Client confidentiality concerns with AI tools on Mac: what lawyers can do in 2026

How much lawyers use AI now, and why privacy comes up

AI use is climbing. The ABA 2024 Legal Technology Survey found that 30% of lawyers in private practice used AI, up from 11% in 2023. Solo practitioners reported 18%. Firms with 10 to 49 attorneys reported 30%, and firms with 100 or more reported 46%. The gap between firm sizes is wide, and it matters for privacy too.

Speed is the pull. A 2026 study of upper-level law students found that AI significantly boosted productivity in five of six legal tasks, with gains of 50 to 130%. Those students were not practicing lawyers. Read the gains as a sign of how attractive these tools are, not as a forecast for your practice.

Privacy concerns by firm size in the ABA survey

Privacy is not the top worry in the survey. Accuracy came first, named by 75% of lawyers. Reliability was next at 56%. Data privacy and security came third at 47%, up from 45% in 2023. That overall number hides a sharp split by firm size.

Solo practitioners named data privacy at 24%. Firms of 10 to 49 attorneys named it at 56%, and firms of 100 or more at 54%. The survey does not explain the gap. One reading is that a solo lawyer working alone has fewer people pasting into fewer accounts. The survey does not test that idea. Larger firms have more accounts, more files, and more people who can make the mistake.

What ABA Formal Opinion 512 asks you to do

ABA Formal Opinion 512 was issued on July 29, 2024. It says you must evaluate the risks before you put client information into a generative AI tool. It also warns that self-learning tools may inadvertently disclose confidential information. This summary comes from the Maine State Bar Association, which reports the opinion's wording.

The opinion says clients must be informed in advance before you input information about the representation into such tools. It also says informed consent is needed. Sources differ on how far consent goes, so read your state's ethics rules and your engagement terms before you decide how to handle it.

Opinion 512 also covers outside help. If a vendor or contractor uses a generative AI tool for your work, you must make reasonable efforts to confirm they follow your professional duties. That includes checking the vendor's security policy and confidentiality terms. A tool with no written confidentiality terms is hard to defend under that standard.

Check which account you are signed in to before you paste

Account type matters more than the app name. The same chatbot can behave differently under a work sign-in than under a personal login. Look at the account menu before you type. A personal Gmail or iCloud address usually means a personal account, even when you use it for firm work.

Cyberhaven reports that 32.3% of ChatGPT usage happens through personal accounts. That figure covers workers in general, not law firms. It still shows how common the pattern is. Personal accounts are where training defaults and retention rules are most likely to differ from a firm contract.

Read the vendor's own page, not a summary of it. Microsoft's Mac help page does not say enterprise protection applies inside the Mac app. Ask IT before you assume the Mac app is covered. Vendor terms change, so check the date on any page you rely on.

Where a Mac leaks more than you expect

A screenshot copies everything on screen, including sidebars, previews, and notifications. If a client name sits in a corner of your screen, it ends up in the file. Crop before you share. On a Mac, ⇧ ⌘ 4 captures a selected area instead of the whole screen.

Email threads carry the whole chain. A reply in Apple Mail can quote earlier messages, signatures, and addresses. Trim the thread to the message that matters before you copy anything. Some clipboard managers keep a history of what you copied, so check yours if you use one.

Check your Apple Intelligence privacy settings on Mac

Start with Apple's own record. Open System Settings > Privacy & Security > Apple Intelligence Report. The default view covers the last 15 minutes. Change the duration to the last 7 days to see more. Export Activity saves the list as Apple_Intelligence_Report.json.

Then check your browser. In Chrome, open Chrome > Settings > Privacy and security. Review site permissions for any AI site you use. Remove extensions you do not recognize, because many extensions can read the pages you open. Do this on the Mac you use for client work, not only the one you use at home.

In Apple Mail, open Mail > Settings > Accounts and confirm which account a message belongs to before you copy text from it. Mixing personal and firm accounts is how client facts end up in the wrong chat.

Redacting a client email before you ask for help

Say a client sends an email about a lease dispute. It names the client, the landlord, the property address, and a deadline. You want help listing your options. Start with a new plain text file, not the full thread. Copy only the facts the question needs.

Swap names for roles. Client stays Client, and the landlord's company becomes Landlord. Replace the street address with the rental unit. If timing matters but the exact day does not, write Date 1 and Date 2. Keep the legal question intact, because that is what the tool needs.

Then ask in general terms. A good prompt reads like this: a tenant received a notice to cure a late rent payment, what should counsel check first? Check the answer against your state's law and your own file. AI output is a draft, not a source.

Mistakes that put client data at risk with AI

1. Pasting a whole PDF when one paragraph answers the question. Every extra page is extra risk, and it is rarely needed.
2. Assuming a paid plan removes every risk. Read what the plan says about training, retention, and logs before you use it on a matter.
3. Testing a tool with a personal login on a live client file. A test should use a made-up fact pattern instead.
4. Trusting a citation without pulling the case yourself. A confident citation in a chat window can still be wrong.
5. Believing a summary is anonymous. A short fact pattern can point to a specific person, especially in a small town or niche practice.
6. Keeping one long chat open across several matters. Start a fresh chat for each matter so facts do not carry over.
7. Skipping the client conversation about AI use. Opinion 512 describes advance disclosure, so decide how you will handle it before the first upload.

What court decisions show about AI errors

Damien Charlotin's AI Hallucination Cases database listed 2,266 cases worldwide as of 5 November 2026. Of those, 1,535 came from US courts. The database tracks decisions that involve AI-generated errors. It is a monitoring list, not an official count.

Lawyer and pro se party counts in that database overlap, so do not add them together. For you, the lesson is about checking. A tool that writes a confident citation can still be wrong. Pull each case from a database you trust before you cite it.

Where Lazur fits, and where it does not

Lazur is a voice assistant for the Mac. Hold ⌥ ⌘ to ask about what is on screen, and Lazur answers out loud. Lazur listens and looks only while you hold the shortcut. Audio and the screenshot go to AI providers to answer, and they are not stored. Lazur's stated policy is that content is never used to train AI models.

That is still a transfer. If the screen shows privileged material, a client's name, or a draft you cannot share, a screenshot goes to a provider when you hold the shortcut. Check your firm's policy first. If your firm keeps client files off AI tools, do not hold the shortcut over them.

Lazur does not click buttons, edit files, or operate other apps for you. It points, explains, and guides you one click at a time. Dictation works differently. Hold fn to dictate clean text at the cursor in any app. Dictate internal notes with no client identifiers, and check how your firm treats voice data.

Lazur keeps a memory of what you asked and did, on your Mac. That helps when you want to find a past answer. It also means the Mac holds a record. Treat that record like any other file that may hold client context.

Limits of the evidence on AI and client data

No controlled study we found shows that a written policy, training, or an enterprise tier reduces client data leaks. The guidance here rests on the ABA opinion, vendor help pages, and survey data. Surveys measure what people report, not what they do with a tool on a busy afternoon.

Vendor terms and tool behavior change. Legal research tests from 2024 covered the tools of that period. Treat each number on this page as a snapshot from November 2026, and check its source before a decision depends on it.

Where to go next for your practice

If you draft in Word, read Microsoft Word for lawyers on Mac. For Outlook, see Outlook for lawyers on Mac. To see how Lazur handles Word, read AI assistant for Word on Mac. Your firm's rules on client data come first, in every app.