Del Biaggio v. Bansen: Why Cite-Checking AI Is the Lawyer's Job, Not the Paralegal's

What Del Biaggio v. Bansen actually held

Del Biaggio v. Bansen (Cal. Ct. App., 1st Dist., Div. 4, July 10, 2026) made two rulings that matter here. First, paralegal fees are recoverable under a contract authorizing "attorneys' fees," reversing their categorical exclusion. Second, a lawyer must personally review AI-generated citations before filing; delegating that check to a paralegal does not meet the duty. Counsel was sanctioned $1,500 for fictional quotations ^[1].

The underlying dispute was a contract case. Daniel Del Biaggio won a $52,850 jury verdict against Pete and Mary Ann Bansen over a dairy-succession agreement, then sought $115,533 in attorney fees, including 86.1 hours of paralegal time. The trial court reduced the lawyer's hours and excluded all paralegal fees on the ground that the contract's "attorneys' fees" language did not reach them ^[1].

On appeal, the court reversed that exclusion. It held that fees for paralegal work are recoverable as a component of attorney fees under the contract's fee-shifting provision, consistent with a long line of California authority treating paralegal time as compensable legal work. That holding matters to the AI question because it establishes what a paralegal is: a skilled contributor whose billable work the law recognizes and pays for.

Then the court addressed the appellant's opening brief, which "rested in part on fictional quotations from the case law." It rejected the argument that the appeal was frivolous, but ordered Del Biaggio's counsel to pay a sanction to the court for the misstatements. The recognition of paralegal value and the refusal to let a paralegal shoulder the lawyer's verification duty sit side by side in the same opinion, and that is the point worth holding onto.

Why delegating the cite check to a paralegal failed the test

The court found the delegation plan improper on its own terms, apart from the miscommunication. Counsel intended to have his paralegal "simply check the citations before filing the brief." The court wrote that this "plan would have been inappropriate even if it had not gone awry," because the State Bar guidance places the review duty on the lawyer, and reading the cited authorities is a fundamental attorney obligation ^[1][2].

Floyd's declaration said his office uses generative AI "consistent with State Bar guidance," with "protocols requiring human verification of all outputs using primary sources." The court took that claim apart. Even if his paralegal had caught the additions, the protocol he described did not comply with the guidance he cited, which requires that a lawyer, not any human, review AI outputs before they reach the court ^[2].

The court was blunt about the arrangement. It gleaned from the declaration that Floyd "used generative AI to make additions to the opening brief during trial recesses in another case without ever planning to review personally, or even to have any other lawyer review, the case law he added." His plan was to have the paralegal check the citations. Being in trial, and therefore unavailable to supervise, made it "even less appropriate for him to expect his paralegal to assume responsibility for verifying what he wrote about cases he apparently had not laid eyes on himself."

There was also a signal the court thought he should have caught: the AI-supplied quotations articulated the exact proposition he wanted, "including purportedly from the California Supreme Court 26 years ago," yet had eluded him when he researched the same issue at trial. A source that appears only when a tool conjures it, and never in your own research, is a red flag, not a gift.

The sanction here rests on existing duties of candor and competent supervision, not on any new AI-specific rule. Del Biaggio applies the State Bar's Practical Guidance and long-standing authority; it does not announce a novel certification requirement.

Where the verification duty sits in the rules

The duty flows from the lawyer's candor and supervision duties, not a special AI rule. Rule 3.3(a)(1) forbids failing to correct a false statement of law made to a court. The State Bar guidance and Noland v. Land of the Free establish that reading the authorities you cite, to confirm they stand for what you claim, is a fundamental duty that AI does not shift onto staff ^[2][3][4].

The court anchored the reading duty in Noland v. Land of the Free, L.P. (2025), quoting its statement that "it is a fundamental duty of attorneys to read the legal authorities they cite in appellate briefs or any other court filings to determine that the authorities stand for the propositions for which they are cited" ^[4]. That duty predates generative AI by decades. It applied to associates, to commercial research services, and now to model output, without changing shape.

The candor duty carried its own weight in the sanction. The State Bar guidance, quoted by the court, requires counsel to "correct any errors or misleading statements made to the court, regardless of whether such outputs were generated with or without real-time human direction," which the court paired with Rule 3.3(a)(1) ^[2][3]. Floyd eventually filed a letter withdrawing the misstatements, but it came only days before oral argument and more than three months after the opposing motion put him on notice. The court rejected his claim that the delay caused no prejudice.

Read together, the rules describe a chain that always ends at the signing lawyer. A paralegal can research, draft, and format, and the firm can bill for that work. What the lawyer cannot delegate is the certification that the authorities are real, accurately quoted, and still good law. That is the specific act Del Biaggio protects.

