for
No court anywhere asks for a detector score. What courts ask for is disclosure and verification, and in New South Wales, the record of the prompts. Twenty-five entries, read at source.
AI rules for lawyers, country by country →