Where AI Helps in a California Practice, and Where It Will Burn You

Last reviewed: July 28, 2026 Checked against the current California code on this date.

The beginner’s guide in this section covers getting started. This one is for lawyers who are already using these tools and want a straight accounting: what actually pays off, what quietly costs you, and how to check your work.

The honest ledger

What genuinely pays off

Compression. Turning three hundred pages into something you can hold in your head. Depo summaries, document review triage, chronologies built from records you already have.
First drafts of structured, repetitive writing. Discovery responses, routine motions, meet and confer letters. The form is known. You are supplying judgment, not prose.
Adversarial pressure testing. “Here is my argument. Make the strongest version of the other side.” The most underrated use, and it costs you nothing.
Translating register. Turning your analysis into something a client can actually read, or turning a client’s rambling email into a clean set of facts.

What quietly costs you

Anything that requires knowing rather than writing. Research you intend to rely on. Local rules. Whether a case is still good law. It will answer confidently and it will sometimes be wrong.
Anything where the failure is invisible. A summary that omits the one bad fact reads exactly like a summary that does not.
Volume without judgment. It will happily generate forty pages. Forty pages is not better. Judges read hundreds of filings.

The under-discussed risk is not fabrication, it is omission. A hallucinated case gets caught the moment somebody pulls it. A summary that silently drops the harmful paragraph gets caught at deposition, by opposing counsel, in front of your client.

Checking your citations

Every sanctions story follows the same shape. A brief cites cases. Nobody pulls them. The cases do not exist.

The prevention is not a better tool or a cleverer prompt. It is a rule you apply to yourself, your associates, and your staff:

No citation goes in a filing unless a human opened it and read the relevant passage.

Being connected to a real database helps, but it does not solve this. It reduces outright fabrication. It does not stop mischaracterization, and a real case cited for a proposition it does not support will get you in just as much trouble under Rule 3.3.

If you have a research connector

If you are running Claude with the CoCounsel connector, or an equivalent setup that reaches a real legal database, there is a workflow worth building into your habits.

At the end of a drafting session, before you finalize anything, tell the assistant to collect every piece of legal research it relied on and write you a prompt for a different chat. Then run that prompt in a fresh session, pointed at the connector, and have it put every citation through verification and hand you back an analysis of what it found.

The reason for the second chat is the whole point. A fresh session has none of the first one’s assumptions. It is not invested in the draft being right. You are getting an independent check rather than asking the same context to grade its own work, which it will do generously.

If you do not have one

Then you check the cites the way you always did. Pull every one. Read the passage. Run them through whatever cite-checking process you already trust.

A real example, and it is the reason this section exists. A research assistant connected to a paid legal database was asked to settle which bill added a particular Penal Code section. It answered confidently and cited a Penal Code section that does not exist. Everything else it produced in that session, statutes and case law alike, checked out.

Sit with that for a second. It fabricated a citation while answering a question about which citations are real. A database connection lowers the rate of invention. It does not eliminate it, and it cannot tell you which of the answers in front of you is the invented one. The only thing that caught it was somebody pulling the section.

Not checking your cites is lazy. If you do not want to check cites, you should not be a lawyer. There is no version of this technology that makes verification optional, and the lawyers who have learned that the hard way learned it in a published order with their name on it.

Read more: building a verification workflow that survives a busy week

The rule is easy to state and easy to skip at 11pm before a filing deadline. What makes it hold:

Keep a citation table separate from the draft. Every authority, with a column for who pulled it and when. If the column is empty, the brief is not done.
Require a pincite for every proposition. This is the single highest-leverage habit, because you cannot produce an honest pincite without opening the case. It converts a vague obligation into a specific one.
Have someone other than the drafter do the pull. The person who wrote the sentence already believes it. Fresh eyes catch the case that is real but does not say what the sentence claims.
Give AI-assisted drafts MORE citation scrutiny, not less. This is counterintuitive and it matters. The prose comes back clean and confident, which makes it read as more authoritative than it has earned. Polish is not accuracy.
Check the department, not just the jurisdiction. Some judges have standing orders about AI use in filings. That is a five minute check.

If you find the problem after filing: Rule 3.3 imposes a duty to correct a false statement of material fact or law previously made to the tribunal. The candor obligation points strongly toward telling the court promptly and on your own initiative. That conversation is unpleasant. It is considerably less unpleasant than the version where opposing counsel raises it first.

Six rule changes California lawyers should read now

California is moving toward putting AI obligations into the Rules of Professional Conduct. Note carefully what is actually happening: the California Supreme Court directed the State Bar to consider incorporating the guidance into the comments to the rules, not into the black letter rules themselves. That distinction matters when you are reading anyone else’s summary.

