Is ChatGPT Confidential? What Counts as a Breach in 2026

Is ChatGPT confidential enough for your company's NDA? Their business terms provide confidentiality protections, but whether an upload breaches your NDA depends on the agreement, the account, and how the information is used. Turning off model training answers only part of that question.

For an in-house lawyer, the problem often arrives after the upload. Sales has summarized a customer's term sheet, or engineering has pasted source code into a personal account. You need to establish what was shared, who could receive it, and whether the company had permission.

We built GC AI, enterprise legal AI software for in-house counsel, to help teams review agreements, apply company positions, and turn source-backed analysis into advice for the business. More than 2,100 legal teams use GC AI, including companies like TIME, Nestle, Helix, Snyk, and SKIMS.

Co-founder and CEO Cecilia Ziniti began exploring early GPT while serving as Replit's general counsel. Her experience testing and teaching legal AI showed the need for precise citations, document sourcing, and professional legal writing. She and co-founder Bardia Pourvakil built GC AI around the work in-house teams do, including applying their own negotiating positions and preparing advice for colleagues outside legal.

That purpose matters when an NDA review has to produce a usable decision: what information the company may share, with whom, and under which conditions. Your team's approval should specify the information the platform may receive and the work it may perform.

What “Confidential” Means in an NDA

An NDA's definition determines which information it protects. An illustrative definition might read:

“Confidential Information” means any non-public information disclosed by one party to the other, whether disclosed in writing, orally, or by inspection, that is marked confidential, identified as confidential at the time of disclosure, or that a reasonable person would understand to be confidential under the circumstances.

The receiving party may then promise to:

1. Hold the Confidential Information in strict confidence.
2. Use it only for the Permitted Purpose.
3. Limit access to employees, advisors, and representatives who need to know and are bound by sufficiently protective confidentiality obligations.
4. Return or destroy it at the end of the engagement or upon written request.

Your signed NDA may define these obligations differently. Read the signed agreement, including exceptions for information already known, independently developed, or lawfully public. Our confidentiality clause guide explains those drafting choices.

For an AI upload, check the NDA's permitted recipients and purpose. Determine whether service providers may receive the information and what confidentiality obligations they must accept. Check for consent requirements or prohibited uses. A provider's confidentiality commitment can help satisfy those requirements, but it does not amend the NDA.

A data processing agreement, or DPA, addresses personal-data processing. Commercial confidentiality can also cover source code, pricing, and strategy that contain no personal data. Review both sets of obligations when both apply.

Is ChatGPT Confidential Under Its Business Terms?

OpenAI's Services Agreement applies to business services including ChatGPT Business, ChatGPT Enterprise, and the API. Its confidentiality provisions cover Customer Content, including inputs and outputs. They restrict use and disclosure, while allowing specified disclosures to personnel, contractors, and other recipients and disclosures required by law.

That is a contractual commitment counsel can assess against an NDA. A plan name alone cannot establish that every upload is permitted.

Personal plans have terms of service. Paying for Plus or Pro, however, does not by itself establish an approved business deployment or permission under your company's NDA. OpenAI's Data Controls FAQ explains training controls; the ChatGPT privacy guide covers the settings and retention questions in more detail.

OpenAI says it does not train on business data by default. Its enterprise privacy documentation also describes administrator access and retention controls for both Business and Enterprise. Business administrators can access, export, and delete end-user conversations; Enterprise offers compliance access and audit capabilities. Confirm the controls in the workspace you are approving, including who can inspect the work.

For an API application, review each endpoint and storage layer. OpenAI's API data documentation distinguishes abuse-monitoring logs from application state and limits Zero Data Retention to eligible configurations. ZDR does not establish that the application, connected services, or every endpoint stores nothing. Private Safety Processing is another scoped safety-processing control to assess where available; it does not authorize a disclosure under an NDA.

What Does Astra for Law Change for Confidential Work?

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.

The legal offering includes API Zero Data Retention and exclusion of ChatGPT Enterprise usage from human review by default for eligible firms. Ask which services and endpoints your agreement covers and what access exceptions remain. OpenAI's general Enterprise policy permits limited employee access for incident resolution, recovery with explicit permission, and legal requirements.

