Using Claude for Legal Work? Here's What Every Attorney Needs to Know Before They Start
Blog post description.
5/29/20268 min read
Using Claude for Legal Work? Here's What Every Attorney Needs to Know Before They Start
By Angeli Fitch, AI Compliance & Ethics Attorney | 20+ Years of Trial Experience | Creator of the State Bar-Approved CLE Course "AI Ethics for Attorneys"
What is Claude, and why are lawyers using it?
Claude is a large language model AI assistant built by Anthropic. It is capable of drafting documents, summarizing lengthy materials, analyzing contracts, conducting research, answering legal questions, and engaging in extended, complex reasoning tasks. It is available directly through Claude.ai, through Anthropic's API for developers, and increasingly through enterprise integrations built into platforms attorneys already use.
Lawyers are gravitating toward Claude for several reasons. Anthropic has positioned Claude as a particularly capable tool for long-form analysis and document-heavy tasks — the kind of work that defines legal practice. Claude can hold a very large amount of text in its working context at once, which makes it useful for reviewing contracts, synthesizing deposition transcripts, or analyzing document sets without the attorney having to manually break the material into pieces.
It is a genuinely useful tool. That is not the question. The question is what an attorney's ethical obligations require before, during, and after using it on client matters — and those obligations do not bend because the tool is impressive.
Does using Claude create confidentiality risks under Rule 1.6?
Yes, and this is the first question every attorney should answer before using Claude on any client matter.
When you input client information into Claude — a contract, a case summary, a set of facts, a client's name and circumstances — you are transmitting that information to Anthropic's systems. Whether that transmission is permissible under your duty of confidentiality depends on how Claude is being accessed, what Anthropic does with that data, and whether your use is consistent with the reasonable expectations of your client.
The specifics matter here. Anthropic offers different products with materially different data practices. Claude.ai's free and Pro consumer tiers have historically used conversation data for model improvement by default, though users can opt out. Anthropic's API, used by developers and enterprise integrations, operates under different terms — data submitted through the API is not used to train models by default. Enterprise agreements through Anthropic's Claude for Work and Claude for Enterprise offerings include enhanced data protections, BAA availability for applicable use cases, and contractual commitments about data handling.
An attorney who opens Claude.ai on a personal browser account and pastes a client's confidential information into it without understanding which product tier they are using, what the current data retention terms are, and whether those terms are consistent with Rule 1.6 has not met their confidentiality obligation. The tool being reputable does not substitute for the attorney's duty to understand how it handles client data.
Before using Claude on any client matter, review Anthropic's current privacy policy and terms of service for the specific product you are using. If your firm has an enterprise agreement, understand what it covers. If you are accessing Claude through a third-party legal platform, understand whether that platform's agreement with Anthropic governs data handling or whether the consumer terms apply. These are not technical questions — they are professional responsibility questions.
What does the duty of competence require when using Claude for legal research?
Rule 1.1 requires that attorneys exercise the legal knowledge, skill, thoroughness, and preparation reasonably necessary for competent representation. Applied to Claude, this means two things: understanding enough about how Claude works to use it responsibly, and verifying its output before relying on it.
Claude hallucinates. This is not a criticism unique to Claude — it applies to every large language model currently available, including the ones marketed specifically for legal use. Hallucination means the model generates content that sounds authoritative and is factually wrong. In legal research, the most dangerous version of this is fabricated case citations — plausible-looking references to opinions that do not exist, or accurate-looking summaries of cases that misstate the holding.
Claude is generally less prone to confident fabrication than some competing models, and Anthropic has invested significantly in reducing hallucination rates. That investment does not eliminate the risk. It reduces it to a level that still requires attorney verification — not because Claude is untrustworthy relative to other tools, but because the competence standard does not grade on a curve.
Every case citation Claude produces must be verified through a citator service — Shepard's, KeyCite, or BCite — before it appears in any work product that goes to a client, opposing counsel, or a court. Every legal proposition Claude states as settled should be checked against primary sources. The attorney's name goes on the filing. The standard is the attorney's standard.
Can Claude give legal advice to clients, and what are the unauthorized practice implications?
Claude can produce text that resembles legal advice. Whether it constitutes legal advice in a professionally and legally meaningful sense — and what that means for the attorney deploying it — depends on how it is used.
If an attorney uses Claude to draft a client communication and reviews and approves that communication before it is sent, the attorney has given the advice and Claude was the drafting tool. That is a permissible use of technology, subject to the usual competence and confidentiality obligations.
If a firm deploys Claude in a client-facing application — a portal, a chat interface, an intake tool — where Claude responds to client questions about their legal matters without meaningful attorney review of each response, the analysis becomes more complicated. The responses may constitute legal advice. The attorney or firm deploying the system is responsible for that advice under their professional responsibility obligations. If the system is client-facing and automated, the supervision obligation under Rules 5.1 and 5.3 requires that the attorney have designed the system with adequate safeguards — not merely reviewed the technology before deployment and assumed it would perform correctly thereafter.
The unauthorized practice of law analysis runs against the attorney, not Claude. Claude does not have a license that can be violated. The attorney who deploys Claude in ways that result in unreviewed legal advice reaching clients is the one whose professional obligations are at issue.
What are Claude's limitations that lawyers specifically need to understand?
Several limitations are particularly relevant to legal practice.
