Is ChatGPT Private? A Lawyer's Guide to Data Controls

Is ChatGPT private enough for legal work? The answer depends on your plan, settings, connected services, and the material you share. Consumer conversations may help train OpenAI models unless you opt out; ChatGPT Business, Enterprise, and the API exclude business data from training by default.

For in-house counsel, approval also depends on where copies remain and who can retrieve or receive them. A training opt-out addresses model improvement. Retention, contractual confidentiality, and attorney-client privilege require separate analysis.

GC AI is enterprise legal AI software built for in-house counsel, used by more than 2,100 legal teams. Our customers include TIME, developer-security company Snyk, and healthcare-technology company Helix. We bring your company's documents, review standards, and business context into legal research, drafting, and contract review, with published security controls your team can assess.

That focus grew from co-founder and CEO Cecilia Ziniti's experience working in-house. While at Replit, she experimented with early GPT models to draft contract language and review agreements. Teaching lawyers to use AI helped shape the product she built: company-specific playbooks, sources a lawyer can inspect, and work written for business stakeholders.

For a privacy review, assess both parts of that proposition. The workflow should help your lawyers carry their approved standards through the matter; the contract and configuration should govern what happens to the information they use.

Is ChatGPT Private on Your Plan?

Start with the workspace where the conversation will happen. An individual subscription and a company workspace have different defaults, even when they offer access to the same model.

Service	Model Training	Retention and Access to Check
Free, Go, Plus, and Pro	Consumer content may support training unless you opt out.	Saved chats remain until deletion; Temporary Chat has separate rules. Review sharing, Memory, and app settings.
ChatGPT Business	No training on business data by default.	Administrators can set retention and access workspace conversations.
ChatGPT Enterprise	No training on business data by default.	Administrators control retention and can use compliance tools to audit conversations.
OpenAI API	No training by default.	Abuse logs, stored application data, and approved retention controls depend on the endpoint and features.

Business and Enterprise controls: OpenAI's business privacy commitments, checked October 1, 2026.

Business and Enterprise provide substantive protections. OpenAI's Services Agreement includes confidentiality obligations for covered business services, including its API. Review your organization's applicable agreement, enabled features, and configuration before approving particular data.

Plus and Pro are ad-free consumer plans. Paying for either leaves your model-improvement settings and your company's approved retention policy unchanged.

What Does Astra for Law Change?

Astra for Law combines GPT-6 Astra with a legal search index and instructions for legal analysis and writing. The initial rollout is limited to selected U.S. law firms through Trusted Access in ChatGPT and Codex; API access is coming soon. In-house teams should confirm eligibility before treating it as an available deployment.

For eligible firms, OpenAI's legal offering includes Zero-Data-Retention for API usage and excludes ChatGPT Enterprise usage from human review by default. These are separate protections on different services. They do not make every ChatGPT conversation zero-retention or extend the legal offering to every business account.

Confirm eligibility and the protections enabled for your workspace, API endpoints, and connected tools. The general Enterprise policy also permits limited employee access for incident resolution, recovery with explicit permission, and legal requirements. Review how your firm's terms address those exceptions before approving client material. In-house teams should verify access rather than assume that employing lawyers qualifies the company for the law-firm program.

You can also connect your GC AI workspace through the GC AI plugin for ChatGPT. It supports document-grounded answers, drafting, review, and team playbooks. Availability depends on your plan and workspace settings, including required administrator approval. Evaluate the ChatGPT conversation and GC AI workspace separately when you document access and retention.

What ChatGPT Keeps After a Conversation

Saved chats remain in your account until you delete them or a workspace retention policy removes them. When you delete a chat, OpenAI schedules its permanent deletion within 30 days, subject to security or legal exceptions and the exception for content already de-identified and disassociated from your account.

Temporary Chats stay out of history while temporary, and OpenAI does not use them for model improvement. OpenAI may retain a copy for up to 30 days for safety. Saving a Temporary Chat converts it to a regular conversation, with the corresponding account settings and retention rules.

Files need a separate check. Deleting a chat does not delete a file you saved in Library. Files attached to projects or custom GPTs also have their own lifecycle. Map the copies you created before treating a chat deletion as completion. OpenAI's retention guide explains these distinctions.

Legal holds can change ordinary deletion practices. OpenAI's litigation update says the obligation to retain new consumer data indefinitely ended September 26, 2025; a limited historical dataset remains under preservation. The notice's October update date was not the date that obligation ended.

For a matter with preservation duties, coordinate deletion with counsel and your records policy. Privacy housekeeping should follow the same legal-hold instructions as the underlying documents.

How to Make ChatGPT More Private

Use the controls available to your account and workspace, then confirm the resulting settings before entering work material:

1. In Settings, open Data controls and turn off “Improve the model for everyone” if you use a consumer account. The choice applies to new conversations; it leaves existing chat history in place.

2. Review Memory and personalization separately. They control the context ChatGPT uses for you, while the model-improvement setting controls eligible training use.

3. Choose Temporary Chat when you do not want a conversation in history. Check its personalization choice before starting, and consider the change in handling before saving it.

4. Review shared conversations, connected accounts, and action approvals. Confirm the people and services that will receive the content.

