The Lawyer Cannot Delegate the Part That Makes the Argument Legal.

He won important parts of the appeal

The underlying case had nothing to do with artificial intelligence. Daniel Del Biaggio prevailed in a dispute arising from a dairy operations succession agreement and sought contractual attorney fees. The trial court reduced the requested lawyer hours, categorically excluded paralegal fees, and imposed fees against counsel in connection with a motion for reconsideration.

On July 10, California's First District Court of Appeal affirmed the reduction in lawyer hours but reversed the categorical exclusion of paralegal fees. It also vacated the trial-level sanctions. The court held that the contractual authorization for attorneys' fees could encompass reasonable paralegal fees and concluded that the reconsideration motion was procedurally permissible and meritorious.

That context matters. This was not a meritless appeal assembled by a machine. Counsel was right about consequential issues. The court nevertheless ordered him to pay $1,500 to the clerk because the opening brief misrepresented case law through fictional quotations. A legally correct destination did not excuse an unreliable route.

The fabrications carried the argument

The false material was not decorative. The brief attributed invented language to PLCM Group, Inc. v. Drexler, a real California Supreme Court case, and to another real appellate decision. The quotations appeared to say exactly what counsel needed them to say about recovery of paralegal fees. One additional authority was misidentified.

Counsel acknowledged using generative AI and told the court that he had asked a paralegal to verify the additions before filing. According to his explanation, the paralegal believed he had already completed that review. He characterized the failure as a communication error within an office that otherwise required human verification against primary sources.

The Court of Appeal found more than a broken handoff. It said the protocol itself was inadequate because the planned reviewer was a paralegal rather than counsel or another lawyer. Citing California's AI guidance and a prior appellate decision, the panel described reading the authorities relied upon in a court filing as a fundamental and nondelegable attorney duty.

Delegation is not the problem

Read too broadly, the decision could produce an unhelpful rule: if AI touched the work, a lawyer must personally repeat every verification step. That would confuse accountability with duplication. Law firms properly delegate factual research, cite formatting, quotation comparison, docket checks, record references, and many other quality-control tasks. California Rule 5.3 expressly contemplates work by nonlawyer assistants, provided lawyers establish appropriate measures and supervision.

AI should not erase that operating model. A capable paralegal can determine whether a case exists, whether the reporter citation is correct, and whether quoted words appear on a particular page. Those tasks do not become mystical because a language model proposed the citation first.

The limit is professional judgment. Whether an authority governs, whether its procedural posture matters, whether later history weakens it, whether the quoted passage is holding or dicta, and whether it honestly supports the proposition are not merely citation checks. They are the legal analysis the lawyer presents under the lawyer's name.

Existence, accuracy, and support are different tests

Many firm protocols use the word verify as though it describes one act. It describes at least three. Existence asks whether the authority is real. Accuracy asks whether the citation, quotation, and record reference are correct. Support asks whether the authority actually justifies the legal proposition for which it is offered.

The first two can often be documented mechanically. Open the opinion from a reliable source. Match the court, date, reporter, page, and quoted language. Check subsequent history. Record who performed the review and when.

The third test requires reading. A sentence can be quoted perfectly and still be misleading when removed from a dissent, a party's argument, a factual summary, or a passage the court later rejected. A real case can be cited accurately for a proposition it does not establish. Generative AI increases the number of nonexistent authorities, but the deeper professional risk has always been superficial reading.

Del Biaggio is therefore not best understood as an anti-delegation case. It is a warning against delegating the conclusion. Support cannot be confirmed by checking that words exist on a page. The lawyer must decide whether those words do the legal work the brief claims they do.

The human in the loop needs authority

Technology policies often promise that a human remains in the loop. The phrase sounds responsible while revealing almost nothing. Which human? Reviewing what? With what training, source access, time, and authority to stop the filing? A person placed after the model but before the filing button is not automatically a safeguard.

In this case, the proposed review failed at two levels. The assignment was misunderstood, and the person assigned the task was not the person the court believed had to exercise the relevant judgment. A workflow diagram would have shown a human checkpoint. It would not have shown that the checkpoint was capable of discharging counsel's duty.

A defensible process identifies the decision owner, not merely the reviewer. Staff can validate objective attributes and escalate discrepancies. The lawyer responsible for the argument must read the central authorities, resolve the discrepancies, and decide that the proposition is supportable. The final human cannot be a name added to satisfy policy language.

The correction came too late

The fabricated material was only part of the court's concern. Counsel learned of the problem but waited more than three months to withdraw the misstatements, filing a corrective letter shortly before oral argument. The panel tied that delay to the duty of candor reflected in California Rule 3.3.

This is a separate systems failure. Verification protocols focus on preventing an erroneous filing. Firms also need an incident-response rule for what happens after one is discovered. Who assesses materiality? Who informs the client? Who decides whether to notify opposing counsel or the tribunal? How quickly must the corrected filing or letter be prepared? A policy that ends at submission is incomplete.

Delay is especially dangerous when the original mistake implicates credibility. A prompt correction demonstrates that the lawyer's loyalty is to the integrity of the proceeding, even when disclosure is embarrassing. Waiting until the court or an adversary forces the issue makes an avoidable error look like a choice.

A protocol that respects how legal work is actually done

The answer is not to prohibit delegation or require partners to recreate every research path. It is to divide the work honestly. Any AI-proposed authority should be opened from a reliable legal source rather than trusted from the generated text. Staff may confirm identity, citation form, quotation accuracy, record location, and negative treatment. Exceptions should be marked, not silently repaired.

The lawyer using the authority should read enough of the decision to understand the issue, posture, reasoning, and limits. Authorities carrying the central proposition deserve full review. The lawyer should then confirm the proposition-to-authority relationship in the brief, not merely approve a list of citations. A second lawyer can perform that judgment where responsibility is deliberately transferred and supervision is clear.

Finally, the firm should preserve a short verification record and a correction path. That record is not bureaucratic theater. It prevents the phrase ‘I thought someone else checked it’ from becoming the firm's explanation to a court.

AI did not create the duty

Courts are not inventing a special professional obligation for artificial intelligence. Lawyers have always been responsible for reading the authorities they ask judges to follow. AI simply exposes how often the profession treated a citation as a formatting object rather than the foundation of an argument.

The useful lesson from Del Biaggio is narrower than ‘never trust AI’ and more demanding than ‘keep a human in the loop.’ Delegate the checking that can be checked. Supervise it. Document it. But the part that asks whether the law supports what the brief tells the court remains legal judgment.

That is not because paralegals are incapable. It is because the lawyer is the person asking the tribunal to act on the argument. Responsibility should stay where authority sits.

General educational information only. Del Biaggio v. Bansen applies California law and addresses a particular appellate record. Professional obligations vary by jurisdiction and circumstance. This article is not legal advice.