Can Lawyers Use Claude? Ethics, Confidentiality, and Client Data

Short answer: Yes, lawyers can potentially use Claude ethically. Neither the ABA Model Rules nor ABA Formal Opinion 512 categorically prohibits lawyers from using generative AI. But a lawyer should not treat that answer as permission to put any client information into any Claude account.

Can lawyers enter client information into Claude? Sometimes. The lawyer must evaluate the information, matter, task, Claude plan and settings, governing terms, client consent, and any restrictions imposed by law, a court, a protective order, an engagement agreement, or the client.

Which Claude plan is better for confidential legal work? Claude Team provides a materially stronger contractual foundation than Claude Pro. Anthropic's Commercial Terms say it "may not train models on Customer Content from Services" and designate Customer Content as the customer's Confidential Information. Claude Pro users can opt out of ordinary model improvement, but the Consumer Terms and Privacy Policy retain exceptions for feedback and safety-reviewed material.

A solo or small-firm lawyer can nevertheless make a credible ethics case for using Claude Pro with model improvement disabled when the lawyer selects appropriate tasks, minimizes sensitive information, controls integrations, verifies every material output, and obtains informed client consent when required.

That is not the same as saying Claude is "ABA approved," that every use is permissible, or that using an AI service can never affect privilege. The ABA Model Rules of Professional Conduct are model rules, not the binding rules of every jurisdiction. ABA Formal Opinion 512 is guidance, not a product certification. Lawyers must also consider their jurisdiction's adopted rules and opinions, applicable law, protective orders, court-specific AI rules, engagement terms, and client instructions.

But the primary sources support a more useful conclusion than "lawyers should never use AI with client work." The ABA applies a fact-specific, reasonable-efforts standard. It recognizes both the risks of generative AI and its potential to improve the efficiency and quality of legal services.

This article explains how that framework applies to Claude Pro and Claude Team, using Anthropic's current terms and the ABA's own rules and ethics opinion.

Source check date: July 10, 2026. Anthropic's terms, product names, settings, and features can change.

The Answer in One Table

Plan details and pricing can change. Anthropic's current Plans and Pricing page lists Claude Pro as an individual productivity plan and Team as a plan for 5 to 150 users. The pricing page is useful for features and availability; the legal commitments come from the governing terms.

A Five-Question Test Before a Lawyer Uses Claude

1. Will the prompt disclose information relating to a representation? If not, ordinary uses such as summarizing a published opinion or generating generic ideas usually do not present the same Rule 1.6 consent question. If yes, continue the analysis.
2. Can the task be completed with less information? Prefer public materials, narrow excerpts, redaction, pseudonyms, or a neutral hypothetical when they will work.
3. Is the information privileged, sealed, exceptionally sensitive, or restricted? Review privilege law, protective and sealing orders, client instructions, engagement terms, and outside-counsel guidelines before disclosure.
4. What account, feature, and terms govern? Confirm whether the account is Pro or Team, whether model improvement is disabled, what retention applies, and whether Cowork, connectors, plugins, or external services will receive access.
5. What authorization and safeguards are required? Decide whether informed client consent is required, restrict permissions, preserve necessary work product, and independently verify every material output.

The practical outcomes are not limited to "safe" or "unsafe." A lawyer may proceed with ordinary verification, proceed only with minimization and documented safeguards, prefer a commercial plan and conduct further matter-specific analysis, or keep the information out of Claude unless additional authority or consent is obtained.

What Does the ABA Actually Require When Lawyers Use Generative AI?

ABA Formal Opinion 512, issued July 29, 2024, addresses lawyers' use of generative AI. It does not prohibit the technology. Instead, it organizes the analysis around duties lawyers already have:

• competence;
• confidentiality;
• communication;
• supervision;
• meritorious claims and candor to tribunals; and
• reasonable fees.

The opinion says lawyers do not need to become AI experts. They do need a "reasonable understanding of the capabilities and limitations" of the specific tool they use. ABA Formal Opinion 512, at 3. It also says GAI may be used as a "springboard or foundation for legal work," but the lawyer remains fully responsible for the work. Id. at 4.

For confidentiality, the opinion rejects a universal rule in favor of a factual inquiry. The lawyer should consider the likelihood of disclosure or unauthorized access, the information's sensitivity, the difficulty of implementing safeguards, and whether those safeguards interfere with the representation. The relevant facts include the client, matter, task, tool, and configuration.

