What information should not be input to generation AI? Judgment criteria and actions for business data
Bottom line: Do not input passwords or private keys to the generating AI. Input of personal information of customers and employees, undisclosed contracts and business documents will be withheld until the purpose of use, contract, internal approval, and service handling conditions are confirmed. Public information also requires confirmation of usage conditions and output.
The risk of information leakage cannot be determined simply by distinguishing between "free version is prohibited, and paid version allows anything." This article explains what the person in charge should check before entering materials, and what to do if you make a mistake.
What information should not be input to generation AI?
Check Personal Information Protection Commission's Alert Regarding Generated AI Services and IPA's Security Materials for AI Users, and apply them to your company's information classification and service conditions. The following is an initial judgment example for internal use, and does not indicate uniform legal permission or prohibition.
Personal Information Protection Commission's Alert Regarding Generated AI Serviceshttps://www.ppc.go.jp/news/careful_information/230602_AI_utilize_alert/
IPA's Security Materials for AI Usershttps://www.ipa.go.jp/digital/ai/security/ai_security_tips.html
| Information | Examples seen in work | Where to judge and check before inputting |
|---|---|---|
| Password/private key/access token | Screens and configuration files containing login information | Do not enter. Replace with fictitious values instead of real ones |
| Customer/employee personal information | Meeting minutes including name, contact information, performance reviews, and health information | Pending. Confirm the purpose of use, internal regulations, and handling of services with the person in charge of personal information protection |
| Confidential Information/Confidential Materials | Unpublished prices, pre-closing figures, design documents, customer contracts | Pending. Check information owner, confidentiality, and usage environment |
| Materials created by third parties | Proposals, paid reports, and images received from customers | Pending. Check the contract/license and necessary scope |
| Public information | Published information on your company's website | Check reprint conditions and input purpose before using in an approved environment |
"Pending" does not mean that you can just hide your name and enter it right away. Q&A of the Personal Information Protection Commission indicates the need to thoroughly check the scope of the purpose of use and the service provider's use for machine learning for prompts that include personal data. Please make decisions with the responsible department of your company.
Q&A of the Personal Information Protection Commissionhttps://www.ppc.go.jp/all_fad_index/fad2-3-3-q1/
What should I check before entering?
If you cannot answer the following three questions, please stop pasting materials or uploading files.
- Whose information is it? Is it public information, customer/employee information, or confidential? Also check minutes, hidden columns in tables, corners of images, and comments in files.
- Can it be passed to an external service for that purpose? Check customer contracts, internal regulations, purpose of use, and terms of use of third-party materials.
- Which environment will it be sent to and who will check it? Check the product, plan, account, link destination, save and share settings, and the person in charge of checking the output.
If your company's decision standards are not yet in place, first decide who will approve and who to consult using the template provided in How to create internal rules for generative AI.
Is it safe to set it to "not used for learning"?
Whether it is used for model training and whether it is saved on the service are two different things. Separately check the retention period, scope that administrators can view, external linkages, logs, and deletion methods. Even if you operate from an internal environment, data may be sent to the connected API or external application.
Confirmation results will be recorded as "Product name/Plan/Company or personal account/Used functions/Connection destination/Confirmation date". Don't assume that just one switch on the screen will handle all functions the same way.
Can I enter customer information if it is anonymized?
Even if you change the customer name to Company A, you may be able to guess who the customer is based on the department in charge, date of order, and unique project details. Information can also be left in the corners of images, hidden columns in tables, and comments in files.
First, extract only the parts necessary for input. For in-house training, it may be easier to make decisions if you create materials with the same structure using fictitious people, projects, and numbers, rather than making slight changes to the original documents. It is important not to rely on anonymization while leaving the availability of the original data ambiguous.
What should I do if I enter confidential or personal information incorrectly?
- Stop additional input and external sharing.
- Organize usage environment, date and time, type of information, and scope of sharing.
- Report to the internal information management desk and respond according to instructions.
- If authentication information is included, request the administrator to revoke or reissue it.
- Determine whether it is necessary to delete, cancel linkage, or contact the service provider.
Simply deleting a chat cannot guarantee that the information has disappeared from all storage locations. The trail will not be deleted without permission, and the person in charge will confirm the deletion scope and contract terms.
If personal data is involved, the person in charge of personal information protection in the company will refer to the Personal Information Protection Commission's Leak Response Materials and determine whether a report or personal notification is necessary. It is important not to assume that all incorrect entries will be reported immediately, but to immediately contact the person in charge.
Personal Information Protection Commission's Leak Response Materialshttps://www.ppc.go.jp/personalinfo/legal/leakAction/
What do we practice during in-house training?
During the training, we prepare public materials, fictitious business meeting memos, explanatory notes on materials containing actual personal information, and categorize them into Enterable,'' Waiting for confirmation,'' and ``Input unavailable.'' There is no need to include actual confidential information in the materials. Evaluate whether you can answer the reason and where to check.
FAQ
If it's a company account, can I include customer data?
You can't judge based on your account alone. We also check customer contracts, purpose of use, terms of service, and internal approvals.
Is AI useless without inputting information?
You can start by creating formats that use fictitious data, organizing public materials, and considering general work procedures. We will first consider a design that will not pass on unnecessary personal information.
Can I paste the entire minutes on ChatGPT, etc.?
Minutes may include customer names, contact names, terms and conditions, and non-public policies. First, categorize the information and check the approved usage environment and purpose of input. If you are unsure, do not paste it and try the work procedure on a hypothetical meeting minutes.
Before consulting ZETTAi
It will be easier to make decisions if you separate the materials you are considering entering into public information, internal information, personal information, etc. Regarding ZETTAi's AI training for corporations, you can consult with us about how to incorporate pre-input confirmation into work exercises. The scope of individual support will be confirmed at the time of consultation.
Articles you may also want to read
- How to create internal rules for generative AI | Usage guidelines and operation examples
- How to create dummy data for generative AI training? Exercises similar to actual practice and confirmation before input
Reference materials
The following are published materials from public institutions that are referenced in the text. Enter the document name and link destination separately.
- Precautions when using generative AI: Personal Information Protection Commission "Caution regarding the use of generative AI services" https://www.ppc.go.jp/news/careful_information/230602_AI_utilize_alert/
- Personal data input judgment: Personal Information Protection Commission "Q&A when entering personal data into prompts" https://www.ppc.go.jp/all_fad_index/fad2-3-3-q1/
- Response to leaks, etc.: Personal Information Protection Commission "Response to leaks, etc. and useful materials" https://www.ppc.go.jp/personalinfo/legal/leakAction/
- Safety measures for users: IPA “Security Tips for AI Users” https://www.ipa.go.jp/digital/ai/security/ai_security_tips.html
Reference date confirmed: November 14, 2026. Please separately check the data usage conditions of the AI service you will actually use and your company's information management regulations.