Claude for Lawyers (2026): What It Does Well and Where It Falls Short

Claude, the assistant from Anthropic, is a reasonable choice for an in-house lawyer who wants help drafting, summarizing, and reasoning through long documents. What it cannot do is tell you what the law is. A language model predicts plausible text, and the American Bar Association's ethics opinion on generative AI warns that some tools are "prone to 'hallucinations,' providing ostensibly plausible responses that have no basis in fact or reality." This guide covers where Claude helps, which plan decides whether you can paste a client matter, the two ways lawyers use it, and why every citation still needs a check against the primary source.

Reviewed and corrected on 12 October 2026. Vendor facts below come from the Anthropic pages named under Sources, as read on 12 October 2026, and from ABA Formal Opinion 512. Vendors change their terms, so check those pages before relying on this. This is general information, not legal advice.

TL;DR

• Claude is a drafting and analysis assistant, not a source of law. It is useful for long-document work, summaries, and first drafts. It is not a way to find authority or confirm a holding.
• A made-up citation looks like a real one. ABA Formal Opinion 512 lists "citations to nonexistent opinions, inaccurate analysis of authority, and use of misleading arguments" among the problems that have arisen with lawyers' use of generative AI.
• Consumer and business plans run under different contracts. Anthropic's Consumer Terms (Free, Pro, Max) let it use your chats for training "unless you opt out of training through your account settings." Anthropic's privacy center says that if you allow training it may keep your data in a de-identified format for up to five years. Team and Enterprise run under Commercial Terms, and Anthropic says it does not use their inputs or outputs for training by default.
• ABA Formal Opinion 512 already sets the duty. Understand the tool, read its terms, get informed client consent before inputting information relating to the representation where the opinion requires it, and review every output.
• There are two ways lawyers use Claude. The chat assistant at claude.ai for drafting and analysis, and connectors (including MCP servers) that let it read real sources.

What Claude is, and the plan that matters for confidentiality

Claude is a family of large language models built by Anthropic, reached through the chat app at claude.ai, desktop and mobile apps, and an API. ABA Formal Opinion 512 describes the text-producing tools in this class as "prediction tools that generate a statistically probable output when prompted." That one fact explains both the strengths and the failure modes.

For a lawyer, the question before any of that is which plan you are on, because the plan decides the contract. Anthropic's August 2025 announcement says its consumer update applies to Free, Pro and Max, and that it does not apply to services under the Commercial Terms, including "Claude for Work, which includes our Team and Enterprise plans" and the API.

Consumer plans (Free, Pro, Max). Section 4 of the Consumer Terms (effective 8 October 2025) says Anthropic may use your Materials "including training our models, unless you opt out of training through your account settings." It adds that even if you opt out, Anthropic will use Materials for training when you provide feedback or when they are flagged for safety review. Anthropic's retention article for consumer products (1 July 2026) says that if you allow model improvement, it may keep your data "in a de-identified format for up to 5 years in our model training pipelines." Deleting a chat removes it from your history immediately and from back-end storage within 30 days, but if you submit thumbs up/down feedback, Anthropic stores the related conversation for up to 5 years. Anthropic's pages do not say how the training toggle is preset for a given account, so check your own Privacy Settings.

Business plans (Team, Enterprise) and the API. Anthropic's commercial privacy article (18 August 2026) says: "By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models." It adds an exception for feedback you explicitly submit or data you otherwise choose to allow. The Commercial Terms (effective 17 June 2025) say "Anthropic may not train models on Customer Content from Services" and that data is processed under Anthropic's Data Processing Addendum, which is incorporated by reference.

