> ## Content Index
> Fetch the complete content index at: https://www.plaintiffedge.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Confidentiality First: HIPAA, Privacy, and What Can Go Into an AI Tool
- URL: https://www.plaintiffedge.com/ai-confidentiality-hipaa-plaintiff-lawyers/
- Published: 2026-10-17T04:29:52.000Z
- Updated: 2026-10-17T04:29:52.000Z
- Description: HIPAA probably isn't what binds your plaintiff firm, but plenty else does. Here's what applies, what vendors do with your data, and what can go where.
- Author: Travis Siegel
- Tags: Start Here, Confidentiality, HIPAA, Ethics

*Current as of October 2026\. Vendor terms and ethics guidance change, so check the links below before you rely on anything here.*

*New to AI tools? Start with the first post in this series,* [*Choosing Your First AI Tool*](https://www.plaintiffedge.com/choosing-an-ai-tool-plaintiff-lawyers/)*.*

The first real question a lawyer has to answer about AI isn't which tool is smartest. It's what happens to client information once you type it in. In a plaintiff practice that handles injury, neglect, and elder abuse cases, almost everything in the file is sensitive: medical records, diagnoses, family conflicts, settlement numbers.

The good news is that you don't need to be a privacy lawyer to use AI safely. You need to know which rules apply, what the vendor does with your data, and which account your client information is allowed into. This post covers all three.

Here's how I handle it in my own work. I keep two accounts. My firm's business account has a signed business associate agreement with the vendor, and that's the only place client records go. My personal account, which runs on consumer terms, is for everything else: drafting this blog, testing ideas on public documents, and learning new features. Keeping them separate takes no effort once it's a habit, and it keeps the hardest question from ever coming up.

## The rules that actually apply to you

There's no special AI exception to the duties you already have. The same rules that govern what you say in an elevator govern what you paste into a chat window.

- **Confidentiality.** In California, that's [Business and Professions Code section 6068(e)](https://leginfo.legislature.ca.gov/faces/codes%5FdisplaySection.xhtml?lawCode=BPC§ionNum=6068&ref=plaintiffedge.com) and [Rule 1.6](https://www.calbar.ca.gov/Portals/0/documents/rules/Rule%5F1.6-Exec%5FSummary-Redline.pdf?ref=plaintiffedge.com). Typing client information into an AI tool is a disclosure to the vendor, so the question is whether that disclosure is protected well enough.
- **Competence.** [Rule 1.1](https://www.calbar.ca.gov/Portals/0/documents/rules/Rule%5F1.1-Exec%5FSummary-Redline.pdf?ref=plaintiffedge.com) requires you to understand the tools you use, including their risks. You can't hand that off to the vendor's marketing page.
- **Supervision.** If your staff use AI, Rules 5.1 and 5.3 make you responsible for having a policy and training them on it.

Three newer sources apply those duties to AI directly.

**The State Bar of California's 2026 Practical Guidance.** The State Bar replaced its 2023 guidance this year, partly to address AI tools that act on their own ("agentic" AI). On confidentiality, it says you generally shouldn't put confidential client information into a tool that poses material confidentiality or security risks without the client's informed consent. It also says you have to make reasonable efforts to learn how the tool collects, uses, stores, and shares what you enter, and that generalized marketing assurances aren't enough. [Read the guidance here.](https://www.calbar.ca.gov/sites/default/files/portals/0/documents/ethics/Generative-AI-Practical-Guidance.pdf?ref=plaintiffedge.com)

**ABA Formal Opinion 512 (July 2024).** The ABA's national opinion reaches a similar place. If a tool can learn from what you enter, you need the client's informed consent before entering information about the representation, and a boilerplate line in your engagement letter doesn't count. [Read Opinion 512 here.](https://www.americanbar.org/content/dam/aba/administrative/professional%5Fresponsibility/ethics-opinions/aba-formal-opinion-512.pdf?ref=plaintiffedge.com)

