AI Hallucinations and Other Pitfalls in Legal Work
Invented citations get the headlines, but they are only one of several ways AI can trip up legal work. Here is how each happens and what to do about it.
Generative AI predicts plausible text. Most of the time plausible and accurate overlap. When they do not, you get what are known as AI hallucinations: confident statements, citations or quotations with no basis in fact. In legal work, where a single wrong citation can embarrass a firm or harm a client, understanding why these errors happen is as important as knowing how to catch them.
This guide explains hallucinations and five other pitfalls that show up when lawyers use AI, with a safeguard for each. It complements a verification checklist by focusing on why things go wrong, so you can anticipate problems rather than only catching them afterwards.
This article is general information, not legal advice.
Pitfall 1: Hallucinated authority
What happens
Language models generate text one piece at a time based on patterns. Asked for supporting cases, a model may produce citations with realistic names, reporters and page numbers that do not exist, or attribute a holding to a real case that never said it.
Why it matters
Courts in several US jurisdictions have sanctioned lawyers for filing briefs containing fictitious citations generated by AI. Professional and court rules place the duty to verify on the lawyer. Federal Rule of Civil Procedure 11, available on Cornell's Legal Information Institute, is one example of a rule requiring reasonable inquiry into legal contentions.
Safeguard
Never rely on an authority you have not located and read in a trusted source. Be most suspicious when a case fits your facts perfectly.
Pitfall 2: Stale or incomplete law
What happens
General models are trained on data up to a cut-off and may not reflect recent statutes, amendments or decisions. Even tools connected to current sources may have coverage gaps for particular jurisdictions or courts.
Safeguard
Ask yourself what has changed recently in the area. Check the effective date of every statute and run citator checks on every case. Note the tool's stated coverage and cut-off in your research log.
Pitfall 3: Misplaced confidence
What happens
AI output reads fluently regardless of accuracy. There is often no signal in the tone that distinguishes a solid answer from a fabricated one. Users under time pressure tend to trust fluent text.
Safeguard
Ask the tool to separate what it is citing from what it is inferring, and to state where it is uncertain. Treat the output as a draft from a capable but unreliable assistant, not as an authority.
Pitfall 4: Confidentiality leaks
What happens
Client information entered into a tool may be stored, reviewed by vendor staff or used for training, depending on the tool's terms. Staff using personal accounts can expose client data without realising it.
Safeguard
Use only tools approved for client information after reviewing their data terms, and maintain a clear list of what is allowed. The ABA's Model Rule 1.6 on confidentiality frames the reasonable-efforts standard in most US jurisdictions.
Pitfall 5: Losing the thread in long documents
What happens
In long drafts or large document sets, AI tools may drift: changing defined terms, contradicting earlier sections, missing a document or summarising the wrong version.
Safeguard
Work in smaller sections, provide definitions upfront and check consistency afterwards with searches for each defined term. For document review, spot-check summaries against originals and confirm that every document was processed.
Pitfall 6: Deskilling and over-reliance
What happens
If juniors always start from AI output, they may not develop the research and drafting skills needed to recognise when that output is wrong. Over time the team's ability to check the machine weakens.
Safeguard
Keep training on core skills, and require juniors to explain the reasoning behind AI-assisted work in their own words. Rotate tasks so people still do some work from scratch.
A quick reference
| Pitfall | Early warning sign | First response |
|---|---|---|
| Hallucinated authority | Unfamiliar reporter or perfect fact match | Locate and read the source |
| Stale law | Recently active area of law | Check effective dates and citators |
| Misplaced confidence | No citations or vague sourcing | Ask for sources; verify independently |
| Confidentiality leak | Unclear vendor terms, personal accounts | Stop; check approved-tools list |
| Document drift | Inconsistent terms or missing items | Section-by-section review |
| Over-reliance | Cannot explain the reasoning | Rework without the tool |
Reducing risk with tool choice
Some risks shrink with the right tool. Tools that ground answers in identifiable sources and show citations make hallucinations easier to detect. eLaw provides cited answers on firm-isolated data across legal research and document analysis. That helps with checking and confidentiality, but none of the safeguards above become optional.
Turn this into team practice
- Share the quick-reference table with everyone who uses AI on client work.
- Add a "tool used and verified by" line to your matter templates.
- Review your approved-tools list against Pitfall 4.
- Schedule a training session on spotting fabricated authority.
FAQ
Why do AI tools make up legal citations?
General-purpose language models generate text based on patterns rather than retrieving verified records. When asked for authority, they can produce text that looks like a citation without checking that the source exists. Tools that retrieve from defined legal databases reduce this risk but do not eliminate it.
Are some AI tools free of hallucinations?
No tool can guarantee error-free output. Tools grounded in specific legal sources and showing citations generally make errors easier to catch, but they can still misread holdings, miss recent changes or summarise incorrectly. Verification is still needed whatever the vendor claims.
What should I do if I find a hallucinated citation in work already sent?
Act quickly. If it was filed with a court, consider your candour obligations and whether a correction is needed, and inform the supervising lawyer. If it went to a client, correct the advice. Then review other work from the same session for similar errors and record what happened.
How can firms train lawyers to spot AI errors?
Practical exercises work well: give lawyers AI output containing planted errors and ask them to find them. Combine that with refresher training on citators and reading holdings closely. Reviewing real near-misses within the firm, without blame, also helps people recognise warning signs.