A few practical notes:

• Price does not change the contract. Max is an individual plan. Anthropic's pricing page lists model training for Free, Pro and Max as "Opt-out."
• An opt-out is a setting. A signed commercial agreement is something you can point to if a client asks how the data was handled. Anthropic's retention article says that turning off model improvement stops your previous and new chats from future training runs, but "Your data will still be included in model training runs that are already in progress, or in models that have been trained."
• Zero data retention is narrow. Anthropic's article says some Claude Platform (API) and Claude Code for Enterprise customers, "subject to Anthropic's approval," may have arrangements under which it does not store inputs or outputs except where needed to comply with law or combat misuse. It says Anthropic still keeps safety classifier results, and that the arrangements cover eligible APIs, products using a commercial API key, and Claude Code on Enterprise plans. It does not describe zero data retention as a setting for ordinary Team or Enterprise chat.
• Plans shift. For current features and prices, see Anthropic's pricing page.

We unpack the broader data question, where your inputs travel and what to ask a vendor, in we do not train on your data: how to verify it. For the full Anthropic walk-through see is Claude private for legal work. The same consumer-versus-business split applies to the other big chatbot in is ChatGPT confidential for legal work.

Where Claude is strong for legal work

These are tasks where a general assistant saves time and the verification burden stays lower because you supply the substance. Opinion 512 says the right amount of verification "will necessarily depend on the GAI tool and the specific task that it performs."

• Long-document analysis. Anthropic's pricing page lists a context window of "Up to 1M" that "varies by model," so a long lease, credit agreement or transcript can be pasted in and queried for key dates, obligations or one party's risk. It is working from your text, not its memory. You still check the answer against the document.
• First drafts. Hand it a fact pattern and ask for a demand letter, an internal policy memo, a deposition outline, or a board update. You get a structure quickly, then you edit and source it.
• Summarizing and reorganizing. Turn an email thread into a timeline, or condense a long report into a one-pager.
• Reasoning through a problem you frame. Counterarguments, issue-spotting checklists, negotiation angles, a list of questions to ask a vendor. You are using it to think, not to cite.
• Tone and plain-language work. Soften a blunt note, or explain a concept in plain English that you will confirm elsewhere before relying on it.

For a head-to-head on the drafting tier specifically, see our best AI chatbot for legal writing comparison.

Where Claude is risky, and why it matters

The failure modes cluster around exactly the work that goes to a court or a client.

Citation hallucination

Asked for authority that supports a point, a language model can generate a citation-shaped string: a plausible case name, a real-sounding court, a reporter cite in the right format. Sometimes it matches a real case. Sometimes it does not. Opinion 512 says output from a GAI tool "must be carefully reviewed to ensure that the assertions made to the court are not false," and that lawyers must review "analysis and citations to authority" and correct "misstatements of law and fact, a failure to include controlling legal authority, and misleading arguments."

We track court decisions on fabricated AI citations in our AI hallucination sanctions tracker and lay out a checking workflow in how to verify AI legal citations before filing.

No grounding in your matter or in verified law on its own

Claude does not see your matter files unless you paste them or connect them. Without a connected source, an answer about the law comes from what the model learned in training, not from a document you can open. Opinion 512 notes that if the data a tool was trained on is "limited or outdated," the tool "might produce unreliable, incomplete, or discriminatory results."

Jurisdiction and currency

Laws change, and a model's training data has a cut-off. If you do not name the state and the date that matter, you cannot tell from the answer which jurisdiction or which version of a rule it describes. Treat every statement of law as something to confirm in the official text for your jurisdiction.

A court has already read an AI vendor's privacy policy against a user

Judge Jed Rakoff of the Southern District of New York held in United States v. Heppner (ruling from the bench on 10 February 2026, memorandum dated 17 February 2026) that a criminal defendant's own exchanges with Claude were not protected by attorney-client privilege or by the work product doctrine. The court framed the question as one about communications with a "publicly available AI platform," reasoned partly from Anthropic's privacy policy as of 19 February 2025, and noted that the defendant's lawyers conceded they had not directed him to use it. The opinion says nothing about how a Team or Enterprise plan would fare. Our privilege guide covers the analysis.