That framework matters. The ethical question is not simply, "Did information leave the lawyer's laptop?" Lawyers routinely use email, cloud storage, e-discovery platforms, practice-management systems, experts, and litigation-support vendors. The question under Model Rule 1.6(c) is whether the lawyer made reasonable efforts to prevent inadvertent or unauthorized disclosure or access.

Formal Opinion 512 specifically tells lawyers to read and understand a generative AI tool's terms of use, privacy policy, and related contractual terms, or consult someone who has. Id. at 6. It also tells lawyers to investigate configuration, enforceable confidentiality, breach notification, reliability, security measures, retention, proprietary rights, and cyberattack risk. Id. at 11–12.

Those instructions make the differences between Claude Pro and Claude Team ethically relevant.

That is the central comparison in this article. The ABA has not approved Claude Team—or any other Claude plan. Team presents the stronger ethics case because its contract answers more of the vendor-diligence questions Formal Opinion 512 tells lawyers to investigate.

Why Using Claude Well Can Improve Client Outcomes

Confidentiality is essential, but it is not the only ethical consideration. The starting rule is Model Rule 1.1: a lawyer must provide the legal knowledge, skill, thoroughness, and preparation reasonably necessary for the representation. Comment 8 to Rule 1.1 adds that lawyers should keep abreast of changes in law and practice, including the benefits and risks associated with relevant technology.

Responsible use of Claude can help a lawyer:

• build a chronology from a large record;
• compare documents and identify inconsistencies;
• prepare deposition topics and test both sides of an argument;
• draft from attorney-selected authorities and record materials;
• check whether a draft addresses every element or claim; and
• explain complicated issues more clearly to a client.

These uses can support competence, diligence, and client communication. Formal Opinion 512 recognizes that emerging tools may produce work of "distinctively higher quality" or allow lawyers to work "markedly faster and more economically." Id. at 4. Responsible use can therefore improve representation as well as create risks.

Can a Solo Lawyer or Firm With Fewer Than Five Users Ethically Use Claude Pro?

There is a credible argument that it can. Claude Pro is an individual plan, while Team starts at five seats, leaving smaller firms to use Pro with limitations or pay for unused Team seats.

The strongest Pro-plan argument has five parts.

1. Turn off model training

Anthropic's Consumer Terms say it may use inputs and outputs to improve services and train models unless the user opts out through account settings. Anthropic's Privacy Policy likewise says it may use inputs and outputs for model improvement unless the user opts out.

This setting matters because Formal Opinion 512 gives special attention to "self-learning" tools that could later reveal information from one representation. Disabling training reduces the ordinary risk that client material will become training material, but it is not absolute: the Consumer Terms retain exceptions for feedback and safety-reviewed materials.

Turning off model improvement changes the risk analysis. It does not by itself resolve client consent, professional confidentiality, privilege, retention, human access, or disclosure through third-party integrations.

A Pro policy should prohibit feedback on client-related conversations and recognize the safety-review exception.

2. Match the information to the risk

Rule 1.6 protects all information relating to a representation, not only privileged communications. But summarizing a published opinion is different from uploading a psychiatric record, trade secret, sealed discovery, or protected material.

For Pro, a lawyer can reduce risk by:

• using public or already-disclosed information where possible;
• removing unnecessary identifiers and using consistent pseudonyms;
• extracting only the passages needed for a task instead of uploading the entire file;
• using a neutral hypothetical when the task does not require actual facts;
• keeping exceptionally sensitive or restricted information out of the service; and
• deleting conversations when they are no longer needed.

Anthropic's Privacy Policy says a deleted conversation is removed immediately from conversation history and automatically deleted from Anthropic's backend within 30 days, subject to the policy's stated limitations and exceptions. Deletion is therefore a useful lifecycle control, not a representation that no copy can ever remain anywhere.

3. Treat Claude as supervised assistance, not the lawyer

Anthropic's Consumer Terms warn that outputs may be inaccurate, and Formal Opinion 512 leaves lawyers fully responsible for client work.

For litigators, that means independently checking:

• cases, statutes, rules, quotations, and citations;
• factual statements against the record;
• whether authority remains good law and whether controlling authority is omitted;
• whether arguments have a nonfrivolous basis and requested relief fits the evidence and procedure; and
• whether the court has a local rule, standing order, or certification requirement concerning AI.

These controls implement Model Rule 3.1, Model Rule 3.3, and the prohibition on dishonesty in Model Rule 8.4(c). They are mandatory regardless of plan.

