ChatGPT for Lawyers: What to Know Before You Use It With Client Work

If your practice is experimenting with faster ways to prepare routine legal work, it is normal to be unsure where useful assistance ends and professional risk begins. You do not need to reject the technology or give it unrestricted access to client matters. This guide explains how to classify common uses, review the main risks and put sensible controls around ChatGPT before it becomes part of everyday client work.

Start by separating useful drafting from sensitive legal work

The central question is not whether lawyers can use ChatGPT, but which tasks can be performed with an acceptable level of risk. A general-purpose chatbot can rewrite, summarise, classify and generate text quickly. Those capabilities can save time, but the risk changes sharply when the prompt contains identifiable client information, privileged communications, litigation strategy, unreleased transaction details or facts that will be relied on in advice or court material.

Before setting any rules, Alex's small commercial practice used ChatGPT for almost everything, from client letters to internal case summaries. No distinction was made between formatting an already-approved paragraph and uploading a case file for analysis. After reviewing the workflow, Alex separates lower-risk uses such as formatting, brainstorming headings and general research planning from higher-risk uses such as processing client files, analysing evidence or producing legal propositions. The practice then chooses a managed plan with contractual and administrative controls suitable for professional work.

Why ChatGPT creates specific risks for lawyers

Legal work combines sensitive information with outputs that may affect a client's rights, money or reputation. That makes ordinary chatbot weaknesses more serious in a law practice than they may be in low-stakes marketing or administration. The main issues are confidentiality, factual and legal accuracy, supervision, transparency with clients, billing and the security arrangements surrounding the account.

A confident answer from ChatGPT can still be incomplete, outdated or invented. It may combine a correct general principle with a false citation, misstate a jurisdictional rule or overlook a fact that changes the analysis. The tool does not know which parts of its response are safe to rely on, and fluent wording can make weak material look finished.

Confidentiality risk arises earlier in the workflow. The problem is not limited to publishing information. Copying client material into a third-party service is itself a disclosure to a service provider, which means the firm needs to understand the product, the applicable terms, the purpose of the disclosure and the safeguards around the data.

What professional guidance says

Professional bodies generally apply existing duties to generative AI rather than treating it as an ethics-free category. In the United States, the American Bar Association's Formal Opinion 512 discusses competence, confidentiality, communication, supervision, candour and reasonable fees when lawyers use generative AI. It explains that lawyers need a reasonable understanding of a tool's capabilities and limitations, and that client information may require informed consent before it is entered into a system.

The opinion does not create one universal rule for every jurisdiction or matter. It provides a useful framework for asking the right questions. Local bar rules, professional conduct rules, court directions, client contracts and sector-specific privacy requirements may set different or additional expectations.

Courts are also emphasising personal responsibility for filed material. The Courts and Tribunals Judiciary of England and Wales states in its 2025 judicial guidance on artificial intelligence that users remain personally responsible for material produced in their name and should be alert to inaccurate or fabricated content. Although that guidance is written for judicial office holders, the underlying warning is directly relevant to lawyers preparing submissions, evidence summaries and authorities.

Confidentiality and privilege need separate attention

Removing a client's name does not automatically make a prompt safe. A matter can remain identifiable through dates, transaction values, locations, counterparties, medical facts, unusual allegations or a combination of details. A copied email chain may also contain metadata, signatures, personal contact details and information about people who are not the firm's client.

Privilege adds another layer because the consequences of disclosure depend on the jurisdiction, the circumstances and the controls used. A firm should not assume that a confidentiality promise in a marketing page settles privilege, professional secrecy or contractual obligations. The practical step is to obtain advice appropriate to the firm's jurisdiction and create a clear internal rule for material that cannot be entered without approval.

For matters that genuinely require AI processing of client information, the firm should examine the provider's contract, retention settings, access controls, subprocessors, security certifications, data location options, deletion process and incident response commitments. OpenAI's enterprise privacy information states that organisations own and control their business data and that covered business data is not used for model training by default. Those statements are relevant, but the firm still needs to assess whether the available controls match its own duties and client commitments.

Accuracy requires a verification workflow, not a warning label

A policy that merely says "check the output" is too vague for legal work. Staff need to know what checking means. For legal propositions, verification should normally involve reading the primary authority in an approved legal research source, confirming that the authority exists, checking jurisdiction and currency, and assessing whether the cited passage actually supports the proposition.

For factual summaries, the reviewer should compare the output against the source documents and identify omissions, contradictions and uncertain statements. For calculations, dates and procedural requirements, the answer should be checked independently. For client correspondence, a responsible lawyer should confirm that the tone, facts, instructions and legal position are correct before the document leaves the practice.

ChatGPT can still be useful inside this process. It can create an issue list, turn verified notes into a clearer structure, suggest questions for further research or produce alternative wording. The safer pattern is to use it as a drafting assistant around verified material, not as the source of truth.

A simple traffic-light model for legal tasks

This model is a starting point, not a legal conclusion. A task that appears lower risk may move into a higher category once real client information is added. The firm should tailor the categories to its practice areas, local rules, client requirements and technical environment.