Safe vs unsafe uses for lawyers

The line tracks one question: are you asking Claude to write from law you supplied, or to supply the law itself? ABA Formal Opinion 512, issued 29 July 2024, applies the Model Rules to generative AI.

Opinion 512 covers competence (Rule 1.1), confidentiality (Rule 1.6), communication (Rule 1.4), meritorious claims and candor toward the tribunal (Rules 3.1, 3.3, 8.4(c)), supervision (Rules 5.1, 5.3) and fees (Rule 1.5). Three points matter most here:

• Competence. Lawyers "should either acquire a reasonable understanding of the benefits and risks of the GAI tools that they employ in their practices or draw on the expertise of others who can provide guidance about the relevant GAI tool's capabilities and limitations." The amount of verification depends on the tool and the task, and "regardless of the level of review the lawyer selects, the lawyer is fully responsible for the work on behalf of the client."
• Confidentiality. The opinion says that because many self-learning tools are designed so that their output could lead directly or indirectly to disclosure of information relating to a representation, "a client's informed consent is required prior to inputting information relating to the representation into such a GAI tool," and that "merely adding general, boiler-plate provisions to engagement letters" is not sufficient. It also says lawyers "should read and understand the Terms of Use, privacy policy, and related contractual terms and policies of any GAI tool they use." Consent is not necessary where nothing relating to the representation is input, for example when using a tool for idea generation.
• Candor. Review the output, including citations, before it goes to a court.

We translate the full opinion into plain English in our ABA Formal Opinion 512 guide. The opinion applies the ABA Model Rules, so check your own state bar, since several have issued their own guidance.

The two ways lawyers use Claude

There is a beginner path and a power-user path, and they solve different problems.

1. The chat assistant at claude.ai

This is what most lawyers mean by "using Claude." You open the app, paste your text, and draft or analyze in a conversation. It needs no setup. Use it for the reasonable tasks above: scaffolding, summaries of your own text, rewrites, and reasoning you frame. Because it is not reading a verified source, treat every legal proposition it produces as a claim to verify.

2. Connectors and MCP servers

The Model Context Protocol (MCP) lets an AI client call out to outside tools and data sources during a chat. Anthropic's pricing page lists Connectors on every plan, and Claude Code on Pro, Max, Team and Enterprise (not Free). Anthropic's consumer privacy article refers to "remote and local MCP servers" as connectors and says raw connector content is not part of the chat data it may use for improving models, "though data may be included if it's directly copied into your conversation with Claude." So what you copy into the chat still counts.

Retrieving from a real source gives you a document to check, which is a better starting point than a predicted citation. It does not remove the duty to review. If you want to set it up, we have walkthroughs: use Claude for US legal research with MCP covers the connection end to end, and the legal research in Cursor and Claude Code MCP guide shows the developer-tool setup.

Claude and a retrieval-based legal tool

A general chatbot and a tool that retrieves source documents can look identical from the outside. Both are a text box. The difference is what happens before the answer appears.

Grounding. A retrieval-based tool pulls real source documents, such as statutes or opinions, and answers on top of them. A chatbot with no connected source answers from training data.

Citations you can check. When a tool links a cite to the text of a section of the US Code, you can open it and read it. A cite that is only a prediction has to be rebuilt from scratch, which is the step people skip under deadline. Retrieval does not make verification unnecessary. It makes it quicker.

Data terms. Whatever tool you use, the contract that governs your data matters as much as the model. Ask for the data processing agreement and the retention terms in writing. See what a DPA review involves and our roundup of the best legal AI tools for in-house counsel.

So a reasonable division of labor:

• Claude (the chat assistant): first drafts where you supply the law, rewrites, summaries of your own text, brainstorming, long-document analysis.
• A source you can open: anything that asserts what the law is and needs a citation a court can check, or confirming statutory text.

