Question 1

Can you tell if a document was written by AI?

Accepted Answer

Not reliably from the text, and no honest examiner will claim to (§4). AI-text detectors produce false positives and negatives and are easily defeated, so a "this was written by AI" opinion based on the prose does not survive scrutiny. What can be established, objectively, is provenance: whether the document's metadata and creation history support its claimed origin, and whether it exists where a genuine document would (§5). That, not text analysis, is the reliable answer.

Question 2

How do you prove a document is fabricated?

Accepted Answer

By provenance and corroboration, not prose (§5, Example 1). A genuine email exists in mailboxes, on servers and in backups; a genuine file has a consistent creation and revision history and copies in the cloud and with others. A fabricated document typically exists only where the party produced it, contradicts its own creation metadata, and is absent from every system that should hold it (§8). That objective absence and inconsistency, not how the text reads, exposes the fake.

Question 3

Why do we need the native file and not a printout?

Accepted Answer

Because provenance lives in the native file and its system context (§5). A printout or screenshot strips the metadata, the author in g records, the creation and revision history, the very things the analysis depends on, and flattens the document into an image that cannot be tested. To examine provenance you need the electronic original with its metadata intact, obtained from the system where it lives, with integrity preserved.

Question 4

Our opponent's submissions cite cases we cannot find. What does that mean?

Accepted Answer

It may mean the authorities were hallucinated by an AI tool used in drafting and never verified (§3, Example 2), an increasingly common and serious problem. The check is the simplest one: verify each c it at i on against a primary source. Authorities that exist nowhere are a real risk to the party relying on them, engaging duties to the court, and the safeguard is verification, not any attempt to "detect AI" in the writing.

Question 5

Can metadata itself be faked?

Accepted Answer

Yes, metadata can be altered, which is why it is never taken alone (§6). Its power comes from corroboration: a creation date is tested against independent system records, and a document's claimed metadata is checked against the mailboxes, servers and backups that should independently reflect it. Consistent corroboration across systems supports authenticity; contradiction and absence expose fabrication. Metadata is a strong tool, used with the same corroboration discipline as any artefact.

Question 6

Is using AI to draft documents a problem in itself?

Accepted Answer

No, honest, checked and, where required, disclosed AI-assisted drafting is legitimate, and this guide does not treat the tool as the wrong (§7). The problems are fabrication, using AI to manufacture false evidence, and negligence, relying on invented authorities or facts without verifying them. The line is between assistance that is checked and disclosed and deception or carelessness that is not, and provenance and verification, not hostility to the technology, are how that line is policed. cflab. u k · e-disc ove r y. u k ©2026 Computer Forensics Lab Ltd ·cflab.uk ·e-discovery.uk ·info@cflab.uk ·+44 (0)20 7164 6915 Page 15 of 17

AI Generated Documents Provenance And Detection