The sequence, current as of this writing: the State Bar issued its Practical Guidance on November 16, 2023. On August 22, 2025 the California Supreme Court directed the Bar to consider incorporating that guidance into the comments to the rules. COPRAC approved a package on March 13, 2026 and a first public comment period closed May 4, 2026. On May 14, 2026 the Board of Trustees approved a revised Practical Guidance and depublished the 2023 version. The rule proposals then went out for a second public comment period on June 12, 2026, closing August 6, 2026, with rules 1.1, 1.4, and 1.6 substantively revised in the interim.

These are proposals. They are not rules yet. They remain subject to modification and ultimately to the California Supreme Court, which has final authority over the Rules of Professional Conduct. The list below describes the package as approved in March 2026; rules 1.1, 1.4, and 1.6 were revised again before the second comment period, so read the current text before relying on the detail.

What is in the package:

Rule 1.1, Competence. Two new comments on technology and AI. The operative requirement is that a lawyer must independently review, verify, and exercise professional judgment over AI output used in a representation.
Rule 1.4, Communication. A new comment requiring disclosure to the client where AI use presents a significant risk or materially affects the scope, cost, manner, or decision making of the representation. Note what this is not: it is not a blanket duty to tell every client you used AI. It turns on significance and on client sophistication.
Rule 1.6, Confidentiality. A new comment expanding what counts as revealing information to include exposing client data to AI systems where there is a material risk of unauthorized access or retention. This one has the most practical bite, because it reaches your vendor choices and your retention settings.
Rule 3.3, Candor. A new comment requiring verification that cited authorities are not fabricated, misstated, or taken out of context. Note “misstated” and “out of context,” not only “fabricated.”
Rule 5.1, Supervisory lawyers. Managerial lawyers must establish procedures governing AI use. In practice: your firm needs a written policy.
Rule 5.3, Nonlawyer assistants. Supervision duties extend to AI use by paralegals and staff.

If you read that list and thought “we do not have a written AI policy,” that is your action item. Five and six are the ones most firms are furthest from, and they are also the cheapest to fix before the rules land.

Read more: what a workable small firm AI policy contains

A short policy people actually read beats a long one nobody opens. Cover:

Approved and prohibited tools. Name them. “Use good judgment” is not a policy.
What may never be entered into any tool. Client identifiers, privileged material, anything under a protective order.
The verification rule. Citations and factual assertions, who checks, and how it gets recorded.
Who reviews output before it leaves the office. Name a role, not a hope.
Staff disclosure. If a paralegal used AI in work product they hand up, you want to know that.
Vendor defaults. Your practice management, email, and document systems are adding AI features and switching some on by default. Someone has to own checking that.
A review date. This area moves fast enough that an annual look is already too slow.

Billing

Rule 1.5(a), Fees for Legal Services

“A lawyer shall not make an agreement for, charge, or collect an unconscionable or illegal fee.”

This is the area where the guidance changed most, and a lot of what is circulating is out of date:

The flat prohibition is gone. The 2023 guidance said plainly that a lawyer must not charge hourly fees for time saved by using generative AI. That sentence was deleted in the 2026 revision. Anyone still quoting it is quoting a depublished document.
You may bill for time you actually spend working with the tool. Writing and refining prompts, reviewing output, correcting it, and turning it into something filed is real work and it is billable like any other.
Cost pass-through narrowed. The 2026 revision is tighter about charging clients for the tools themselves, not looser. Do not assume a subscription passes through. Put the basis in the engagement letter and treat it as a deliberate decision.

The rule did not move, only the guidance did. Rule 1.5(a) still prohibits an unconscionable fee, and billing a client for hours you did not work is a Rule 1.5 problem regardless of what any guidance document says. What you do with the recovered time is a business decision the profession has not settled.

What I would tell a skeptical partner

Two things.

The tools are more useful than the skeptics say and much less useful than the vendors say. The gap between those two claims is where the actual work sits, and nobody selling you anything is going to point at it.

And the risk is manageable, because it is the risk you already manage. Verify your citations. Protect client confidences. Supervise your people. Bill honestly. None of those obligations changed. The only thing that changed is how fast you can now generate something unverified, which means the discipline has to be deliberate where it used to be automatic.

Everything here aimed at lawyers rather than clients is collected in the for attorneys section.

This is general information for other lawyers. It is not legal advice, not an ethics opinion, and not a substitute for your own reading of the Rules of Professional Conduct and the State Bar’s guidance. The rule amendments described above were proposals as of July 2026 and had not been adopted. Confirm the current status before relying on any of it.