LLM for Law Firms: Confidentiality, Ethics and Safe Use
An LLM for law firms is useful for the reading and first-draft work that fills associate hours: reviewing documents, summarizing depositions, producing a first version of a clause or a letter, and pointing research in the right direction. Two problems come with it. Client information typed into a public AI tool may be stored, reviewed or reused by the provider, which puts confidentiality at risk. And every model can produce confident, wrong output, including citations to cases that do not exist.
Both problems are manageable. The first is mostly about where the model runs and who can see what you send it. The second is about review: nothing a model writes goes to a client or a court until a lawyer has checked it against the source.
This post is general information, not legal advice. Your firm's ethics counsel and your jurisdiction's rules decide what you must do.
Why public AI tools are a confidentiality risk
When a lawyer pastes a draft agreement or a deposition excerpt into a consumer chat tool, that text goes to the provider's servers. What happens next depends on the provider's terms, the plan the user signed up for and the user's settings. Depending on those terms, inputs may be retained for a period, reviewed by staff for abuse monitoring, or used to improve the provider's models. Terms also change over time, and individual lawyers rarely read them.
That creates three risks:
- Disclosure: client information leaves the firm's control and sits with a third party the client never agreed to.
- Reuse: if the provider trains on inputs, fragments of one client's matter could in principle influence answers given to others.
- Privilege: whether sharing privileged material with a technology vendor affects privilege is a question for your own counsel and your jurisdiction. The safer working assumption is to keep privileged material in systems the firm controls.
What ABA Formal Opinion 512 says
On July 29, 2024, the American Bar Association's Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, Generative Artificial Intelligence Tools. It applies existing Model Rules rather than creating new ones. As summarized by the National Conference of Bar Examiners' review of the opinion, the main points are:
- Competence (Rule 1.1). Lawyers need a reasonable understanding of what a tool can and cannot do. The level of independent checking depends on the tool and the task, and generative AI cannot replace a lawyer's own competent work.
- Confidentiality (Rule 1.6). Before entering client information into a tool, a lawyer should assess the risk that people outside the firm can access it. For self-learning tools, which may use inputs to improve, the opinion says a client's informed consent is required before inputting information relating to the representation. "Boilerplate waivers will not suffice as informed consent."
- Due diligence on the tool. Read the terms of use and privacy policy, or rely on a colleague or expert who has, and consult IT or security professionals where needed.
- Supervision (Rules 5.1 and 5.3). Managerial lawyers must set clear firm policies on permissible use and make reasonable efforts to ensure lawyers and staff follow them, with training on the tools, ethics and data security.
- Fees (Rule 1.5). Hourly billing covers actual time spent. Time spent learning a tool is generally not billable to the client.
State bars have issued their own guidance too, and it is not identical everywhere. Check the rules in each jurisdiction where you practice.
Use cases that work, and how to review them
| Use case | What the model does | Required review |
|---|---|---|
| Document review | Flags documents that mention a topic, party or date range; summarizes each one | A lawyer samples flagged and unflagged documents to check recall |
| First drafts | Produces a first version of a clause, letter or memo section from your instructions and templates | Line-by-line edit by the responsible lawyer |
| Deposition summaries | Condenses a transcript into a page-and-line summary by topic | Spot-check every cited page and line against the transcript |
| Research support | Suggests issues, search terms and possible authorities | Every authority read in full in a primary source before use |
Document review
Models are good at reading a large set of documents and saying which ones discuss a given issue. They are less reliable at proving a negative: a document the model did not flag may still matter. Treat model output as a first pass that a lawyer tests by sampling, not as the review itself.
First drafts
Drafting goes best when the model works from your own material: the firm's precedent clause, the term sheet, the client's instructions. Give it those in the prompt and ask for a draft in that style. The lawyer still owns every word that leaves the firm.
