“Show your work” is the only acceptable standard for AI in litigation
Every few months another sanctions order makes the rounds: a brief with hallucinated case citations, an expert report quoting a document that doesn’t exist, a chronology that collapses the moment opposing counsel checks a date. The lesson gets framed as “be careful with AI.” That’s the wrong lesson.
The right lesson is architectural: an AI output that cannot show its work is not evidence of anything. It’s a draft of a guess. The problem isn’t that lawyers used AI — it’s that they used AI systems designed to produce fluent text instead of systems designed to produce verifiable records.
Fluency is not the product. The record is.
Litigation runs on a simple chain of custody for ideas: a fact matters because a document says it, and the document matters because it’s in the record. Any tool that breaks that chain — summarizes without citing, asserts without linking, paraphrases without preserving the original — creates work you have to redo before you can rely on it.
That’s why we built FactMarshal citation-first rather than chat-first:
- Every extracted fact carries a pointer to the exact page and span it came from.
- Facts whose text is verbatim from the source are machine-verified at ingest — the system checks the string against the document, not against its own confidence.
- Everything else lands in a review queue where a human confirms or corrects it in one click.
- When the assistant answers a question, the answer is assembled from that verified record, and each sentence cites its way back down the chain.
Verification is a workflow, not a disclaimer
Most legal AI products handle reliability with a footer: “AI can make mistakes. Check important information.” That pushes the entire verification burden onto you, with no tooling to do it.
The alternative is to make verification cheap and visible. If checking a fact means clicking once and seeing the source page with the passage highlighted, verification stops being a chore and becomes the default reading mode. If a chronology distinguishes verified facts from unreviewed extractions, you always know which parts of the timeline can go in front of a judge.
The question to ask any legal AI vendor is not “how accurate is the model?” It’s “when the model is wrong, how fast will my team catch it?”
What this buys you in practice
A fact record with provenance changes what the downstream tools are worth. Contradiction detection is only useful if you can open both sources side by side. A deposition outline is only safe to use if each question’s factual predicate links to an exhibit. Settlement analysis is only credible if the facts it weighs are the ones you verified.
Show your work. It was the standard in law school, it’s the standard in court, and it should be the standard for the software in between.