Why Verification Is Non-Negotiable
The duty of competence requires attorneys to understand the tools they use. The duty of candor to the tribunal requires that cited authority be real and accurately characterized. Both of these obligations apply when you use AI for legal research — and neither can be satisfied by trusting that the AI got it right. Every AI tool currently available for legal research has a non-zero hallucination rate. The question is not whether you should verify AI output, but how to do it efficiently enough that the time savings of AI still apply.
Step One: Know Which Type of AI Tool You Are Using
Not all legal AI tools carry the same risk profile. The first step in any verification workflow is understanding what kind of tool you are working with. General-purpose language models — ChatGPT, Claude, Gemini used directly — generate citations from training data and should never be used for citation generation without independent database verification. Purpose-built legal research tools that retrieve from verified case databases have a different, lower risk profile for citation accuracy, though analytical accuracy still requires verification.
The key question to ask about any tool: does it retrieve from a verified database, or does it generate from memory? If you cannot answer that question about a tool you are using, find out before you rely on its output.
Step Two: Check the Grounding Score if Available
Some legal AI tools now provide grounding scores or confidence metrics that tell you how closely the AI's analysis tracks its source material. CaseMatch AI's Hallucination Check assigns a percentage score to each AI-analyzed case — a "Verified · 90%" badge means 90% of the AI's extracted claims map directly to language in the original court opinion. When this information is available, use it to prioritize your reading. High grounding scores mean you need to read the source to confirm nuance; low grounding scores mean you should read the source before relying on anything the AI said about it.
Step Three: Verify Citations Exist
For every case you intend to cite in a filed document, run the full citation — case name, reporter, volume, and page number — against at least one verified legal database. This takes thirty seconds per citation and eliminates the risk of filing a fabricated case citation entirely. This step is non-negotiable regardless of which tool you used to find the case.
Step Four: Read the Holding for Every Case You Plan to Cite
The AI's summary of what a case holds is a starting point, not an endpoint. For any case you intend to use as authority in a brief, argument, or client advice, read the actual holding in the opinion. You are looking for three things: whether the court actually said what the AI claims it said, whether the holding was narrow or broad, and whether there is language in the opinion that cuts against your position that the AI did not surface.
This does not mean reading every case the AI retrieved. It means reading the cases you intend to use. The AI's retrieval and initial filtering can still save you significant time even if your final verification step involves reading the opinions that make your shortlist.
Step Five: Document Your Verification
As AI use in legal practice attracts more bar association attention, being able to demonstrate that you verified AI output may become an important part of demonstrating competent practice. A simple internal note — "citation verified in Westlaw on [date]" or "holding confirmed against full opinion" — creates a record that the AI was used as a starting point, not a final authority. Some firms are building this documentation into their matter management workflows.
Building Verification into Your Workflow, Not Onto It
The verification steps above are not additional work on top of AI-assisted research — they are a replacement for work you would have done anyway if you had researched manually. The goal is not to use AI and then do all the same manual research you would have done without it. It is to use AI to compress the search and shortlisting phase dramatically, then apply your verification effort to a much smaller set of high-confidence candidates. Tools that provide grounding scores help by telling you where that verification effort is most needed.