The pairing that makes this opinion unusual

Most AI-sanctions opinions do one thing: they punish a lawyer for unverified output. Del Biaggio does two, and they point in opposite directions. It expands what a firm can recover for paralegal work, holding 86.1 hours of paralegal time compensable under a contract's "attorneys' fees" clause, while simultaneously refusing to let a paralegal carry the verification duty ^[1].

It would be easy to read the verification holding as a demotion of paralegals. The opinion says the opposite, and it says it first. The court reversed the trial court's categorical exclusion of paralegal fees, holding that work billed by paralegals is recoverable as a component of attorney fees. The firm had sought $115,533 in fees including 86.1 hours of paralegal time, and the trial court had zeroed that component entirely on the theory that "attorneys' fees" means fees for attorneys ^[1].

That reversal is a statement about what paralegal work is worth. The law treats it as skilled legal labor that a paying party can be ordered to reimburse. A court that thought paralegals were clerical would not have written that half of the opinion.

So the verification holding is not about competence or status. It is about who signs. The signature on a brief is a certification, and a certification can only be made by the person whose name is on it. That is why the same court can say, in one opinion, that a paralegal's 86.1 hours are worth real money and that a paralegal cannot perform the lawyer's final read.

The practical consequence for a firm is a clean division rather than a restriction. Paralegal time spent assembling a verification record is billable, recoverable work that strengthens the file. The lawyer's read on top of it is the part that cannot be moved. A firm that structures its workflow that way loses nothing economically and gains the one thing Del Biaggio actually demands.

Who gets sanctioned when an AI citation is fabricated

The signing lawyer does. Del Biaggio ordered counsel to pay $1,500 to the court for the fictional quotations ^[1]. It sits in a growing California line: People v. Alvarez imposed $1,500 for a fictitious quotation and a nonexistent case, and Schlichter v. Kennedy imposed $1,750 for fabricated authority where the court found denials of AI use not credible ^[1].

The dollar figures are modest; the pattern is not. The Del Biaggio court cited its own recent authority in a single string: Alvarez at $1,500, Schlichter at $1,750, and the Ninth Circuit's LNU v. Blanche, which sanctioned two attorneys $2,500 each for briefs with "multiple nonexistent cases, misattributed quotations, and gross misrepresentations of real cases" and then denying that AI might have produced the errors ^[1]. Public trackers such as the Damien Charlotin AI Hallucination Cases database catalog well over a thousand United States filings involving AI-fabricated or misused citations as of mid-2026, a count best read as approximate and rising ^[6].

Notice what draws the sanction in each case. It is not that a tool was used. It is that unverified output reached the court under a lawyer's name, and, in several instances, that the lawyer was slow or less than candid about the failure once it surfaced. Alvarez and Schlichter both turned on unverified AI citations; Del Biaggio added the delegation wrinkle and the delayed correction.

For a signing attorney, the takeaway is that the reputational cost now outruns the monetary one. A published opinion naming you, quoting your fictional citations, and explaining that your verification protocol was inadequate is a durable record. The way to stay out of that line is not to avoid AI. It is to make the lawyer's verification real and provable.

The correction clock: why the delay cost more than the error

Del Biaggio was not sanctioned only for filing fictional quotations. The court weighed the fact that counsel withdrew the misstatements just days before oral argument, more than three months after the opposing sanctions motion put him on notice, and it rejected his argument that the delay caused no prejudice ^[1].

There are two failures in this case, and firms tend to fixate on the first. The first is that unverified AI output reached a filing. The second is what happened in the three months after someone pointed it out. The court treated both as sanctionable, and the second is the one entirely within a lawyer's control after the fact.

The State Bar guidance the court quoted is explicit that counsel must "correct any errors or misleading statements made to the court, regardless of whether such outputs were generated with or without real-time human direction" ^[2]. Paired with Rule 3.3(a)(1), which forbids failing to correct a false statement of law to a tribunal, the duty does not wait for a convenient moment ^[3].

Floyd did eventually file a letter withdrawing the misstatements. The timing is what the court noticed: days before argument, after more than three months of notice. He argued the delay produced no prejudice. The court disagreed, and the sanction covers both the misstatements and the failure to correct them promptly ^[1].

This reframes what a firm's AI incident response has to do. Discovering a fabricated citation in your own filed brief is bad but survivable. Sitting on it converts a verification failure into a candor problem, and candor problems are the ones that generate published opinions with your name in them. The correct move on discovery is immediate written notice to the court and opposing counsel, before anyone has to ask twice.

