About 1,600 court filings with AI-invented material have now been logged worldwide.

and hallucinations in filings rose EIGHTFOLD in a year.

That's from Damien Charlotin's tracker, as reported by AFP in June. So if you're a law student picking up legal AI tools this summer, the tool matters less than your checking routine.

The legal AI tools law students will actually meet

Here's the landscape you'll bump into on placements, vac schemes and (more and more) in law school:

  1. Harvey: the big-firm assistant. Harvey runs a "Preparing for Legal Work" course for students whose school gives them access, plus a tool-agnostic "AI for Legal Basics" course for everyone else (Harvey)
  2. Legora: Harvey's main rival. Cornell Law announced partnerships with both in May
  3. Vincent AI (Clio vLex): Cornell's existing partner before those deals
  4. CoCounsel: now owned by Thomson Reuters (the Westlaw people)
  5. Lexis+ AI: LexisNexis's assistant, sitting on its own case law database
  6. General chatbots (ChatGPT, Claude, Gemini): everywhere, cheap or free, and the riskiest for legal research

That last point isn't us being snobby.

Charlotin found lawyers and litigants mostly used mainstream chatbots, especially ChatGPT, when things went wrong. Specialised legal tools look more reliable, but they're not immune (AFP).

And the benchmark data backs that up. A Stanford study found:

→ Lexis+ AI and Ask Practical Law AI hallucinated on more than 17% of test queries → Westlaw's AI-Assisted Research on more than 34% → general chatbots on 58% to 82% of legal queries, in the authors' earlier work

So better, not bulletproof.

Why "the tool said so" is no defence

Two recent cases, same lesson:

  1. Louisiana, May. A lawyer used Claude to draft a brief. A law clerk flagged fake quotes, the lawyer asked Claude to fix them, then filed without checking. Seven quotes turned out to be missing from the rulings or misrepresented. Result: a $1,000 fine and a three-hour course on AI in practice
  2. Mississippi, June. Judge Sharion Aycock sanctioned lawyers on BOTH sides of the same case. $8,000 in fines in total, and two of them barred from that court for two years

The Louisiana judge put it bluntly: "Ignorance of the risks of AI usage is no longer an excuse" (AFP).

The UK isn't different. In Ayinde the High Court referred solicitors to the SRA (the solicitors' regulator) after fake citations turned up in submissions. We went deeper on the sanctions wave in our piece on hallucinated citations.

What law school won't teach you about legal AI

→ legal research classes teach you to FIND authority, not to audit someone else's → most uni AI policies say "don't cheat", not "here's how to verify" → moots and problem questions rarely test whether a quote actually appears in the judgment

Meanwhile Harvey's own advice to students puts "build strong verification habits" near the top of its list (Harvey). Even the vendor is telling you to check the vendor.

How to check legal AI output: a 5-step routine

Use this every time, on every tool:

  1. Pull every authority. Open each case in a primary source: Find Case Law (the National Archives' free judgments site) or BAILII for the UK, CourtListener for the US. No source, no sentence
  2. Search for the exact quote. Ctrl+F the words inside the judgment. The Louisiana brief failed on seven quotes, not on made-up case names
  3. Check it's still good law. Use your library's Westlaw or Lexis citator (the feature that shows whether a case was appealed, overruled or followed)
  4. Ask what it didn't read. Which jurisdiction? Which dates? Which documents were outside its view?
  5. Never ask the AI to fix its own fake. That's exactly how the Louisiana lawyer got fined. Re-check by hand

What to do this month

  1. Find out what your law school licenses. Check the library page or ask a law librarian. Paid tools are often free while you're a student
  2. Run a hallucination drill. Ask a general chatbot for five cases on a topic you know well, then check each one on Find Case Law or BAILII. Count the misses
  3. Read one sanction order end to end. Pick one from Charlotin's database. Honestly it's the best warning label there is
  4. Keep a verification log of every AI error you catch. Same principle as studying with AI without outsourcing your thinking: you keep the reasoning, the tool just speeds you up

More on the profession in our law hub.

so TLDR: learn the tools, but the skill firms will pay for is checking them. Pull the authority, find the quote, every time.

Which legal AI tool does your law school give you, and have you ever caught it inventing something?

Sources

Get the next issue by email

theVANTAGEd is written by students to bring critical thinking to AI across real industries. One issue every two weeks.