Learn

AI-generated documents: what you can actually verify

AI has made fabricating a convincing document nearly free. A plausible employment letter, invoice, or reference takes one prompt, with no template-site fingerprints and no copy-paste artifacts, and the volume of synthetic paperwork in applications and disputes is rising accordingly. The uncomfortable half of this story: tools claiming to detect AI-written text are unreliable, with false positives and misses that make them unusable for decisions about money or people. The useful half: detection was always the wrong question. The right question is provenance, and provenance can still be verified.

Why content-based detection fails

AI text detectors look for statistical patterns in wording, and those patterns are neither stable across models nor absent from human writing. Polished human prose flags as AI; lightly edited AI passes as human. No serious decision should rest on one. Visual inspection fails for the same underlying reason: the content of a fake is now essentially perfect, so the content is no longer where the evidence lives.

Provenance still works

A document's claim was never really "this text is human-written." It was "this document came from this institution through this process." That claim is checkable regardless of what wrote the sentences. A genuine bank statement still carries the bank's production fingerprint, arrives on the statement cycle, and increasingly carries the bank's signature; an AI-fabricated one carries none of that, whatever its prose sounds like. The file-level checks, producer chain, dates, edit layers, signatures, work identically on AI fakes and hand-made fakes, because fabrication can imitate appearance but not the source system's records.

This is also where document trust is heading. Cryptographic provenance, institutions signing what they issue, verifiable at the file level by anyone, is the durable answer to synthetic content, and the infrastructure is spreading through banks, registrars, and government issuers now.

The practical screening posture

Treat document content as cheap and provenance as the test. Concretely: check the file's records against its claimed source, prefer signed documents wherever the issuer offers them, and route anything consequential through source verification, the portal login, the direct delivery, the callback. None of this requires knowing or caring whether AI wrote the words. It requires the document to prove where it came from, which honest documents increasingly can and fabricated ones can't.

FAQ

Does DocVerdict detect AI-generated text?

No, and we'd encourage skepticism of tools that claim to. DocVerdict reads what's verifiable: signatures, producer chains, dates, and edit history, the provenance evidence that synthetic documents lack and genuine ones carry.

Will AI eventually fake provenance too?

Metadata, yes, it already can, which is why metadata alone was never proof. Cryptographic signatures, no: forging one requires the issuer's private key, not a better model. That asymmetry is why provenance is the durable layer.

What about images and scans made by AI?

A generated "scan" carries no issuing-system fingerprint, the same hollowness as any scan, with perfect content. The screening answer is the same: scans of consequential documents are unverified by definition; go to the source.