Those protections are relevant to vendor approval, but the signed NDA still determines permitted recipients, purpose, and any required consent. Confirm the firm's eligibility and actual configuration, including connected services and the audience for outputs. The offering does not itself establish attorney-client privilege or permission to disclose a counterparty's information.

The GC AI plugin for ChatGPT can connect an authorized GC AI workspace for document-grounded questions, drafting, and review against company standards. Assess that connected workflow separately: the plugin does not establish that every ChatGPT copy inherits the retention terms of direct GC AI use or receives legal privilege.

Check Connected Apps and the Conversation Audience

Connecting a document repository adds another part of the disclosure analysis. Check the source permissions, the account authorizing access, what the app can retrieve or send, and the provider's terms. A tool's permission to read a file and its retention of retrieved content are separate questions. OpenAI's connected-apps documentation explains that app authorization, source permissions, and workspace controls all matter.

In shared @ChatGPT conversations in Slack or Microsoft Teams, enabled connections use the configured company account's permissions, which may differ from each participant's personal access. Because shared replies are visible to conversation participants, approve both the connected sources and the conversation audience before using confidential material. OpenAI's shared-conversation guide describes this setup, and our guide to who can see the answers @ChatGPT posts in Slack and Teams covers the audience question in more detail.

For example, permission to retrieve a deal document does not mean everyone in the channel is an authorized recipient under the deal's NDA. Include the output's audience in the approval.

Six Types of Information That Need a Closer Look

These common workplace examples help employees recognize material that needs approval. Each calls for a review of the actual information and obligations:

1. Before using AI to debug source code or review engineering specifications, check ownership, trade-secret safeguards, and confidentiality commitments to customers or development partners.
2. Summaries of customer contracts and pricing can expose negotiated discounts, commitments, or strategy. Check the customer agreement as well as any separate NDA.
3. An investigation summary or compensation spreadsheet can raise confidentiality, privacy, and employment-law issues. Removing names may leave employees identifiable.
4. Draft forecasts, board materials, and acquisition plans may have restricted recipients and purposes. Confirm authorization before uploading an excerpt.
5. Involve counsel before submitting advice, witness notes, or case strategy. Privilege and work-product protection require their own analysis.
6. A supplier's technical proposal may belong to the supplier and carry use restrictions even when your company stores it.

An unauthorized disclosure may breach a contractual obligation without proving that the material became public. Counsel should assess breach, remedies, trade-secret consequences, and any notification duty separately.

Confidentiality and Legal Privilege Are Different Questions

An NDA is a contractual obligation. Attorney-client privilege and work-product protection govern whether particular material can be withheld from disclosure in legal proceedings. A no-training promise, security audit, or lawyer's participation cannot establish those protections by itself.

In United States v. Heppner, the Southern District of New York rejected privilege and work-product claims for documents a defendant created through his own use of consumer Claude. The February 17, 2026 opinion considered the absence of an attorney relationship with the chatbot, the relevant privacy terms, and the fact that counsel had not directed the work. Sending the resulting documents to counsel later did not make them privileged.

The ruling concerned those facts and that use of Claude. It did not establish a universal rule for every AI deployment, and its discussion of possible counsel-directed use did not create a safe harbor.

ABA Formal Opinion 512 likewise calls for a fact-specific assessment of the tool, matter, and disclosure risks. Lawyers should evaluate the relevant terms and determine whether informed client consent is required. Get that analysis into the workflow before sensitive material is submitted.

What to Do After an Employee Has Uploaded Information

The first job is to understand the disclosure well enough to respond. Give employees a clear reporting route to legal and security, then work through these steps:

1. Identify the account, workspace, model or application, settings, connected tools, exact material, recipients, and time of the upload. Capture the relevant terms and configuration.
2. Keep an appropriate incident record and consider legal-hold obligations before deleting evidence or conversations.
3. Stop additional sharing, review shared links and connected access, and request deletion where appropriate. OpenAI's retention policy describes deletion timing and exceptions; deleting a chat does not undo a completed disclosure.
4. Counsel should review the NDA, applicable law, privilege, and notification provisions. Determine who must be told, what the notice must contain, and the actual deadline.

Record the decision and any corrective steps. Avoid promising a counterparty that no copy exists unless the evidence supports that statement.