Deposition summaries
Ask for a fixed format, such as topic, summary, page and line. Page and line references make the summary checkable, and they make errors obvious when a reviewer opens the transcript. A 300-page transcript, assuming roughly 250 words a page, is about 75,000 words, which is around 100,000 tokens (at about 1.3 tokens per word), so it will usually need to be split into sections that fit the model's context window and summarized section by section.
Research with verification
A general model is not a legal database. It can suggest how to frame an issue or what to search for, but anything it presents as authority has to be found and read in a primary source.
Hallucinated citations and why human review is mandatory
An LLM writes text that is statistically plausible, and a case citation is easy to make plausible: a party name, a reporter, a volume and a page. The model can produce one that looks right and does not exist.
Courts have sanctioned lawyers for exactly this. In Mata v. Avianca, Inc., decided June 22, 2023 in the Southern District of New York, the court found the lawyers had submitted "non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT", then kept standing by them after the court questioned their existence. The court imposed a $5,000 penalty under Rule 11.
The lesson is not that one tool is bad. Any model, including one you host yourself, can do this. The fix is a rule with no exceptions: every citation is pulled up and read in a primary source by a person before a document is filed or sent. Retrieval-augmented generation (RAG), where the model answers from documents you supply, reduces invented content but does not remove it, so the rule stays.
Why a private LLM for law firms helps confidentiality
A private LLM is an open-weight model running on infrastructure the firm controls: its own servers, its cloud account or an isolated network. That changes the confidentiality analysis in concrete ways.
- No third party processes the text. Prompts, documents and outputs stay inside the firm's environment.
- The model does not learn from your matters. A model served for inference has fixed weights: it reads your prompt, answers, and keeps nothing in the model itself. Whether that changes your consent analysis under Opinion 512 is for the firm to decide with ethics counsel, but it removes the training-on-inputs question.
- You control access and records. Single sign-on, role-based permissions and an audit log show who used the model and when, which supports the supervision duties in Rules 5.1 and 5.3. Our post on LLM governance covers the controls in detail.
A private model does not fix hallucinations or replace the competence duty. It narrows the confidentiality problem so the firm can focus on review.
If you want the comparison with hosted APIs on cost and control, see private LLM vs OpenAI. For a chat interface lawyers can use day to day, a self-hosted ChatGPT alternative explains the options.
Adoption checklist for law firms
- Write a firm policy on generative AI: approved tools, permitted uses, prohibited inputs and required review. Name who owns it.
- Decide where client data may be processed. If the answer is "only on firm systems", rule out public tools for client work in writing.
- Read the terms of any tool you approve, or have IT and security do it and record their findings.
- Decide, with ethics counsel, when client consent is needed and how you will obtain it. A boilerplate clause in the engagement letter is not enough on its own.
- Make citation checking a written step in every workflow that can reach a court.
- Train lawyers and staff before they get access, and keep a record of who completed it.
- Turn on logging and review usage regularly for uses nobody approved.
- Update billing guidance so time and costs follow Rule 1.5.
- Re-test when you change the model, and keep a small test set of tasks with known good answers.
Where AI Refinery fits
AI Refinery, a self-hosted LLM platform, installs inside your own infrastructure: your cloud, your VPC, on-premise or a fully air-gapped network, and nothing phones home. You choose open-source models from a curated catalog with license documentation and security scoring, deploy one with one click on your own GPUs, and connect applications through an OpenAI-compatible endpoint. RBAC, SSO, approval workflows and a full audit log are built in, and a dashboard shows request volume, latency, token usage and errors.
It does not include a chat interface or a document search system. Firms pair it with an open-source chat UI such as Open WebUI pointed at the endpoint, and with their own document pipeline. It does not check citations, provide legal research content or satisfy any professional rule on the firm's behalf: those duties stay with the lawyers.
It is not the right fit for a small firm with no IT support and no hardware, where a legal-specific hosted product under a carefully reviewed contract may be simpler. It is also more than you need if you only plan to use AI for non-confidential work such as marketing copy.