Why every AI-drafted fact should link to its page

AI can draft a demand, a petition or a chronology in minutes. Checking it still takes as long as it takes. That gap is where firms get hurt, and the fix is less about the AI than about what it hands back: every fact tied to the page it came from.

What courts have done

Two sanctions from 2025 show where the risk sits.

In February 2025 a federal judge in Wyoming sanctioned three Morgan & Morgan lawyers in Wadsworth v. Walmart over motions in limine that cited nine cases, eight of which did not exist. The drafting lawyer had used an AI platform the firm had built. The court revoked his admission to the case and fined him $3,000, and fined the two lawyers who also signed the motions $1,000 each. The firm itself wasn't sanctioned, in part because it had put policies in place requiring lawyers to verify AI output (Bloomberg Law; Law360).

In July 2025 the California Workers' Compensation Appeals Board sanctioned a firm $750, reduced from $2,500 after the firm promptly admitted its errors, over a petition that failed to cite the evidentiary record and appeared to cite authorities that did not exist. The judge speculated that AI may have been used. The firm said there was no finding about AI, that the attorney who filed it had left, and that it had added safeguards (Business Insurance).

Both cases were about legal citations. The same failure applies to facts from the record, and it's harder to spot there. A case that doesn't exist can be caught with a search. A date of injury that's off by a month, a diagnosis the doctor never made, or a TD rate copied from the wrong period looks exactly like the truth.

What the ethics rules ask

The ABA's Formal Opinion 512, issued in July 2024, says lawyers who use generative AI have to understand what a tool can and can't do, and review what it produces before relying on it. How much independent checking is enough depends on the tool and the task, but uncritical reliance isn't enough (ABA Formal Opinion 512). The State Bar of California's practical guidance says the same: a lawyer has to review, validate and correct AI output before using it (State Bar of California). Our guide to ABA Formal Opinion 512 for plaintiff firms goes through it duty by duty.

None of this is new law. A signature on a filing has always meant the lawyer checked it. What's new is how much text can arrive on a desk without anyone having read the record behind it.

Verification is the cost nobody prices

If an AI drafts a chronology in two minutes and checking it takes two hours, the firm saved very little and took on new risk. The real cost of an AI tool is the time it takes a person to confirm it got things right.

That time depends almost entirely on one thing: how far the reviewer has to travel from a statement to its source.

• No citation. The reviewer rereads the record to find where each fact came from, or doesn't, and hopes.
• A citation to the document. Better, but a 400-page record set is still 400 pages.
• A citation to the page and the words. The reviewer clicks, reads two lines and moves on.

The last one is the only version where review takes less time than writing it yourself. It's also the only version where a busy case manager will actually check every fact, every time, which is the point.

What a useful citation looks like

Not every "source" link is equal. A citation earns its place when it does these things:

1. Names the document, the page and the exact words. Not a summary of the page. The words on it.
2. Opens to the passage. One click from the fact to the evidence.
3. Shows conflicts instead of resolving them quietly. If two records give different dates, the reviewer should see both.
4. Separates what the record says from what was inferred. A gap in treatment is an inference. The dates on either side of it are facts.
5. Says what wasn't read. If a file couldn't be read, the reviewer needs to know the AI never saw it.

How it works in PlaintiffStack

PlaintiffStack is built around that kind of citation.

• Every fact an agent proposes cites the document, the page and the exact words it came from, so checking takes a click.
• A proposed fact doesn't land in the case until a person accepts it. Agents propose facts, deadlines and tasks; people decide.
• Drafts, such as a demand letter, a QME cover letter or a WCAB petition, cite the page each fact came from, and wait for review. Nothing goes to a client, a carrier or a court until a person sends it.
• Case tools flag what doesn't add up, like dates that disagree or bills with no records, and say where they found it.
• In California, deadlines are computed and shown with their rule and basis, so your team can confirm them.

And the limit, said plainly: agents read searchable PDFs, Word and text files, not scans or faxes. A file with no readable text is flagged in the document list. Scanned pages inside an otherwise searchable PDF aren't read or flagged yet, so a person still has to know what's in them.

See how documents are read and cited, how drafting works, or the AI questions in our FAQ.

A review habit that scales

Whatever tool your firm uses, a few habits keep AI drafting safe:

• Check every number and every date against its page. These are the errors that look most like the truth.
• Check legal citations separately. A citation to a record page tells you where a fact came from. A citation to a case or statute needs its own check against the source, whatever tool produced it.
• Read what the AI didn't. Know which files it couldn't read, and read them yourself.
• Make the reviewer's name part of the record. The person who accepted a fact or approved a draft should be on the record, not just the tool.

Fast drafting is only useful if review is fast too. A link to the page is what makes that possible.