A workflow that keeps you safe

1. Pick the plan with the right contract for the sensitivity of the matter, and read its current terms.
2. Draft and analyze in Claude, supplying the facts and the law yourself.
3. Strip or anonymize anything client-identifying before pasting if you are on a consumer plan, or get informed client consent where Opinion 512 requires it.
4. Verify every legal proposition against the primary source, never from the model's memory.
5. Treat Claude's output as scaffolding. You own every cited claim in the final document.

For the US statutes, regulations and court rules these laws sit in, see the US primary law API.

FAQ

Is Claude good for legal work? For some tasks. Claude is useful for long-document analysis, first drafts, summaries of text you provide, and reasoning you frame. It is not a source of law, because a model can produce citations that do not exist. Use it for drafting and analysis, then verify the law in the official source.

Can Claude do legal research? On its own, the chat assistant predicts text and does not retrieve a verified source, so its citations must be checked one by one. Connecting it to a source through a connector or MCP server gives it something real to read. Either way, you verify every cite before filing.

Does Anthropic train Claude on my data? It depends on the plan. Anthropic's Consumer Terms (Free, Pro, Max) say it may use your Materials for training unless you opt out in your account settings, and it still uses them when you give feedback or a chat is flagged for safety review. If you allow training, Anthropic's retention article says it may keep de-identified data for up to five years. Anthropic's commercial privacy article says it does not use inputs or outputs from its commercial products (Claude for Work, the API) to train its models by default, with an exception for feedback you submit or data you choose to allow.

Is Claude confidential enough for client matters? That depends on your plan, your agreement and your client's consent, not on the model. A personal Free, Pro or Max account has no commercial agreement behind it. For client data, find out whether you are on Team or Enterprise under the Commercial Terms, read the current terms and your admin settings, and decide on consent under ABA Opinion 512.

How long does Anthropic keep my chats? On consumer plans, deleted chats leave back-end storage within 30 days, and if you allow model improvement Anthropic may keep de-identified data for up to five years. Chats flagged for violating its Usage Policy can be kept for up to two years, per its retention page. On Team and Enterprise, chats stay in the product until deleted. See is Claude private for legal work for the detail.

Claude or ChatGPT for legal work? Neither is grounded in verified law by itself, so use either for scaffolding and supply the law yourself. Choose by the data terms on the plan you would buy and what your client has agreed to. We compare them in ChatGPT for lawyers.

Why does Claude make up cases? Because it generates statistically probable text rather than looking up a document. Asked for supporting authority, it can produce a citation-shaped string with no source behind it. That is why Opinion 512 expects you to review citations before submitting them.

Claude or a purpose-built legal tool for in-house work? Many teams use both: Claude for drafting and analysis where they own the facts and the law, and a source they can open for anything that has to cite the law accurately. The split is whether the task needs a checkable source or just good prose.

Sources

All Anthropic pages read on 12 October 2026.

• Updates to Consumer Terms and Privacy Policy (Anthropic, 28 August 2025)
• Consumer Terms of Service (effective 8 October 2025) and Commercial Terms of Service (effective 17 June 2025)
• Anthropic Privacy Center, consumer: Is my data used for model training? (16 March 2026) and How long do you store my data? (1 July 2026)
• Anthropic Privacy Center, commercial: Is my data used for model training? (18 August 2026), How long do you store my organization's data? (1 July 2026), zero data retention agreement and custom data retention for Enterprise plans (both marked "Updated this week")
• Claude pricing (Anthropic)
• United States v. Heppner, No. 25 Cr. 503 (JSR), ECF No. 27 (S.D.N.Y. 17 February 2026)
• ABA Formal Opinion 512, Generative Artificial Intelligence Tools (American Bar Association, 29 July 2024)

Last updated: 12 October 2026.

Arshita Anand

Co-Founder & CEO · Attorney

Arshita leads product and strategy at Vaquill AI, the US primary law data API. As an attorney, she decides which law we carry, how each section is sourced and cited, and what legal AI and compliance teams need from the data.