Practical controls to put in place

A small practice does not need an enterprise governance department, but it does need clear ownership and repeatable rules. Start with a short list of approved tools and accounts. Personal accounts should not become an unofficial route around the firm's settings, record keeping or purchasing decisions.

1. Appoint an owner. Give one partner or senior lawyer responsibility for approving tools, reviewing incidents and updating the rules.
2. Define allowed and prohibited data. Use examples from the practice, including client names, contract drafts, discovery documents, personal data and commercially sensitive information.
3. Use the minimum information needed. Replace real details with neutral placeholders where the task does not require them.
4. Set a verification standard. State how authorities, quotations, facts, calculations and procedural information are checked.
5. Require human sign-off. No AI-generated material should be sent, filed or relied on merely because it reads well.
6. Control access. Use firm-managed accounts, multi-factor authentication, role-based access and a process for removing former staff.
7. Keep an incident route. Staff should know what to do if sensitive information is entered by mistake or an inaccurate output reaches a client.
8. Review client and court requirements. Some clients, matters or courts may require disclosure, consent or restrictions that go beyond the firm's general rule.

Client communication, supervision and billing

Using ChatGPT does not remove the lawyer's responsibility for the work or for the people using the tool. Partners and supervising lawyers should treat AI-assisted work like other delegated work, with instructions, review and accountability appropriate to the task. Staff training should include realistic examples of unsafe prompts and unreliable outputs, not just a general reminder to be careful.

Whether a client should be told about a particular use depends on the jurisdiction, the engagement and the significance of the tool to the work. ABA Formal Opinion 512 notes that some uses may require consultation or informed consent, particularly where client information is disclosed or the tool materially affects the representation. A firm can make this easier by deciding in advance which uses require matter-specific discussion rather than leaving each staff member to improvise.

Billing also needs honest treatment. If a task takes substantially less time because of AI assistance, a time-based invoice should reflect the time actually spent, subject to the applicable fee rules and engagement terms. The value of the lawyer's judgement, review and responsibility remains real, but a chatbot should not become a reason to bill fictional hours.

How global regulation fits into the picture

Professional conduct is only one part of the risk assessment. A law practice may also need to consider privacy, data protection, employment, records management and AI-specific rules, depending on where it operates and whose information it handles. For example, the European Union's Artificial Intelligence Act uses a risk-based framework, while the General Data Protection Regulation applies to the processing of personal data within its scope.

Not every use of ChatGPT by a lawyer will be classified as high risk under the EU AI Act. The classification depends on the system, purpose and context. The useful business lesson is broader: identify the purpose, the people affected, the data involved and the level of human oversight before adopting an AI workflow.

International practices can also use governance frameworks such as ISO/IEC 42001 as a reference for managing AI responsibilities, risks and controls. Certification is not required for a small firm to borrow the basic discipline of assigning ownership, documenting uses, assessing risk and monitoring performance.

Questions to ask before using ChatGPT on client work

A short pre-use review can prevent most avoidable mistakes. Ask what problem the tool is solving, whether client information is genuinely necessary, which account and contract will apply, how long data is retained, who can access the workspace and how the output will be verified. Then decide whether the task is approved, needs escalation or should stay outside the tool.

The result should be a practical rule that staff can follow during a busy day. A twelve-page policy that nobody remembers is less useful than a clear one-page matrix supported by training, managed accounts and a named person who can answer questions.

Frequently asked questions

Related reading: our AI governance by region.

Try our free AI Tool Selector to get a personalised AI tool recommendation for your business.

Can lawyers use ChatGPT for client work?

Lawyers can use generative AI in many jurisdictions, but the ordinary duties attached to legal practice still apply. The appropriate use depends on the task, the information entered, the account controls, local professional rules, client instructions and the lawyer's verification process. Treat the tool as supervised assistance, not an independent legal professional.

Is it safe to paste a client document into ChatGPT?

Do not assume it is safe merely because the document is being used for work. Review confidentiality, privilege, privacy, contractual and security considerations first. A managed business product may offer stronger contractual and administrative controls than an individual account, but the firm still needs to assess the specific service and use case.

Does anonymising a prompt solve the confidentiality problem?

It can reduce risk, but weak anonymisation may leave a person or matter identifiable through context. Remove unnecessary facts, use placeholders and consider whether the combination of details could identify the matter. Some tasks can be completed with fictional or generic information instead of client data.

Can ChatGPT be used for legal research?

It can help identify issues, generate search terms and organise a research plan. It should not be relied on as the authority itself. Check every case, statute, quotation and legal proposition in an approved primary or authoritative source, and confirm jurisdiction and currency before relying on it.

Do lawyers need to tell clients that ChatGPT was used?

There is no single global answer. Disclosure or consent may depend on local professional rules, the engagement, the nature of the use and whether client information is shared with the service. Firms should define which uses require matter-specific discussion and seek jurisdiction-specific guidance.

Should a small law firm use ChatGPT Business or Enterprise?

A managed business workspace is generally a more appropriate starting point than unmanaged personal accounts because it can provide central administration and business-data protections. The final choice should follow a review of contract terms, retention, security, access, data residency needs and the sensitivity of intended use. Product names and features can change, so confirm the current vendor documentation.