Can you use AI for legal work without breaking client confidentiality?

The two ethics rules that actually apply

Two ABA Model Rules govern almost every question about AI and client information, and most states have adopted close analogs.

Model Rule 1.6 (confidentiality) requires a lawyer to not reveal "information relating to the representation" and to make "reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to," that information. That standard is broader than attorney-client privilege — it covers essentially anything you learned in the course of representing the client, not just privileged communications.

Model Rule 1.1 (competence), through Comment 8, requires lawyers to keep abreast of "the benefits and risks associated with relevant technology." In practice, that means you're expected to understand how an AI tool handles your inputs before you use it on client matters.

Several state bars and the ABA's Formal Opinion 512 (2024) have applied these rules directly to generative AI. The through-line: using AI is permitted, but you remain responsible for confidentiality, for verifying outputs, and for understanding the tool.

What goes wrong when you skip the safeguards

There are two distinct failure modes, and they're often confused.

The confidentiality failure is on the input side: you paste a client's name, a settlement figure, medical details, or an entire pleading into a consumer AI tool whose terms allow it to retain and train on what you submit. The sensitive information is now outside your control.

The accuracy failure is on the output side: the AI invents a case that doesn't exist, or cites a real case for a proposition it never held, and the fabricated authority ends up in a filing. This one has become impossible to ignore.

The two failures compound: a tool that isn't safe to feed client data into is often the same tool that hallucinates citations, because consumer chatbots are optimized for fluency, not for grounding every claim in verifiable law.

Does AI use waive attorney-client privilege?

Privilege can be waived when privileged material is voluntarily disclosed to a third party. Whether an AI vendor is that kind of "third party" depends entirely on the arrangement:

• Consumer tools that may retain inputs and use them to train models are the highest-risk category — the disclosure is real and the data may persist.
• Enterprise tools operating under an agreement where your data is not used for training, is encrypted, and stays within a controlled environment present a much stronger position — closer to using any other vendor bound by confidentiality (an e-discovery host, a cloud provider) under Rule 5.3.
• De-identified content sidesteps the question: if the text sent to the model contains no client-identifying information, there is no confidential client information in the disclosure to begin with.

Because waiver law is still developing and varies by jurisdiction, the conservative practice is to assume any identifiable client detail sent to an outside model could be scrutinized later — and to avoid sending it.

How to use AI on case files safely

Four practices, in order of importance:

1. De-identify before you send. Remove or tokenize names, addresses, SSNs, dates of birth, account and case numbers before the content reaches any AI. This is the single highest-leverage safeguard because it protects confidentiality regardless of the vendor's terms.
2. Use tools that don't train on your data. Choose enterprise AI that contractually excludes your inputs from model training and runs in a private, encrypted environment — ideally one that is SOC 2 audited and, where health data is involved, HIPAA-eligible.
3. Validate every citation. Never file an AI-drafted citation you haven't confirmed against the primary source. Prefer tools that check citations automatically and remove unverified ones rather than leaving them for you to catch.
4. Keep a human attorney accountable. AI assists; it does not sign the filing. Competence and candor to the court remain yours.

Where document anonymization fits

Anonymization is what makes step one practical instead of aspirational. Manually scrubbing a 40-page deposition is error-prone; missing a single name defeats the purpose. Automated anonymization detects the identifiers and replaces each with a reversible token — for example, every occurrence of a client's name becomes [[PERSON_001]]. The AI analyzes the de-identified text, and you rehydrate the real values in the result afterward, so you keep a usable work product without the model ever seeing who the client is.

This is the approach ClientShield is built around: a Presidio-based pipeline anonymizes documents before analysis, the AI runs on Claude via Amazon Bedrock in a private VPC where inputs are never used for training, and case law research validates every citation against primary sources — addressing the confidentiality failure and the accuracy failure in one workflow. If your concern is narrowly about catching bad citations in briefs you've already drafted, CiteVerify handles that piece on its own.

Frequently asked questions

Is it ethical for a lawyer to use ChatGPT on client documents?

It can be, but pasting identifiable client information into a public consumer tool generally is not, because it risks unauthorized disclosure under Rule 1.6. Use an enterprise tool with a no-training agreement in a private environment, or de-identify the information first.

Does using AI waive attorney-client privilege?

Disclosing privileged material to a third party can waive privilege. An enterprise arrangement where data isn't used for training and stays in your controlled environment is far lower risk than a consumer tool. Anonymizing the content before sending removes the identifying information entirely.

What's the safest way to use AI on case files?

De-identify client PII first, use tools that never train on your data, validate every citation against primary sources, and keep a human attorney responsible for the final product.

This guide is general information about legal-ethics considerations and technology, not legal advice, and it doesn't create an attorney-client relationship. Ethics rules vary by jurisdiction and change over time — consult your state bar's rules and opinions, and your own professional-responsibility counsel, for guidance on your situation.