Give Employees a Policy They Can Follow

An AI policy should tell employees which account they can use for a given task. Legal, security, and the business owner can take these steps:

• Name the approved workspace, authorized tasks, and data categories. Explain how employees request an exception.
• Identify material requiring additional approval, such as privileged advice, investigation records, or counterparty information subject to consent.
• Ask employees to use only the information needed and to check whether redacted material remains identifiable or confidential.
• Require approval for apps, company accounts, channel audiences, and external sharing.
• Specify who verifies a contract summary, proposed clause, or legal answer before anyone relies on it or sends it externally.
• Assign responsibility for deletion, access removal, and return-or-destroy requests, subject to legal holds.
• Give employees a named team or channel and a prompt reporting expectation that lets counsel meet the applicable deadlines.
• Reassess approvals when terms, providers, features, or the company's obligations change.

These are policy choices to adapt to your business. A blanket label such as “AI approved” gives employees too little information about the work they can do.

What to Ask an AI Vendor Before Approval

Apply the same questions to OpenAI, GC AI, and any other provider receiving confidential material:

1. Which contract governs this service, and does its confidentiality definition cover inputs, outputs, and derived material?
2. What uses and disclosures are permitted, and which personnel, subprocessors, or other recipients can access the information?
3. How do training, safety processing, application storage, logs, and backups differ?
4. Which retention controls apply to this plan and feature, and what exceptions remain?
5. What can administrators and other users see, export, share, or delete?
6. Which connections can retrieve or transmit data, using whose authorization?
7. How do deletion, account termination, and return-or-destroy requests work?
8. What do the agreement and order form provide for incident notice, liability, indemnities, audit evidence, and service commitments?

Ask for the governing documents and evidence behind the answers. A SOC report helps assess controls; it does not decide whether an NDA permits the disclosure. Treat a vendor's product summary as the start of contract review.

Use GC AI to Review the NDA and Record the Decision

Once your team has approved the environment for the material, GC AI can help turn a confidentiality review into a usable decision for the business. Start with a representative agreement you are authorized to upload and a question: may this information be used in this proposed AI workflow?

At developer-security company Snyk, Senior Managing Counsel Alexis Palmer describes using GC AI to research regulatory requirements that arise in customer agreements and draft language for the negotiation.

Saved prompts let teammates start with the same review approach and apply their own judgment. That is a concrete distinction to test against an ad hoc chat: how much of your team's context and approved method carries into the next review?

1. Include the relevant amendments, the proposed use, the information involved, and the intended recipients. Ask for an obligations table with clause references and a list of unresolved questions.
2. Our Playbooks let you review against defined standards and fallbacks. Compare the NDA's permitted recipients, use restrictions, protection standard, and return-or-destroy terms with the workflow you want to approve.
3. Use Exact Quote to open cited source passages. A lawyer should check the actual language, exceptions, and amendments, then decide whether the use is permitted, needs consent, or requires a contractual change.
4. Draft the requested consent, proposed clause, or internal approval with its conditions. Our GC AI for Word add-in supports working on the contract where the team edits it. Check the draft before sending it.
5. Save the decision in your approval record, including its scope and owner. When counsel approves a reusable contract position, add it to the team's Playbook so later reviews start with that standard.

Tiffany Lee, General Counsel and Corporate Secretary at Liquid Death:

If it sees a missing confidentiality clause, I’ll just ask it to draft one I can drop right into the agreement

The lawyer still decides whether the clause and proposed disclosure meet the company's obligations.

Our model providers are contractually prohibited from training on customer data and we maintain zero data retention agreements wherever feasible. Review GC AI's own storage and deletion terms, the relevant provider configuration, and your agreement when approving a use.

Our subprocessor list distinguishes providers that process Customer Content on a zero-data-retention basis from other subprocessors. Under our DPA, we require subprocessors handling Customer Personal Data to accept written data-protection obligations at least as protective as ours. We remain responsible for their acts and omissions to the extent they cause a breach of our DPA obligations. We notify customers before adding or replacing a subprocessor and provide a 30-day written objection period. Review the applicable agreement alongside the actual data flow.

Test the workflow on one approved NDA. Check the obligations table against the agreement, have counsel resolve the open questions, and see whether the result gives the business a clear permission, condition, or required next step.

Start your 14-day trial