5. Remove unneeded chats, saved files, and memories through their respective controls, subject to your retention duties.

Voluntary feedback needs attention too: if you submit a thumbs-up or thumbs-down, OpenAI may use the associated conversation for training even after a consumer training opt-out. Data controls documentation describes that exception.

Ads have their own controls. On eligible Free and Go accounts, ChatGPT may select ads using the current conversation and, when personalization permits, past chats or memories. Turning off ad personalization can still allow contextual ads. OpenAI's ads FAQ says advertisers receive aggregate reporting that excludes your conversations and memories; information you message directly to an advertiser is a separate disclosure.

What Changes When You Connect Workplace Apps?

A connected app can retrieve company information or perform actions in another service. Review both the access granted by that service and the approvals ChatGPT requests before using it. Approval preferences govern when ChatGPT asks you. The connected account's underlying permissions still determine what the app can access.

Data sent to an app also falls under that app's terms. Depending on the connection, ChatGPT may share relevant conversation context, including information from Memory, to fulfill a request. OpenAI's connected-app guide separates these controls from training and retention. Check the provider-specific rules: Google app data, for example, has additional restrictions on training use.

Sync creates another copy to evaluate. Administrator-managed sync indexes approved workspace content and respects members' permissions in the source system. Disconnecting your individual live account does not remove that workspace-managed index. Ask the administrator which content remains indexed and how to remove it; review existing chats and saved memories separately. OpenAI's sync documentation explains the division.

Shared assistants also change the audience. Where your organization has enabled it, the Slack and Microsoft Teams integration uses a service account and configured company connections. @ChatGPT does not automatically inherit the requester's personal permissions, and personal app access requires separate authorization. Check the participants who can see its reply and the separate permissions on any generated files before requesting a summary of restricted material.

Does Private Safety Processing Mean Zero Retention?

At its September 29, 2026 DevDay announcement, OpenAI described Private Intelligence protections for automated safety review and a Private Inference preview planned for fall 2026. These announcements do not apply one retention rule to every ChatGPT conversation.

The API's ZDR with PSP configuration stores encrypted safety records in customer-controlled cloud storage. Customers must retain those records for at least 30 days. A protected automated runtime performs safety review. Content records stay in the customer's storage; OpenAI retains an operational index.

That distinction matters when someone asks whether “zero retention” means no records exist anywhere. The separate Private Retention with PSP option retains encrypted content on OpenAI infrastructure. Ordinary API abuse-monitoring logs and stored application data have their own rules, and ZDR eligibility varies by endpoint and feature. Have the implementation owner verify the API's applicable data controls, including any downstream application storage.

What Lawyers Should Approve Before Sharing Matter Data

A useful approval identifies the workspace, permitted data, connected services, recipients, and retention requirements for a defined task. For an initial evaluation, use public material or a synthetic agreement. Before using a live contract, investigation file, or board document, confirm your authority to share it and the safeguards the matter requires.

ABA Formal Opinion 512 calls for lawyers to understand the technology and assess confidentiality risks, including whether informed client consent is necessary. A model-training setting does not substitute for that analysis.

Privilege requires its own assessment. In United States v. Heppner, the court rejected privilege and work-product claims over materials the defendant generated through Claude on his own initiative. The opinion did not approve a product category or establish a safe harbor for tools with particular security certifications. Our ChatGPT confidentiality guide covers the contractual and privilege questions in more detail.

For regulated information, confirm that the applicable contract and product features cover the proposed use. For example, OpenAI's business terms require the relevant healthcare agreement for protected health information and warn that some services do not support that use. An organization's approval of one workspace should identify any excluded apps or actions.

Evaluate GC AI on an Approved Legal Workflow

For recurring in-house work, evaluate privacy alongside the review your lawyers need to complete. Our security documentation describes customer data segregation, encryption, and model providers' prohibition on training with your data. We maintain zero-data-retention agreements with our LLM providers wherever feasible. GC AI's own storage is separate: we retain customer data while the account is active and delete it on written request, with legal and operational requirements for other records.

At Helix, a healthcare-technology company, Chief Privacy Officer and Associate General Counsel KT Farley describes using GC AI for privacy research, risk analysis, and drafting. Reusable review instructions help the team bring work back to a lawyer:

Junior teammates now run the checklist prompt first and bring me the output as the predicate for my review.

That is the kind of workflow to evaluate alongside privacy controls: approved context, repeatable instructions, source checking, and a named lawyer making the decision. The customer example illustrates how legal work is organized; it does not establish that every type of data is approved for upload.

A representative vendor agreement gives your team a practical test:

1. Supply an agreement you are authorized to upload and your approved review positions. Use our Playbooks to apply those standards and ask for deviations, proposed language, and unresolved questions with clause references.

2. Inspect the cited language. Exact Quote lets you open a source and examine the quoted passage. Check whether the proposed edit addresses the provision and preserves the intended legal meaning.

3. Decide which changes to accept and which business tradeoffs to escalate. Ask for a draft response that reflects those decisions, then approve the wording and recipients before sending it.

4. Save the reviewed playbook for later matters. If you use Agent Connectors to retrieve files or prepare a response in a connected service, check the account and action permissions. Sensitive actions require approval by default unless you change that setting.

Use this pilot to record the errors you found, the work needed to verify the output, and the decisions counsel retained. Keep your security review alongside that record. Use the trial to assess whether this review process fits your team's work before expanding to additional matters.