**California SB 574.** This is the big one for California lawyers. Governor Newsom signed it on September 30, 2026 and it will take effect on January 1, 2027\. It makes it a statutory duty not to enter confidential, personal identifying, or other nonpublic information into a public generative AI system. Tools restricted to you and the people you authorize, who are themselves bound to keep the information confidential, are treated differently. The bill also requires the attorney responsible for a filing to personally read and verify every citation in it, whether or not AI was involved. [Read the bill here.](https://calmatters.digitaldemocracy.org/bills/ca%5F202520260sb574?ref=plaintiffedge.com)

If you practice outside California, check your own state bar. Most have issued AI guidance by now, and the confidentiality principles are largely the same.

## Does HIPAA apply to a plaintiff firm?

This is the question I hear most, and the honest answer is "usually not directly, but that doesn't get you off the hook."

HIPAA, the federal health privacy law, applies to **covered entities** (health plans, health care clearinghouses, and providers who bill electronically) and to their **business associates**, meaning the outside companies that handle patient information on their behalf. A law firm becomes a business associate when it provides legal services *to* a health care provider or plan and gets access to patient information in the process. The federal health department's own list of examples [includes exactly that](https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/business-associates/index.html?ref=plaintiffedge.com).

A plaintiff firm is usually on the other side. You get your client's records because your client signed an authorization, or because the records were produced in discovery. In that situation, you're generally not a covered entity or a business associate, so HIPAA's rules don't bind you the way they bind the hospital.

That isn't the end of it, for three reasons:

1. **Your own confidentiality duty still applies.** Your client's medical history is about as confidential as client information gets.
2. **California's medical privacy law reaches further.** Under the Confidentiality of Medical Information Act, someone who receives medical information under a patient's authorization generally can't pass it on except as a new authorization or the law allows ([Civil Code section 56.13](https://california.public.law/codes/civil%5Fcode%5Fsection%5F56.13?ref=plaintiffedge.com)). Whether uploading records to an AI vendor counts as passing them on hasn't been settled, so treat it as if it might.
3. **Protective orders bind you.** Records produced under a HIPAA qualified protective order can be used only for that case, and must be returned or destroyed when it ends. Read your order before uploading anything produced under it, and ask whether the vendor's retention terms let you comply.

This is why a business associate agreement (BAA) with your AI vendor matters even if HIPAA doesn't technically require one for your firm. A BAA is a contract in which the vendor agrees to HIPAA-level limits on how it uses, stores, and discloses health information. It's the clearest way to show that the medical records you upload are protected.

## What happens to what you type

When you enter something into an AI tool, four things decide how safe it is. These are the questions the State Bar expects you to be able to answer.

- **Training.** Can the vendor use your conversations to improve its models? On individual plans, the answer is usually yes until you turn it off. On business plans, it's usually off by default.
- **Retention.** How long does the vendor keep your conversations, including after you delete them?
- **Human review.** Can the vendor's employees or contractors read your conversations, for example to check for misuse or to rate answers?
- **A BAA.** Will the vendor sign a business associate agreement covering health information?

Here's what the three biggest vendors say today, in their own published terms.

- **Anthropic (Claude).** On individual plans, if you allow training, de-identified conversations can be kept for model training for up to five years. Deleted chats leave your history right away and are purged within 30 days. Chats flagged for possible policy violations can be kept up to two years. Business plans (Team and Enterprise) don't train on your content by default, and a HIPAA-ready offering is available only on Enterprise (more on that below). [Retention terms](https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data?ref=plaintiffedge.com) · [Plan comparison](https://claude.com/pricing?ref=plaintiffedge.com)
- **OpenAI (ChatGPT).** Individual plans train on your content unless you opt out. OpenAI's HIPAA guide says a BAA can cover ChatGPT Enterprise, ChatGPT Edu, and the API, but not the consumer plans or ChatGPT Business. [HIPAA guide](https://cdn.openai.com/osa/hipaa-guide.pdf?ref=plaintiffedge.com) · [Plan comparison](https://chatgpt.com/pricing?ref=plaintiffedge.com)
- **Google (Gemini).** For personal Google accounts, when the "Keep Activity" setting is on, conversations can be used to train Google's models and a sample is read by human reviewers. Reviewed chats are kept for up to three years. With Keep Activity off, chats still stay with your account for 72 hours. Work accounts under Google Workspace have separate terms. [Gemini privacy hub](https://support.google.com/gemini/answer/13594961?ref=plaintiffedge.com)

