Check these 5 things before AI text enters a filing

The risk is specific, so the check can be too

The failure mode is not that a model is bad at law. It is that a model has no way to distinguish between recalling something and constructing something that fits the shape of the answer. Both come out in the same confident register.

So the check is not "is this good". It is "can each factual claim here be traced to a document I can open". Five passes, in order, and they get faster with practice.

1. Every claim names a source

Take each sentence that asserts a fact and ask what it points at. A passage that cites nothing is not a draft, it is a hypothesis. Send it back or delete it.

This is why it matters that the tool you use attaches sources as it works rather than offering to find them afterwards. Retrofitted citations get attached to whatever is closest, which is a different thing from where the claim came from.

2. Open the source, do not trust the label

A reference that looks right is not the same as a reference that is right. Open the document. Confirm it exists, confirm it says what the passage says it says, and confirm it is the document you think it is rather than a similar one.

For file material a checksum makes this cheap: the same document, verifiably unaltered since it entered the file. For cited authorities there is no shortcut. Look it up.

3. Quotations are verbatim, or they are not quotations

Check quoted text character by character against the original. Models paraphrase while keeping quotation marks, which is the most dangerous error on this list because it survives casual review.

A related trap in multilingual files: a quotation translated into the language of the filing has stopped being a quotation. Quote the original and translate alongside it.

4. Dates and numbers, individually

Dates drift. Figures get rounded, transposed, or attached to the wrong party. These are exactly the errors an opponent enjoys finding, and they are quick to check against the source you already have open from step two.

5. Cut what nobody asked for

Models fill space. Ask for a summary of three documents and you may get a fourth point that reads well and rests on nothing. Anything that does not trace back to a document in step one comes out, however good it sounds.

Build the check into the tool, not into your discipline

A process that depends on remembering to be careful at eleven at night is a process that fails at eleven at night.

Legalnaut attaches the source to every generated statement and marks AI drafts as drafts requiring review, so an unsupported claim has nowhere to hide and nothing leaves the system labelled as finished work. The checking still belongs to a lawyer. What changes is that the material needed to check is already attached.