Is a Free Plan Enough? Understanding Training Opt-Outs for Business Use of Generative AI

Hello, I’m Noriaki Asato, representative attorney at LegalAgent.

Pasting a contract clause into generative AI to ask for revisions, or uploading a customer email to prepare a reply, can be useful. The material may also contain a counterparty’s confidential information or a customer’s personal information. An employee’s familiar personal account may not be an appropriate place to enter it.

“We pay for the service” and “training is turned off” are not, by themselves, enough to approve business use. The plan’s contractual terms, the arrangements for storing and accessing data, and the company’s obligations concerning that information each require examination. This article explains a practical sequence for that review, based on public service information available on November 9, 2026.

1. Check the Plan’s Terms, Not Just Whether It Is Free

Pricing and data treatment do not necessarily align. A paid personal plan may allow model-improvement use by default, while an API may apply different conditions to unpaid and paid services. Payment alone does not establish that data will not be used for training.

OpenAI says sharing is enabled by default in personal Free, Plus and Pro workspaces and can be changed in settings. Business, Enterprise, Edu and API data are generally not used for training by default. Separate consent to data sharing may apply, so identify the actual workspace. (Personal data controls, business data)

Anthropic describes several circumstances affecting personal Claude data, including improvement consent, safety flags and feedback. Google’s Gemini API terms distinguish unpaid and paid services, with regional and other exceptions. An API delivery method alone does not determine the applicable treatment. (Anthropic’s explanation, Gemini API terms)

Record the service, plan, contracting entity, intended features, applicable terms and review date. Personal accounts, company workspaces and APIs accessed through another application may involve different contracts. If a reseller or another application sits between the company and the model provider, examine that intermediary’s storage and use of data as well.

2. Separate Training, History, Retention and Human Access

A training opt-out generally concerns the use of data to train or improve models. Whether a conversation appears in the interface, remains on a server, or can be reviewed by personnel for safety purposes involves separate questions. Disappearance from the screen does not establish deletion from every storage location.

OpenAI explains that conversations remain in history after training is disabled. Temporary conversations can also be retained for safety purposes, and features such as voice have separate controls. Review the actual feature and any special-program exceptions. (Data controls)

Before describing a service internally as “AI that does not store data,” identify what is not stored. Prompts, attachments, outputs, activity logs and backups may have different retention periods. Check post-deletion retention and exceptions for investigating failures or misuse.

Feedback matters too. Rating an answer or submitting a bug report may subject the associated conversation to different conditions. For confidential work, define the permitted scope of feedback, external sharing and connected tools alongside the settings that apply when information is first entered.

3. Confidentiality and Personal-Information Obligations Remain

Is Disclosure Permitted Under the Contract?

If an NDA prohibits disclosure to third parties without prior consent, examine its definition of confidential information, permitted recipients and conditions for disclosure to contractors. A promise not to train on the material does not remove restrictions in the company’s agreement with its counterparty.

Disclosure to contractors might be permitted only if equivalent confidentiality obligations are imposed or prior notice is given. Check whether the AI service’s terms satisfy those conditions. If the agreement expressly prohibits entering information into an external service, consider obtaining consent or changing the proposed use; changing a setting does not resolve the contractual restriction.

Removing company and contact names may be insufficient. Prices, product specifications and negotiating history may themselves be confidential. Where the whole agreement is unnecessary, consider extracting the relevant clause and replacing specific commercial details with hypothetical values. Then check whether the remaining material still identifies a transaction or an individual.

What Legal Basis Applies to Personal Data?

Japan’s Personal Information Protection Commission (PPC) calls for checking whether entering personal information is necessary for the specified purpose of use and whether personal data will be used for machine learning. Removing a name does not necessarily prevent identification through other information. An internal label such as “anonymized” is not conclusive. (PPC Q&A)

Where the activity constitutes provision of personal data to a third party, consider Article 27 of the Act on the Protection of Personal Information (APPI) and its exceptions. Entrustment within the scope necessary to achieve the purpose of use is addressed by Article 27(5)(i), while supervision of the entrusted provider remains relevant under Article 25. Do not assume that every AI service arrangement qualifies without examining the provider’s purposes and contractual terms. (PPC General Rules Guidelines)

If personal data are provided to a third party in a foreign country, Article 28 also requires consideration. The location of a server alone does not settle the issue. Identify who handles the data and examine whether consent or another applicable legal route satisfies the requirements. (PPC Guidelines on Provision to a Third Party in a Foreign Country)

4. Match Actual Input Scenarios to Contractual Terms

For an introduction review, a hypothetical example of the intended task is more useful than a broad statement that “confidential information will be processed.” Comparing contract clauses, classifying customer inquiries and summarizing personnel interviews involve different information and consequences.

Check at least the following, retaining the terms or explanations supporting the answers:

• Training and other use: How are prompts, outputs and attachments used? What defaults, permissions and exceptions apply?
• Retention and deletion: What periods apply to data, logs and backups, including after termination?
• Access and subcontracting: When may provider personnel or subcontractors handle the information?
• Company administration: Does the selected plan support the required user management, authentication, sharing restrictions and activity records?
• Incident response: What notification duties, investigation assistance and contact arrangements are contractual?
• The company’s own duties: Are counterparty confidentiality requirements, explanations or consent concerning individuals, and internal rules satisfied?

Vendor questions should specify the scope rather than stop at “Will you train on our data?” For example:

Under our proposed plan, can prompts, outputs or attachments be used for model training or service improvement? Please identify the contractual provisions, including exceptions for user feedback, particular features and subcontractors.

When we delete a conversation or end the contract, when are the data, activity logs and backups deleted? Please explain retention exceptions and the circumstances in which your personnel may access the material.

Treat unanswered points as constraints on the initial use, rather than as confirmation that there is no issue. One option is to start with tasks that achieve their purpose without personal information.

5. Keep Actual Use Aligned With Approved Settings

Settings reviewed at introduction may not apply to every user. Even where the company has a contract, an employee switching to a personal account may move the work outside the approved conditions. Explain which accounts and features are permitted, what may be entered and where uncertain cases should be referred.

For example, a company might allow proofreading of published text, limit unpublished contracts to an approved environment, and require individual review for applicant or employee information. This is an illustration, not a universal rule. Choose categories based on the information and applicable contracts, and retain the reasons for approval.

Settings records should identify the account, plan and feature as well as the date. Revisit the relevant checks when adding a feature or connection, when terms change, or when a user leaves. Define a contact and initial response for accidental input so that an employee does not simply delete the visible history and assume the matter is resolved.

A training opt-out is an important part of business-use review. Understanding its scope, then assessing the proposed documents against the company’s obligations, provides a more workable basis for continued use.