4. Obtain informed consent when required

Formal Opinion 512 says informed consent is required before putting representation information into the type of self-learning GAI tool described in the opinion—one whose output could disclose one client's information to later users. ABA Formal Opinion 512, at 5–6. When consent is required, merely adding general boilerplate to an engagement letter is "not sufficient." Id. at 6.

Disabling ordinary training improves the analysis, but it does not eliminate every disclosure, storage, access, or safety-review issue under the Consumer Terms. For identifiable or sensitive client information on Pro, the conservative approach is to obtain matter-specific informed consent or keep the information out of the service.

According to the opinion, an informed discussion should explain:

• why the tool will be used;
• what client information will be disclosed;
• the particular risks, possible adverse uses, and who may have access;
• how the tool benefits the representation; and
• what safeguards the lawyer will use.

Client consultation may also be necessary under Rule 1.4 when AI will influence a significant litigation decision, when its use affects fees, when the engagement terms or outside-counsel guidelines require disclosure, or when the client asks.

5. Control Cowork, connectors, and external actions

Confidentiality analysis cannot stop at model training. Connectors, plugins, webhooks, and external APIs can send information to third parties under their own privacy policies, and some integrations retain access until disconnected. Claude and Cowork may also read or modify files, send communications, or take other actions.

A small firm should give Claude access only to the folders, tools, and services necessary for the task. External communications, filings, destructive file operations, and material edits should require human confirmation. Matter files should be backed up, and the lawyer should review both the input scope and the resulting action.

The bottom line on Pro

Claude Pro is not contractually equivalent to Team, but Formal Opinion 512 does not require a particular plan. For firms below Team's five-seat minimum, Pro can be defensible for selected tasks with training opt-out, minimization, narrow permissions, verification, deletion practices, and client consultation where necessary. The more sensitive the information, the stronger the reason to avoid Pro or use a commercial offering with better protections.

Why Claude Team Presents a Stronger Ethics Case

Claude Team changes the analysis in several important ways.

Anthropic contractually prohibits training on Customer Content

Under Section B of Anthropic's Commercial Terms, the customer retains its rights in inputs, owns outputs to the extent permitted by law, and Anthropic "may not train models on Customer Content from Services."

This is stronger than an individual user's opt-out setting with enumerated feedback and safety-review exceptions in the Consumer Terms. Team customers should still review any separate feature-specific terms and should not affirmatively enable programs that permit other data use. For example, Anthropic's Service Specific Terms describe an optional Development Partner Mode under which a customer may permit content to be used for training.

Customer Content is contractually confidential

Section E.1 of the Commercial Terms states: "Customer Content is Customer's Confidential Information." Section E.2 limits use to exercising rights and performing obligations under the agreement, restricts disclosure to need-to-know representatives bound by protective confidentiality obligations, and requires protection with no less than reasonable care.

That commitment maps directly onto Formal Opinion 512's instruction to determine whether confidentiality obligations are enforceable.

A Data Processing Addendum governs customer data

Section C of the Commercial Terms incorporates Anthropic's Data Processing Addendum. The DPA provides, among other things, that:

• the customer is controller and Anthropic is processor for Customer Personal Data, which Anthropic processes under documented instructions subject to applicable law;
• authorized personnel and subprocessors receive confidentiality or substantially protective data-processing obligations;
• Anthropic will maintain reasonable and appropriate technical and organizational security measures;
• Anthropic will notify the customer of a Security Breach "without undue delay," and no later than 48 hours after becoming aware, DPA § G.1; and
• customer data is subject to return and deletion provisions following termination, with stated exceptions.

No contract eliminates operational, legal-process, or cyberattack risk. But these terms answer several of the vendor-diligence questions Formal Opinion 512 tells lawyers to investigate.

Team supports firm supervision

Anthropic's pricing page currently lists central billing and administration, SSO, and administrative controls for remote and local connectors as Team features. Anthropic's Service Specific Terms for Claude for Work also describe Team as an administered offering in which the customer has access to and control over data submitted by its users.

Those controls matter under Model Rule 5.1 and Model Rule 5.3. Formal Opinion 512 tells managerial lawyers to establish clear AI policies, train lawyers and nonlawyers, and make reasonable efforts to ensure compliance.

An administered workspace makes it easier to require firm-controlled accounts, enforce connector settings, provision and revoke access, separate work from personal use, train users under one policy, and respond consistently to incidents.