It is worth noting what the court did not do. It rejected the argument that the appeal was frivolous and vacated a separate trial-court sanction ^[1]. This was not a court looking for reasons to punish. It sanctioned the specific conduct it found culpable: fictional authority, an inadequate verification plan, and a slow correction.

1. Stop and confirm the defect yourselfRead the authority in question personally. Do not accept a staff report or a tool's second opinion as the finding. You are about to make a representation to a court about your own filing.
2. Notify the court and opposing counsel in writing, immediatelyThe duty under Rule 3.3(a)(1) and the State Bar guidance runs from the moment you know. Del Biaggio shows that a correction filed months later, on the eve of argument, does not read as diligence.
3. State plainly what happened and howThe Ninth Circuit in LNU v. Blanche sanctioned attorneys who denied AI involvement, and Schlichter v. Kennedy involved denials the court found not credible. Candor about the cause is treated better than a narrow correction that obscures it ^[1][7].
4. Re-verify the entire filing, not just the flagged citationOne fabricated authority in a brief is evidence about the process that produced the brief. Run the three checks across every cited authority before you represent that the filing is now clean.
5. Record what you fixed and whenThe remediation record is what answers the court's next question. A verification log that shows who re-checked each authority and on what date turns an assertion into something you can hand over.

A verification workflow a signing lawyer can actually run

Personal verification is not re-researching the brief. For every cited authority, the signing lawyer confirms three things: that it exists and is correctly identified, that the quotation and characterization are accurate, and that it is still good law. A paralegal can prepare and organize this record; the lawyer performs and attests to the final review ^[2][4].

The phrase "the lawyer must review" sounds like a demand to redo everything. It is narrower than that. The reading duty in Noland asks the signer to confirm, for each authority, three specific things, the same core checks in our guide on verifying AI-generated case citations before you file, and to be able to show that the confirmation happened.

First, existence and identity: the case, statute, or rule is real and cited correctly. A fabricated authority fails here before any question of what it holds. Second, accuracy: the brief says what the source actually says, not what a model asserted it says. Reliability research underscores why this cannot be skipped; a Stanford study of leading legal AI research tools found inaccurate answers on a meaningful share of queries, with figures reported around 17 percent for one product and 33 percent for another in 2024 testing ^[5]. Third, good law: the authority has not been overruled, reversed, or superseded.

The delegation line from Del Biaggio maps cleanly onto this. A paralegal can pull the sources, assemble the pincites, run the citator, and hand the lawyer an organized packet. The lawyer reads the authorities, confirms the three checks, and signs. What Del Biaggio forbids is the lawyer skipping that read and treating the paralegal's pass as the certification. This is where a verification system earns its place: RankShield Legal certifies which cited authorities exist, are quoted accurately, and remain good law, and records who ran the final review and when, so the lawyer's verification is not just performed but provable if a court asks.

What this means if you do not practice in California

Del Biaggio binds California courts. Its reasoning travels further, because it rests on duties that are not California-specific: the duty to read what you cite, and the duty to correct false statements of law. The Ninth Circuit reached the same place in LNU v. Blanche, sanctioning two attorneys $2,500 each and suspending them for six months ^[1][7].

The holding is jurisdictionally limited and this article does not suggest otherwise. A lawyer in Texas or New York is not governed by a California Court of Appeal decision, and any specific question about a state's supervision rules belongs with a licensed attorney in that state.

What travels is the structure of the reasoning. Del Biaggio did not invent an AI rule. It applied the duty to read the authorities you cite, drawn from Noland, and the duty of candor under Rule 3.3(a)(1) ^[3][4]. Both are versions of Model Rules adopted in substantially similar form across the great majority of United States jurisdictions. A court elsewhere reaching this question is likely to reach for the same two duties, because they are the ones that fit.

The federal picture points the same direction. The Ninth Circuit's decision in LNU v. Blanche, published and precedential, sanctioned two attorneys $2,500 each, suspended them from practice before the court for six months, and required disclosure of generative AI use in future filings ^[1][7]. That is a materially heavier consequence than California's $1,500, reached through the same logic about unverified authority and lack of candor.

The volume question settles the rest. Public trackers catalog well over a thousand United States filings involving AI-fabricated or misused citations as of mid-2026 ^[6]. A firm waiting for its own jurisdiction to publish an opinion before changing its workflow is betting that it will not be the case that generates one.

The practical posture, then, is the same everywhere: treat the signing lawyer's read as non-delegable, treat a discovered error as an immediate correction obligation, and keep a record that shows both happened. None of that depends on which state's opinion you are reading.