### A closer look at Claude, since it's what I use

Within Claude, the Enterprise plan is the only one with adequate protection for medical records, and only after your firm has a signed BAA with Anthropic. Pro, Max, and Team don't qualify. Even on Enterprise, the BAA isn't on by default: the account's owner has to turn on the HIPAA configuration and accept the BAA.

The BAA also doesn't cover every feature. Anthropic's [list of covered products](https://privacy.claude.com/en/articles/8114513-can-i-use-anthropic-s-products-in-connection-with-protected-health-information?ref=plaintiffedge.com) includes regular chat, Projects inside chat, and web search in chat. It excludes connectors to outside apps, Claude in Chrome, and work that runs in Anthropic's cloud rather than on your computer.

The desktop agent features (originally sold as Cowork) are where I'd be most careful. When my firm set up Enterprise, those features weren't covered, so we kept medical records in Chat and Projects. Coverage could be extended to them, but we had to go through Anthropic's sales team to do it. Anthropic's list now spells out the extra step: the desktop agent features and Claude Code are covered only in local mode, meaning running on your own computer, and only after the account owner applies the HIPAA configuration to each of those products separately. Turning on HIPAA for chat doesn't turn it on for them. Either way, confirm with your Anthropic contact exactly which features your BAA covers before any records go into them.

The pattern across all three vendors is the same: individual accounts are built for individuals, and business accounts are built for confidential work. Turning off training on a personal account helps, but it doesn't give you a BAA, admin controls, or contractual confidentiality commitments.

## A simple rule for what can go where

You don't need to make a judgment call every time you open a chat window. Sort your material into three buckets once, and the rest follows.

- **Public information goes anywhere.** Published opinions, statutes and regulations, government inspection reports, Care Compare data, a facility's public cost reports, and your own blank templates. This is the right material for learning a new tool.
- **Firm-confidential information goes only into a business account.** Case strategy, settlement figures, client names, draft pleadings, deposition summaries, and anything else covered by your duty of confidentiality. Use a business or enterprise plan with training off by default and terms you've actually read.
- **Medical records go only into an account with a BAA.** Charts, nursing notes, MDS assessments, imaging reports, pharmacy records, and anything else that is health information about a real person. For most vendors, that means the enterprise tier.

Two cautions about this framework. First, removing the client's name isn't enough to move something into the public bucket. A medical chart with the name blacked out still describes a real person, and the dates, diagnoses, and facility can identify them. Second, a document that's public in one sense can still be covered by a protective order in your case. Check before you upload produced documents, even the boring ones.

## Before client information goes in

If you're setting up AI for your firm, here's what to check, in order.

![](https://storage.ghost.io/c/53/70/53703dd2-2dc9-4c4c-889e-c2cd1370b96d/content/images/2026/10/image-2.png)

The training setting in a personal Claude account.

1. **Check the training setting on any individual account, today.** Individual plans have a training switch, and it's usually on until you turn it off. Even on a personal account you only use for public material, turn it off if you don't want your conversations kept for that purpose. Business and enterprise plans generally don't have the switch at all, because training is off by default; confirm that in the plan's terms instead. In Claude it's under Settings → Privacy; in ChatGPT, Settings → Data controls; in Gemini, the Keep Activity setting.
2. **Read the business plan's terms, not the sales page.** Look for the training default, how long data is kept, who at the vendor can see it, and whether the vendor's subcontractors can.
3. **Ask for a BAA in writing before any medical records go in.** If the vendor won't sign one for the plan you're on, records stay out.
4. **Write a one-page firm policy.** Say which account is for what, who may use it, and what never goes in. The State Bar guidance expects supervising lawyers to have one, and it gives your staff a clear answer instead of a judgment call.
5. **Tell your clients.** Add a plain-language paragraph to your engagement agreement explaining that you use AI tools, how, and how you protect their information. Under ABA Opinion 512, a generic line isn't enough when a tool can learn from what you enter, so be specific, and talk it through with any client who asks.
6. **Check your protective orders.** Before uploading anything produced in discovery, make sure the order allows it and that you can return or destroy every copy, including the vendor's, when the case ends.
7. **Revisit it every few months.** Vendors change plans, terms, and features constantly. Put a reminder on your calendar to recheck the settings and terms quarterly.

