“It’s Fine,” Said the Lawyer Who Didn’t Read the Terms.
I am highly skeptical of the confidence with which lawyers are using AI. Last week I wrote about the vulnerabilities it causes vis-à-vis attorney-client privilege waivers. TL;DR: clients shouldn’t use consumer AI tools for legal work on their cases, and neither should you.
This week I want to discuss a related worry that there are a bundle of ARDC actions waiting to surface against unwary attorneys just looking to reduce costs or increase efficiency.
And if you get nothing else from this chapter, I want you to at least walk away with a robust AI policy that you can take back to your firm so that you get on the track of using AI responsibly early. If you want to model yours off mine, here’s a tool you can use for free.
I attended an estate planning conference two months ago for small firms and solo practitioners. An attorney hosted a breakout session to teach the participants how to build their own virtual executive assistant. Step 1: sign into your consumer account and download Claude desktop.
I raised my hand and asked him to address my concerns about violations of privilege or rules of professional conduct. Oh, it’s fine, he maintained. His reasoning, as I understood it: as long as the AI tool is installed on his computer, and the files never leave his firm’s server, nothing gets uploaded to a website the way it does when you paste text into a chat window using Claude on a browser. The agent just reaches into his folders, does its work, and hands back a summary, like a very fast paralegal who never leaves the building. No upload, no disclosure. No disclosure, no waiver, and no ethical problem either.
After reading Anthropic’s policies, I beg to differ.
The Terms
Anthropic’s documentation says that Claude Code, the engine behind agent sessions in the Claude Desktop app, “sends data over the network,” including “all user prompts and model outputs.”1 Its tool-use documentation explains that the results of every local action, including the contents of any file the agent reads, are sent back to the model. And on a personal account, all of it is governed by the consumer terms.
Using any consumer model of AI in this way, let alone at a conference where you are encouraging other lawyers to use it similarly, is a field of ethical landmines: confidentiality, supervision, client consent, waiving privilege.
Your AI. Your Bar License.
In 2024, the ABA issued Formal Opinion 512 on how lawyers’ use of AI must comply with the Rules of Professional Conduct, which has remained relevant despite how quickly AI has changed since it was issued.2 Here are the duties that form the basis for a grievance:
Rule 1.1 Competence – You don’t have to be an AI expert, but you have to understand what your AI can and can’t do, and keep up with how that changes. Relying on AI output without checking it can violate your duty of competence; you’re “fully responsible for the work” no matter how much of it AI did.
Rule 1.6 Confidentiality – Your data is not confidential on a consumer-tier AI product.
Rule 1.4 Communication – You must disclose your use of AI when the client asks, when it informs a significant decision, when it affects your fee, or before you put client information into a tool that could disclose it (the engagement letter is a “logical place” for this disclosure).
Rules 3.1, 3.3, 8.4(c) Candor to the Court – Check every citation and argument.
Rules 5.1 and 5.3 Supervision – Managerial lawyers “must establish clear policies regarding the law firm’s permissible use of GAI” and staff must be trained. (Click here for my free AI policy builder tool.)
Rule 1.5 Fees – Bill your actual time. If AI turns 2 hours into 20 minutes, bill 20 minutes. Don’t bill for learning how to use AI.
One Prompt. Two Ways to Lose.
ABA Opinion 512 doesn’t once mention privilege. It only addresses an ethical duty to keep client information confidential. It says nothing about the evidentiary privilege. The gap between those two concepts is where lawyers are vulnerable.
Remember Request No. 14 from last week’s blog? Opinion 512 tells you that you had a duty to read your AI’s terms and protect that information before you typed it in. Heppner tells you what happens in court when you didn’t. Put them side by side and a single prompt typed into a personal AI account can become two problems: a grievance with the ARDC and a privilege-waiver motion in your case. One AI thread, two ways to lose.
California raised the stakes when it went further than ABA Opinion 512 by codifying acceptable AI use by attorneys into law. On September 30, 2026, Governor Newsom signed SB 574, effective January 1, 2027. California lawyers may not “delegate the practice of law” to AI. They may not enter confidential or nonpublic client information into an AI tool unless access is restricted to the attorney and others obligated to keep it confidential.
The Solution
Pay for a business-tier AI product. Take the time to read their Terms and Conditions and data use policies. Before your firm puts a single client fact into an AI tool, before it gives access to any email or client file to an AI, you must be confident that:
- The tool will not train on your data
- It treats your data as yours, not sharable
- It keeps your data only for a defined period
- It’s bound by a data processing agreement you can enforce.
Beware of simply paying something and assuming your data is protected. The $20 version runs on the same consumer terms as the free version. Here is how the major tools compare as of today.
| Product | ✅ The version that meets the bar | ❌ The version that doesn’t |
|---|---|---|
| Microsoft Copilot | Microsoft Copilot (formerly Microsoft 365 Copilot) and Copilot Chat, used under your firm’s Microsoft 365 work accounts | Copilot signed in with a personal Microsoft account |
| Claude | Claude Team or Enterprise | Claude Free, Pro or Max |
| ChatGPT | ChatGPT Business or Enterprise | ChatGPT Free, Plus or Pro |
| Gemini | Gemini with a Google Workspace license | Gemini on a personal Gmail account |
| Legal research AI (CoCounsel, Lexis+ AI, Harvey, etc.) | Usually sold under enterprise contracts | Vet each one: ask for the DPA and data-use terms before signing |
These business tiers offer the contractual baseline, not a guarantee. You still have to sign the Data Processing Agreement, set retention, and train staff.
Illinois hasn’t gone as far as California in terms of legislation, but it does have a Supreme Court Policy on Artificial Intelligence that went into effect January 1, 2025,3 in addition to the ARDC’s Implementing AI Guide.4
I encourage you to read them, but if nothing else, I want you to understand the 3 crucial takeaways that I strongly feel every practicing attorney must be aware of.
- Do not allow the use of consumer AI accounts for client work, free or paid. Pick one approved tool and require everyone to use it.
- Promulgate a written AI policy based on ABA, ARDC and ISBA guidance and make sure all staff are trained on the policy.
- Include consent to use AI tools in your engagement letter.
My Freebie for you: Build Your Own AI Policy Kit
I have included a link here for you to create your own AI policy. I hope you take advantage of this free resource. All I ask is that you enter your email so you can continue to receive valuable information like this from me.