Claude's knowledge has a training cutoff. As of the current version, Claude's reliable knowledge extends through mid-2025, with some coverage of events after that date depending on the version and whether web search is enabled. For legal research, this means Claude may be unaware of recent decisions, newly enacted statutes, recent regulatory guidance, or current procedural rules. An attorney who relies on Claude's statement that a particular legal standard governs a question, without confirming that the standard has not been modified by subsequent authority, is not performing competent legal research.
Claude does not have access to your jurisdiction's current court rules, local rules, or standing orders unless you provide them or it has access to current web search. Procedural missteps based on outdated or incorrect rules are among the most avoidable — and most damaging — errors in legal practice.
Claude's analysis reflects the information you give it. This sounds obvious but has real consequences. If you describe a client's facts to Claude and ask for a legal analysis, Claude will analyze the facts you described. If your description omitted something material — because you did not know it was material, or because the client had not disclosed it — Claude's analysis will not catch that gap. A lawyer reviewing the situation independently might ask a follow-up question that surfaces the missing fact. Claude will not, unless you design your prompts specifically to probe for what you might have missed.
Claude cannot tell you what it does not know. Large language models generate text that is coherent and confident-sounding regardless of whether the underlying information is accurate or complete. Claude will not reliably flag when it is operating at the edge of its knowledge. The attorney must supply that skepticism.
How should attorneys handle Claude's outputs before using them in client work?
Treat Claude's output the way you would treat a first draft from a capable but unlicensed research assistant who occasionally invents citations, may be unaware of recent developments, and cannot tell you when they are uncertain.
That framing is not an insult to the technology. It is an accurate description of what the technology is and what responsible use of it requires.
For research and legal analysis, verify every case citation, check every stated legal proposition against current primary sources, and review the analysis for internal consistency and jurisdictional accuracy. Claude's reasoning is often sound — but "often" is not the standard for work product that affects a client's rights.
For drafting, review the substance of what Claude produced, not just the form. A well-formatted contract clause that misallocates risk, or a beautifully structured motion that omits a required element, is not improved by Claude's confident presentation of it. The attorney's review needs to be substantive, not editorial.
For client communications, the attorney should review and approve every communication before it reaches the client, regardless of whether Claude drafted it or a first-year associate did. The duty of communication under Rule 1.4 requires that the attorney be responsible for what the client is told about their matter.
Do attorneys need to disclose to clients that they used Claude?
This is an evolving area, and the honest answer as of May 2026 is: it depends on your jurisdiction, the nature of the use, and the terms of your engagement.
Several state bars have issued ethics opinions addressing AI disclosure, and the trend is toward greater transparency rather than less. Some opinions have concluded that disclosure is required when AI plays a meaningful role in the work a client is paying for. Others have framed it as a matter of client communication and reasonable expectations rather than a categorical rule.
Courts are increasingly requiring disclosure of AI use in filings. Several federal courts have standing orders requiring attorneys to certify whether AI was used in preparing documents submitted to the court and to confirm that all AI-generated content was reviewed by a licensed attorney. Violating a standing order of this kind is a serious matter independent of the underlying ethics analysis.
The safest practice is to address AI use proactively in your engagement agreement — not in fine print, but in terms that give clients a genuine understanding of how their matter will be handled and by whom. An attorney who does this is in a far stronger position than one who must explain AI use after a problem arises.
What is the right way to think about Claude as a tool in legal practice?
Claude is a capable, well-designed tool that can make legal work more efficient and in some respects more thorough. It is not a lawyer. It does not have professional judgment, ethical obligations, or accountability for the outcomes it influences. Every one of those qualities belongs to the attorney using it.
The attorneys who use Claude well are the ones who bring the same critical judgment to Claude's output that they would bring to any other source of information or assistance — who verify, who question, who catch what the tool missed, and who remain the author of the work in every sense that matters professionally.
The attorneys who get into trouble are the ones who mistake fluency for accuracy, and efficiency for adequate supervision.
Claude is a powerful addition to legal practice. Your professional obligations are the frame that makes it safe to use.
Where can attorneys get structured guidance on using Claude and other AI tools ethically?
Anthropic publishes usage policies and privacy terms at anthropic.com, and attorneys should review the specific terms for whatever product tier they are using. Your state bar's ethics opinions on AI are the authoritative source for jurisdiction-specific obligations — most state bars now have published guidance, and many have ethics hotlines for informal guidance on specific situations.
For comprehensive, structured education on AI ethics in legal practice — including how to build AI use policies, evaluate tools against your professional obligations, and train your staff on responsible use — I teach AI Ethics for Attorneys, a State Bar-approved CLE course that addresses these questions in practical, jurisdiction-aware terms. [Contact me] to bring this training to your firm or organization.
Angeli Fitch is an AI Compliance & Ethics Attorney and trial lawyer with more than 20 years of experience. She is the creator of the California State Bar-approved CLE course "AI Ethics for Attorneys" and advises law firms and legal professionals on ethical AI adoption, compliance, and governance. She is Of Counsel at Infinity Law Group and available for speaking engagements, CLE instruction, and advisory work.
A note on accuracy: Product details about Claude's data practices, tiers, and terms reflect publicly available information as of May 2026. Anthropic updates its terms and product offerings regularly. Attorneys should verify current data handling practices directly with Anthropic before using any Claude product on client matters, and should not rely on this post as a substitute for reviewing current terms of service.