## Watch out for

- **Connectors and agents.** Tools that connect to your email, cloud storage, or desktop can read far more than the file you meant to share. Check what a connector can access before you turn it on.
- **Staff using personal accounts.** The most common leak isn't a vendor breach. It's a well-meaning paralegal pasting a chart into a free account to save time. Your policy should say this out loud.
- **Shared links.** Many AI tools let you share a conversation by link. Anyone with the link can usually see the whole thing.
- **"De-identified" records.** Dates, diagnoses, and facility names can identify a person even with the name removed.

## The takeaway

HIPAA probably isn't what binds your plaintiff firm, but your confidentiality duty, California's medical privacy law, your protective orders, and now SB 574 all do. Keep public material in any account, confidential material in a business account, and medical records only where the vendor has signed a BAA. Once that sorting is a habit, you can stop worrying about it and get to work.

## Resources

- [State Bar of California, Practical Guidance for the Use of Generative AI in the Practice of Law (2026)](https://www.calbar.ca.gov/sites/default/files/portals/0/documents/ethics/Generative-AI-Practical-Guidance.pdf?ref=plaintiffedge.com)
- [ABA Formal Opinion 512: Generative Artificial Intelligence Tools](https://www.americanbar.org/content/dam/aba/administrative/professional%5Fresponsibility/ethics-opinions/aba-formal-opinion-512.pdf?ref=plaintiffedge.com)
- [California SB 574 (2026)](https://calmatters.digitaldemocracy.org/bills/ca%5F202520260sb574?ref=plaintiffedge.com)
- [HHS: Business associates under HIPAA](https://www.hhs.gov/hipaa/for-professionals/privacy/guidance/business-associates/index.html?ref=plaintiffedge.com)
- [Anthropic: How long do you store my data?](https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data?ref=plaintiffedge.com)
- [Anthropic: Using Anthropic's products with protected health information](https://privacy.claude.com/en/articles/8114513-can-i-use-anthropic-s-products-in-connection-with-protected-health-information?ref=plaintiffedge.com)
- [OpenAI: HIPAA implementation guide](https://cdn.openai.com/osa/hipaa-guide.pdf?ref=plaintiffedge.com)
- [Google: Gemini Apps privacy hub](https://support.google.com/gemini/answer/13594961?ref=plaintiffedge.com)

**Related posts:** [Choosing Your First AI Tool: A Plain-English Guide for Plaintiff's Lawyers](https://www.plaintiffedge.com/choosing-an-ai-tool-plaintiff-lawyers/)

**Next in the Start Here series:** Telling your AI who you are: standing instructions that make every answer better.

If this saved you some time, subscribe and I'll send a new post each week.

---

*This post is for informational purposes only and is not legal advice. Reading it, or contacting me about it, does not create an attorney-client relationship or any other privileged or confidential relationship. AI tools, laws, and ethics rules change quickly and differ by jurisdiction. Verify everything here independently and make sure your use of AI meets the professional and ethical standards that apply to you.*

**About the author.** Travis Siegel is a plaintiff's attorney in California who represents victims of elder abuse and neglect and their families in cases against nursing homes and assisted living facilities. He uses AI daily in his practice and writes Plaintiff Edge to share what works. More about his practice at [neglectattorney.com](https://neglectattorney.com/?ref=plaintiffedge.com).