Team does not make every input appropriate. A protective order can still prohibit disclosure. A client can still say no. A connector can still send information to another provider. A lawyer can still submit a hallucinated case to a court. But Team provides a materially stronger contractual and governance foundation for routine firm use.

What Claude Team does not solve

Claude Team does not:

• override a protective order, sealing order, client instruction, or outside-counsel guideline;
• decide whether a disclosure waives privilege or work-product protection;
• supply informed client consent when consent is required;
• guarantee that an output is accurate or complete;
• make third-party connectors subject to Anthropic's terms; or
• replace the lawyer's jurisdiction-specific analysis and professional judgment.

Claude Pro Versus Team: Which Should a Small Law Firm Choose?

For one to four users, the practical choice may be Pro with strict limitations or paying for Team's five-seat minimum. The published Team price and seat requirements should be checked at the time of purchase.

Pro is easier to defend when the firm primarily uses Claude for:

• public legal materials;
• sanitized or pseudonymized fact patterns;
• generic issue spotting, argument testing, and rewriting without sensitive facts;
• selected excerpts rather than complete matter files; and
• matters in which the client has given appropriately informed consent.

Team is the better default when the firm expects to:

• process identifiable client material routinely;
• give several lawyers or staff members access;
• use Cowork with matter folders or connect firm systems;
• administer access centrally;
• require contractual confidentiality for customer content;
• rely on a DPA and breach-notification terms; or
• establish an approved firm-wide AI environment.

For a five-to-twenty-person firm, the supervisory and access-control case strongly favors Team over a collection of individual Pro accounts.

Ethics Compliance Does Not Automatically Preserve Privilege

Several different bodies of law and contract can apply to the same Claude workflow. A conclusion under one does not decide the others.

An ethically reasonable use is not necessarily privileged, permitted by a protective order, or compliant with a client's contract. Those questions require separate analysis.

A Litigator's Matter-and-Task Risk Matrix

This is not a substitute for jurisdiction-specific analysis, but it illustrates how Formal Opinion 512's fact-specific standard can work in practice.

The plan is only one factor. Team does not make a prohibited disclosure permissible, and Pro does not make public-information work unethical.

What a Small-Firm Claude Policy Should Require

Formal Opinion 512 expressly calls for clear policies and training. A useful policy should cover at least the following:

1. Approved accounts and settings. Identify when Pro or Team is permitted, require Pro training opt-out, prohibit client work in unapproved accounts, and recheck settings periodically.
2. Information handling. Classify public, confidential, privileged, protected, sealed, and exceptionally sensitive information; require minimization appropriate to each.
3. Client communication. Define when the responsible lawyer must consult the client and document informed consent.
4. Permissions and restrictions. Limit connectors and Cowork access, require confirmation before consequential actions, and review court orders, client directions, and engagement terms.
5. Verification. Require source checking of legal, factual, record, and numerical assertions.
6. Supervision and training. Train lawyers and nonlawyers on the tool, confidentiality, secure handling, and escalation of mistakes.
7. Retention and incidents. Govern chat deletion, preservation of necessary work product, labeling of AI-assisted drafts, and response to unintended disclosure or access.
8. Periodic reassessment. Revisit the policy when Anthropic changes its terms, plans, features, integrations, or privacy practices.

Billing for Work Performed With Claude

Model Rule 1.5 requires reasonable fees and expenses. Formal Opinion 512 is direct: hourly lawyers "must bill for their actual time," including prompting and review—not the time the task would have taken without AI. ABA Formal Opinion 512, at 12.

Any separate AI charge should be explained, preferably in writing. Flat and contingent fees remain subject to Rule 1.5, and AI efficiency is not a license to disguise fifteen minutes of work as three billable hours.

Frequently Asked Questions

Is Claude confidential for lawyers?

It depends on the plan and use. Team's Commercial Terms call Customer Content confidential and prohibit training on it. Pro permits an opt-out from ordinary model improvement, subject to feedback and safety-review exceptions. Lawyers must still evaluate access, retention, integrations, sensitivity, consent, and governing law.

Is Claude Pro safe for attorneys if training is turned off?

Turning off training materially improves the analysis but does not create Team's contractual protections or eliminate storage, disclosure, safety-review, integration, or security risks. Pro can be defensible for controlled uses with public, minimized, redacted, or pseudonymized information.

Does Claude train on client data?

For Pro, Anthropic says inputs and outputs may be used for training unless the user opts out, subject to feedback and safety-review exceptions. Under the Commercial Terms, Anthropic may not train models on Customer Content. Verify the terms governing the account and feature.

