AI tools invent citations because they predict text, not because they look things up.
A general language model writes the most likely next words. Asked for authority, it can produce a case name, a reporter citation, and a confident summary that fit the pattern of real citations without matching any real decision. Tools connected to a research database do better, because they retrieve documents before writing, but they can still misread a holding, attach the wrong citation to a statement, or quote language that is not in the opinion.
The errors fall into a few kinds. A fabricated authority does not exist. A misattributed one exists but does not support the proposition. A misquoted one exists, but the quoted words differ from the text. An outdated one was real and on point but has since been reversed, overruled, or superseded. Each needs a different check, and a citation can pass one check and fail another.
Lawyers have faced sanctions for filing briefs that cited cases that do not exist, and ABA Formal Opinion 512 makes clear that lawyers remain responsible for verifying what AI tools produce. The question is not whether to check, but how to check consistently, with the checking built into the work rather than left to memory when a filing is due.
Check each citation for existence, support, quotation, and status.
Existence: pull the authority from a research service the firm trusts, using the citation exactly as given. If it cannot be found, or the citation leads to a different case, treat the statement as unsupported. Do not ask the same AI tool whether the case is real; it can repeat the mistake with equal confidence.
Support and quotation: read the passage the draft relies on, not just a headnote or a summary. Confirm that the authority holds what the draft says it holds, in the right jurisdiction and procedural posture, and that each pin cite points to the page where the language appears. Compare every quoted phrase with the text word for word, including any omissions marked with ellipses.
Status: run the citator the firm uses to confirm the authority has not been reversed, overruled, or limited, and look for later decisions that change how it applies. Statutes and regulations need the same treatment against the current version and the version in effect at the relevant time. Record the result of each check on the draft, so the reviewing lawyer can see what was confirmed and by whom.
Build the check into the tool so unverified material never reaches the draft.
Manual checking works, but it depends on the time and attention of whoever is reviewing. A better design has the tool do the first pass: after drafting, the system extracts every citation and quotation, looks each one up in a trusted source, and compares the cited text with the statement it supports, all before a lawyer sees the draft.
What happens to a failure matters as much as the check. A citation that cannot be confirmed should not appear in the draft as if it were sound. It should be removed and flagged with a note explaining what was searched and what was not found, so the lawyer can research the point or drop it. That is what a hallucination-free design means in practice: unverified material is flagged, not drafted.
Grounding helps earlier in the process as well. When the tool retrieves from the firm’s briefs and memos and from a research service before writing, and links each statement to the passage it came from, there is less to fail and every check starts from a known document. Automated checks reduce the risk; they do not replace a lawyer’s judgment about whether an authority is the right one for the argument.
Make attorney review specific, recorded, and proportionate.
A lawyer signing a filing should know which citations the system confirmed, which it flagged, and which were added by hand afterward. Show that on the draft itself, with each citation linked to the source passage, so review means reading the authorities rather than hunting for them.
Scale the review to the use. A research memo for internal discussion needs its citations checked before anyone relies on it. A brief, a client letter, or anything else that leaves the firm needs a lawyer to read each cited passage. Summaries of depositions or productions need page references, so each statement can be confirmed in the original.
Keep a record of what was asked, which sources were used, what the checks found, and who reviewed the result. The record helps the firm supervise its use of AI, answer a client or a court that asks how a document was prepared, and spot patterns, such as a kind of question where the tool flags far more than it confirms.
Write the verification rules into the firm’s AI policy.
Verification should not depend on each lawyer’s habits. The firm’s AI use policy can state which tools are approved for research, that every citation is confirmed in a trusted source before use, who may sign off on AI-assisted work, and how supervising lawyers confirm that associates and staff follow the rules.
Training should show each kind of failure with examples from the firm’s practice areas, so people recognize a misattributed citation as readily as a fabricated one. New lawyers and staff should learn the checks as part of how the firm does research, not as an afterthought to using a tool. Baybloom builds legal AI that runs these checks before a draft reaches an attorney, and can help a firm set up the same checks around tools it already licenses.
Professional conduct rules on technology, competence, and supervision vary by jurisdiction. This guide is educational, not legal or ethics advice. Confirm the firm’s approach with its general counsel or ethics adviser and the guidance that applies where it practices.
Put it into practice
Talk through the project you are considering.
Bring the workflow, the systems around it, and the questions you need to resolve. We can help define a useful next step.
Discuss your project