AI Privacy for Legal Teams: Provider Data Rules

AI privacy depends on the account and contract your firm uses. A provider can exclude business conversations from training while using personal-account conversations to improve its models. Approve a named plan for a defined use before staff enter client material.

Start with the ChatGPT privacy guide if staff already use personal accounts for matter work. For a service that keeps no prompt content after processing, read about zero data retention. The data sovereignty guide separates Australian storage from overseas processing. If client material must stay under your firm’s control, compare the options for private AI.

What Happens to Data You Enter Into AI

A prompt or uploaded document can be used for model training, kept on a server, read by a reviewer or processed overseas. Each activity needs its own answer. Turning off training leaves the other questions open.

Before approving an AI tool, record these four controls:

1. Training: whether prompts, uploads and answers can improve the provider’s models, and whether a contract excludes that use.
2. Retention: how long the provider keeps content, what deletion removes and which safety or legal exceptions extend storage.
3. Human review: which staff or contractors can read content, for what purpose and under which access controls.
4. Storage location: where content is stored and where the model processes it, including any connected services.

The chat history you see is only one copy. A deleted conversation can remain in backup or safety systems, or under a legal hold. OpenAI’s business privacy commitments distinguish deletion from its ordinary retention period and explain authorised access. Google’s consumer Gemini privacy notice says previously reviewed conversations can survive activity deletion for up to three years.

For example, removing a client’s name leaves identifying facts elsewhere in a matter summary. Treat de-identification as a review of the whole document before disclosure.

Compare AI Providers on Client Data

Business terms can change a provider’s training rules, but they do not automatically remove stored conversations or overseas processing. The table separates those controls. An application programming interface (API) connects software to a model and can have different terms from the provider’s chat app.

Provider Data Controls

The OpenAI residency rules cover selected content and features, not every integration. Anthropic separates commercial training rules from retention. Its covered-model policy explains the safety-retention exception to some zero-retention arrangements.

Google’s Workspace privacy terms distinguish Gemini in Workspace apps from the Gemini app. Microsoft’s Copilot privacy guidance separates storage commitments from processing location. Perplexity’s enterprise contract excludes model training, while its data processing addendum allows retention to provide the service.

For Claude versus ChatGPT privacy, both have commercial plans that exclude training by default. Their personal-account choices, residency options and API retention exceptions differ. A personal training toggle changes one permitted use of data. A business contract also defines the provider’s obligations to your firm. Neither makes every proposed client upload acceptable.

Professional Duties That Apply

A solicitor’s confidentiality duty applies before the firm discloses client material to an AI provider. In NSW, Victoria and Western Australia, rule 9 of the Uniform Solicitors’ Conduct Rules governs that disclosure. The Victorian regulator’s confidentiality guidance explains the permitted recipients and limited exceptions, including client authority. Identify the rule in your jurisdiction and the authority for the proposed disclosure.

The joint statement on AI in Australian legal practice comes from the NSW Law Society and the WA and Victorian regulators. It warns against putting confidential, sensitive or privileged client information into public chatbots. For commercial tools, it requires careful review of contractual security terms. Paying for a personal subscription does not establish those protections.

Where the Privacy Act applies, personal information brings separate obligations. The Office of the Australian Information Commissioner’s AI privacy guidance covers permitted use or disclosure under Australian Privacy Principle (APP) 6. It also addresses overseas disclosure under APP 8 and security under APP 11. Client confidentiality can cover information that is outside the Privacy Act’s personal-information rules.

Court restrictions remain separate from a provider’s privacy settings. The NSW Supreme Court’s Practice Note SC Gen 23 restricts AI-generated witness evidence and protects certain compelled or suppressed material. Use the AI practice note guide for the relevant court’s rules before using AI on evidence.

Choose an AI Privacy Guide

Use the provider guides below for the account your staff actually use. ChatGPT, Claude, Gemini and Copilot each separate personal use from business arrangements. DeepSeek, Perplexity and Grok need their own assessment of hosted-service terms and the information staff enter.

AI built into existing software needs the same review. Acrobat AI Assistant works with PDFs, while Grammarly checks writing inside other applications. Windows Recall concerns saved screen content. Apple Intelligence and Meta AI raise separate questions about device features and information shared with an assistant. Choose the guide for the feature that will touch the client material, even when staff already have the underlying software.

For local models, start with Ollama’s guide and distinguish a local model from a cloud option. Local processing shifts responsibility to the firm’s device and system controls. After choosing approved tools and plans, use AI governance to set the firm’s operating rules.