Does the ABA prohibit lawyers from putting client information into generative AI?

No categorical prohibition appears in Formal Opinion 512. The opinion requires a fact-specific risk analysis and reasonable safeguards. It says informed consent is required before entering representation information into the type of self-learning tool described in the opinion. Formal Op. 512, at 5–6. It also identifies uses, such as idea generation without representation information, for which consent is not necessary under Rule 1.6. Id. at 6.

Do lawyers need client consent before using Claude?

Not for every use. A generic question without representation information ordinarily does not require consent. Consent may be required when representation information is disclosed and the tool creates the risks discussed in Formal Opinion 512. Consultation can also be required when AI affects an important decision, fees, or client expectations.

Does using Claude waive attorney-client privilege?

Not necessarily, but ethics compliance does not decide privilege. Waiver depends on applicable law, the disclosure, governing terms, and precautions taken. Work-product, privacy, contractual, and protective-order questions are also separate.

Can a lawyer upload a deposition transcript or discovery to Claude?

Potentially. First assess whether the material is public, confidential, privileged, sealed, sensitive, or restricted, and whether excerpts, redaction, or pseudonyms will suffice. Team has stronger terms, but neither plan overrides a court order or client restriction.

Can a lawyer use Claude to draft a brief?

Yes, as a supervised drafting tool. The lawyer remains responsible and must verify the record, law, quotations, citations, controlling authority, arguments, relief, and court rules before filing.

Must a lawyer disclose the use of Claude to a court?

There is no universal disclosure rule. Follow the court's rules, standing orders, certifications, and duties of candor and honesty.

Must a lawyer tell the client that Claude was used?

Not always. Disclosure or consultation is more likely when consent is required, the client asks, engagement terms require it, AI affects fees, or the output will influence a significant decision.

Can a law firm bill a client for time saved by Claude?

An hourly lawyer may bill actual prompting and review time, not hypothetical time saved. Separate tool charges must be reasonable and communicated; flat and contingent fees remain subject to Rule 1.5.

Should a small law firm choose Claude Pro or Claude Team?

Team offers stronger no-training, confidentiality, data-processing, and administrative protections. Firms below its five-seat minimum can limit Pro to appropriate workflows or decide Team's cost is justified by the data they expect to process.

Conclusion

The ABA materials do not require responsible lawyers to avoid Claude. They require lawyers to understand the service, match safeguards to the task and information, supervise its use, verify its work, communicate when required, and remain accountable.

Pro with training disabled can support a controlled workflow, but its consumer terms justify tighter limits on sensitive information. Team is stronger for routine firm use because its commercial terms, confidentiality commitment, DPA, and administrative controls answer more of Formal Opinion 512's vendor-diligence questions. Neither plan replaces the lawyer's judgment.

Primary Sources Reviewed

This article relies on the following primary sources. Anthropic's terms and product features can change; this analysis uses the Anthropic materials available on July 10, 2026.

ABA sources

• ABA Formal Opinion 512: Generative Artificial Intelligence Tools (July 29, 2024)
• ABA Model Rule 1.1: Competence and Comment on Rule 1.1
• ABA Model Rule 1.3: Diligence
• ABA Model Rule 1.4: Communications
• ABA Model Rule 1.5: Fees
• ABA Model Rule 1.6: Confidentiality of Information and Comment on Rule 1.6
• ABA Model Rule 3.1: Meritorious Claims and Contentions
• ABA Model Rule 3.3: Candor Toward the Tribunal
• ABA Model Rule 5.1: Responsibilities of Partners, Managers, and Supervisory Lawyers
• ABA Model Rule 5.3: Responsibilities Regarding Nonlawyer Assistance
• ABA Model Rule 8.4: Misconduct

Anthropic sources

• Consumer Terms of Service, effective October 8, 2025 when reviewed
• Privacy Policy, effective July 8, 2026 when reviewed
• Commercial Terms of Service, effective June 17, 2025 when reviewed
• Data Processing Addendum, effective February 24, 2025 when reviewed
• Service Specific Terms, effective June 8, 2026 when reviewed
• Claude Plans and Pricing, reviewed July 10, 2026

---

About this analysis: Tommy Eberle is the CEO and a co-founder of DocketDrafter. He is not a lawyer. This article is educational information, not legal advice. It addresses ABA model materials and Anthropic's published terms, not the binding law of any particular jurisdiction or the requirements of any particular client, court, or matter.