How are LLMs used in legal work?

LLMs are speeding up legal work by handling first-pass drafts and searches, but every word still needs a lawyer's sign-off before it reaches a judge. Firms and in-house teams use them as fast, tireless research assistants — reviewing contracts, surfacing precedent, drafting routine documents, and sorting through discovery — not as replacements for legal judgment.

What legal tasks do LLMs handle?

The strongest fits are the high-volume reading and drafting jobs that used to swallow billable hours:

• Contract review and redlining: scanning a hundred-page contract in seconds to flag clauses that deviate from standard templates, catching unusual terms a tired eye might miss on page 80.
• Legal research: surfacing relevant precedent and statutes, cutting research that used to take hours down to minutes.
• Drafting routine documents: NDAs, standard filings, and client memos that start as an LLM draft a lawyer then edits, rather than a blank page.
• E-discovery: sorting through massive piles of litigation documents to find the handful that actually matter.

What can go wrong when lawyers use LLMs?

This is one of the few fields where AI mistakes have already caused real professional damage. There have been multiple publicly reported cases of lawyers submitting briefs that cited court cases the LLM simply invented — cases that never existed, with fake names and fake rulings, because the model was confidently making up something that sounded plausible. Those incidents led to sanctions and public embarrassment.

How should a lawyer verify LLM output?

Every AI-drafted legal document needs a licensed attorney to verify each citation and fact before filing — not a quick skim, but real verification against the actual case law. The LLM produces the draft; the lawyer confirms the cited law exists and says what the draft claims, and stays the one accountable to the court. Many firms now pair LLMs with tools that check citations against real legal databases, but the attorney signing the filing is the backstop, not the software.