SB 574 and AI: what California lawyers need to do before January 1
SB 574 starts January 1, 2027. It limits what you can put into AI tools. Here is what it says about client files, the other duties it adds, and a five-step list to follow.
If you use ChatGPT, Claude or any other AI for client work, one California law changes the rules on January 1, 2027. This is the short version for a solo or small-firm lawyer. It is based on the enacted text and the law firm alerts written about it. See the sources at the bottom.
What SB 574 is
- Signed by Gov. Newsom on September 30, 2026. Effective January 1, 2027.
- Adds section 6068.1 to the Business and Professions Code. The author is Sen. Tom Umberg. The enacted text is on leginfo.
- Covers every attorney practicing in California, in-house lawyers included. A firm answers for what its lawyers and staff do.
- No new fine. The law firm alerts say a breach means State Bar discipline or court sanctions under existing law.
The rule that matters for client files
This is the key sentence, from section 6068.1(a)(3)(A) of the enacted text. An attorney who uses generative AI shall:
"Not enter confidential, personal identifying, and other nonpublic information into a generative artificial intelligence system for which access to confidential, personal identifying, or other nonpublic information the attorney inputs into the system is not restricted to the attorney and persons authorized by the attorney under obligations to protect the confidentiality of the information."
In plain words: do not put that kind of information into an AI tool unless access to it is limited to you and people who are bound to keep it confidential.
What the law counts as personal identifying information
Section 6068.1(b)(2) says it includes all of these:
- Driver's license numbers
- Dates of birth
- Social security numbers
- National Crime Information and Criminal Identification and Information numbers
- Addresses and phone numbers of parties, victims, witnesses and court personnel
- Medical or psychiatric information
- Financial information
- Account numbers
- Anything sealed by court order, or made confidential by court rule or statute
Names are not on that list. But the rule also covers "confidential" and "other nonpublic" information. Removing identifiers is not the whole job.
Four more things to know
- "Restricted" is not defined. The law does not say what counts. You need a good basis to say a tool meets it.
- An enterprise plan alone does not prove it. Baker McKenzie says you must confirm who can see the data you submit.
- There is no public-records exception. Baker McKenzie says a witness address copied from a court file still counts.
- Some consumer AI tools may be out. Ropes & Gray says the rule may shut out certain publicly available consumer tools.
The other duties, and what a redaction tool does not do
The law adds more than the data rule. These are the other duties:
- Do not hand the practice of law over to AI.
- Take reasonable steps to check what the AI gives you, including every case and statute citation. Fix anything made up.
- Tell the court when you used generative AI in a court submission.
- Personally verify every citation in a filing you sign, whether or not AI was involved.
A redaction tool does not solve any of these. It cannot check an AI's answer or a citation. That is still your job.
Two ways to keep client details out
Option A: use a restricted-access tool you have vetted
This is the exception the law gives. It takes homework. You read the vendor's terms, work out who on their side can see your inputs, and keep a record of why you decided the tool qualifies. Because "restricted access" has no definition, you are making that call yourself.
Option B: take the identifying details out before you paste
If nothing identifying goes into the AI, the rule is never triggered. That follows from how the rule is worded. The law does not mention anonymizing. It does not call it a safe harbor, and it does not require it. For a solo on regular ChatGPT, it is the route with no contract homework.
One limit. The law covers "confidential" and "other nonpublic" information too, not only the list above. Case facts and strategy can stay confidential after every name is gone. A case can sometimes be traced from details like an employer, an injury date and a small town. The law also does not say how it treats placeholders you can reverse later. So you still review the case facts before you paste.
Five steps for after January 1
- List the AI tools you use. Add the ones your staff use. Include browser extensions and anything built into other software.
- Decide which ones are restricted access. Read the terms. An "enterprise" label is not an answer. If you cannot say yes and give a reason, treat the tool as not restricted.
- Take the identifying details out before anything else goes in. Names, dates of birth, account numbers, addresses and phone numbers, for clients, witnesses and everyone else. Then read what is left. Case facts can still be confidential with names gone, so cut or reword anything that points to your client.
- Check the AI's output and every citation. Look each case and statute up yourself. If the work goes to a court, follow the disclosure duty.
- Keep a note for the file. The date, which tool, what you took out first, and who checked the result. Baker McKenzie tells firms to write down which tools they use and what goes into each.
How to do step 3 with KeptPDF
Anonymize for AI looks for names, dates of birth, Social Security numbers, driver's license numbers, account numbers, addresses and phone numbers. It swaps each one for a numbered placeholder like [PERSON_1]. It runs in your browser. The file is not uploaded to KeptPDF.
It does not find everything on the law's list. It does not read medical or psychiatric details, criminal ID numbers or sealed material out of your text, and it finds financial details only in number formats it knows. You review those yourself.
- Open Anonymize for AI and drop in a PDF or Word file, or paste text.
- Review what it found. Add anything it missed. Detection is automatic, but you are the one who checks it, including the case facts.
- Copy the anonymized text into ChatGPT, Claude or whatever you use.
- Put the real names back. Paste the AI's answer into the restore box and click Restore real values. This works while the tab is open. To restore later, click Save case key first. You can lock the key file with a passphrase. If you forget the passphrase, nobody can unlock the key. Later, even in a new tab, click Load a saved key and pick the file, then restore. The key file holds the real names, so store it with the client file and never give it to an AI tool.
One honest note. The cleaned text you paste does go to the AI company. That is the point. The details you took out stay behind. For the longer walk-through, see how to redact a PDF before pasting it into ChatGPT or Claude.
Sources
The primary source is the enacted bill. The law firm alerts and State Bar pages explain it. Outside the one quote above, we summarize in our own words.
- California Legislative Information: SB 574, chaptered text (September 30, 2026)
- Holland & Knight: California enacts rules governing lawyers' use of generative AI
- CLS Blue Sky Blog (Sullivan & Cromwell): the California bill on lawyers' use of generative AI
- Baker McKenzie: California imposes new AI rules on lawyers
- Ropes & Gray: California enacts SB 574
- WilmerHale: California adopts law to impose new guardrails on attorneys and arbitrators' use of AI
- State Bar of California: Generative AI Practical Guidance (2026). It tells lawyers to understand how a tool collects, stores and shares what you give it.
- State Bar of California: proposed AI changes to the Rules of Professional Conduct. These are proposals. They are not in force.
This is not legal advice. It is a plain-English summary for general information. Talk to your own ethics counsel about your practice.