A 5-Step Citation Verification Protocol for AI-Generated Legal Research
Verification takes about 10-15 minutes per citation. Sanctions take months. The math is straightforward.
I build software systems that generate text. I understand, at a technical level, why they hallucinate. And I can tell you with confidence: verification is not optional. It is not a best practice. It is the price of admission for using these tools professionally.
The legal profession has accumulated over 160 documented cases of hallucinated citations in court filings since 2023. Every single one followed the same pattern: an attorney trusted AI output without adequate verification. The fabrication was discovered after submission. Sanctions followed.
Here is the protocol I recommend. It works whether you are using ChatGPT, Lexis+ AI, Westlaw AI-Assisted Research, or any other generative tool.
The Baseline Reality
Even purpose-built legal AI tools produce errors at significant rates. Stanford's study found Lexis+ AI hallucinated on 17% of queries. Westlaw AI-Assisted Research hit 33%. General-purpose models like ChatGPT hallucinate on legal queries 58-82% of the time.
Let me translate that into engineering terms. If you are using Westlaw's AI and you generate 10 citations, statistical expectation says 3 of them may be wrong. If you are using ChatGPT, 6 to 8 of them may be wrong.
These are not occasional edge cases. They are baseline system performance. Building a workflow that assumes the output is correct and only verifies when something "looks wrong" is building on a statistically unsound foundation.
Step 1: Flag Every AI-Generated Element
Before you verify anything, you need to know what requires verification. This step is about creating a clean separation between what the AI produced and what you produced.
Flag every case citation — name, reporter, volume, page. Every quotation attributed to a case, statute, or regulation. Every factual claim about holdings, procedural history, or statutory language. Every reference to dates, parties, or procedural posture. Every secondary source citation.
Practical method: work from a clean copy of the AI output. Highlight or bracket every verifiable claim. I recommend a different color for each category — citations in yellow, quotations in blue, factual claims in green. Do not rely on memory to track what came from AI versus your own research. Memory is unreliable, and the stakes are too high for unreliable.
Time investment: 2-5 minutes per page of AI output.
Step 2: Verify Citation Existence
Before analyzing whether a citation supports your proposition, confirm it exists. This is the most basic check and catches the most egregious hallucination type: citations to cases that were never decided.
For case citations: check the full citation in Westlaw, Lexis, Google Scholar, or CourtListener. Verify the case name matches exactly. Confirm the reporter, volume, and starting page are correct. Check that the year corresponds to the actual decision date.
For statutes and regulations: access the official code or register. Verify the section number exists. Confirm the language matches current law.
Red flags suggesting fabrication: case name combinations that seem suspiciously on-point for your issue (AI is very good at generating plausible-sounding case names that address exactly the question you asked — too good, in fact). Reporter-volume combinations that do not exist. Courts or jurisdictions that seem unusual for the subject matter. Dates that do not align with the court's history.
Time investment: 1-3 minutes per citation.
Step 3: Validate Substantive Accuracy
A citation that exists may still be mischaracterized. This is the misgrounding problem from the Stanford study, and it is more dangerous than outright fabrication because it passes a superficial check.
For cases: read the actual holding, not just the headnotes. Verify any quoted language appears in the opinion verbatim — AI frequently paraphrases while using quotation marks, which is a form of fabrication that looks like faithful quotation. Confirm the case has not been overruled, distinguished, or limited. Check that the procedural posture matches the AI's characterization.
For page-pinpoint citations: navigate to the specific page and confirm the referenced language appears there. This is tedious. It is also the step that catches the most misgrounding errors.
Time investment: 3-10 minutes per citation, depending on complexity.
Step 4: Assess Current Authority
Even an accurate citation may represent bad law. This step prevents the AI from citing a case that was correct when it entered the training data but has since been overruled, superseded, or distinguished into irrelevance.
Run Shepard's or KeyCite on every case citation. Review for negative treatment. Check for subsequent legislation that supersedes case law. Verify regulatory provisions remain in effect.
Pay attention to superseding statutes, circuit splits where your jurisdiction differs, recent amendments, and pending legislation or rulemaking that may change the landscape.
Document your findings with dates. If significant time passes between verification and filing, re-run these checks. Law changes. Training data does not.
Time investment: 2-5 minutes per case citation.
Step 5: Document Your Verification
Create a record of what you verified and when. This serves three purposes: it ensures completeness during the current project, provides protection if questions arise later, and establishes institutional knowledge for supervision and training.
Documentation should include the date of the AI query, the tool used, specific verification steps performed for each citation, results of currency checks, and any discrepancies identified and how they were resolved.
The appropriate level of documentation depends on the stakes. Routine correspondence may warrant minimal notes. Court filings and dispositive motions warrant comprehensive records. When in doubt, err toward more documentation. You will never regret having it. You may regret not having it.
Time investment: 5-10 minutes to compile documentation for a typical research memo.
The Total Time Investment
For a research memorandum containing 10 AI-generated citations:
| Step | Per Citation | Total |
|---|---|---|
| Flagging | 30 seconds | 5 minutes |
| Existence check | 2 minutes | 20 minutes |
| Substantive verification | 5 minutes | 50 minutes |
| Currency check | 3 minutes | 30 minutes |
| Documentation | — | 10 minutes |
| Total | ~2 hours |
Approximately 115 minutes for 10 citations. Without AI assistance, the same research would likely take 4-8 hours. The net saving is real — but only if you actually do the verification.
When to Skip Verification
Never.
I am not being rhetorical. Given documented hallucination rates of 17-33% for specialized legal tools and 58-82% for general-purpose models, skipping verification means accepting a statistically significant probability of submitting fabricated content.
Some practitioners argue that verification of routine matters can be relaxed. My counterargument: the Mata v. Avianca case involved routine research on a statute of limitations question. A well-established area of law. The kind of thing where you might think "this is so basic, the AI surely got it right."
It did not.
If a citation appears in work product that leaves your office, it requires verification. End of discussion.
Scaling by Work Product Type
Quick internal research queries: Steps 1-2 mandatory. Steps 3-5 proportionate to stakes.
Client correspondence: Steps 1-4 mandatory. Step 5 recommended.
Court filings: All five steps mandatory, enhanced documentation, and consider second-reviewer verification for dispositive motions.
Published materials: All five steps mandatory with extended currency monitoring through publication date.
The Engineering Perspective
I build systems that generate text. I understand the architecture. Large language models are statistical pattern matchers that produce the most probable next token given the preceding context. They do not verify against external databases during generation. They do not know whether a citation exists. They generate character sequences that look like citations because citations appear in their training data.
This is not a bug that will be fixed in the next model update. It is an inherent property of the architecture. Better models will hallucinate less frequently. No model in the current paradigm will hallucinate zero percent of the time.
Your verification protocol is not a temporary measure while the technology improves. It is a permanent feature of working with probabilistic text generation systems. Design accordingly.
Key Takeaways
- Budget 10-15 minutes per AI-generated citation for thorough verification — still substantially less than conducting research without AI
- The five-step protocol covers the full verification surface: flag, confirm existence, validate substance, assess currency, document
- Never skip verification — hallucination rates of 17-33% make it a statistical certainty that some citations will be wrong
- Misgrounding (real case, wrong proposition) is more dangerous than fabrication because it passes superficial review
- Verification is not a temporary measure — it is a permanent feature of working with probabilistic text generation systems