The biggest misconception in enterprises adopting external generative AI is treating whether input data is “stored” and whether it is “used for training” as the same issue. These two processes involve different security risks and control mechanisms. Once AI training data is incorporated into a model, it may not be possible to completely eliminate the risk of unintended exposure even after the original data is deleted.
To address this, companies should not stop at reading a service’s terms of use. They must also verify whether data is used for training and the scope of its storage, logging, and deletion through contracts, technical configurations, and policy settings.
Storage and Training Carry Different Risks
Stored data may be exposed through breaches or unauthorized access by third parties. However, enterprise AI services generally provide some degree of control over access scope, retention periods, and deletion policies through contractual terms and administrative settings. Training creates a different kind of risk. Once data has been incorporated into the training process, deleting the original data may not fully reverse its influence, and there may still be a possibility that sensitive or confidential information could surface through future outputs or internal representations.
The central question therefore goes beyond simply asking whether data is stored. It must also ask whether contracts and configurations can confirm that stored data will not be used to train the model.
Major Services Have Different Data Handling Policies
For consumer services, the scope of data use can be misunderstood based on feature names alone. When ChatGPT’s “Improve the model for everyone” setting is enabled, conversations involving research, work, or personal data may be used to improve and train the model. Claude Free, Pro, and Max allow users to choose whether their data is used during the signup and usage processes. However, the opt-out setting is checked by default, so conversations and coding sessions may be used to improve the model unless users change it. Gemini manages related data use through its “Keep Activity” setting.
| Service | Relevant setting or method | Data use risk to verify |
|---|---|---|
| ChatGPT | “Improve the model for everyone” | Conversations may be used to improve and train the model unless this setting is disabled |
| Claude | Opt-out method enabled by default | Conversations and coding sessions may be used to improve the model unless users change the setting |
| Gemini | “Keep Activity” | Data use is managed through the activity retention setting |
Even when a company pays for employees’ personal subscriptions, it does not automatically gain control over those accounts. If work information is entered without separate security controls, the company’s internal data may be transmitted to and stored by the provider, and it may be used for training in accordance with the service’s policies.
Verify Contracts and Settings Before Deployment
Enterprises should ensure that internal data is not used to train or improve models in external services. They should not overlook storage risks simply because training has been prohibited. The preferred storage method is ZDR, or Zero Data Retention. ZDR is described as a policy or contractual condition that prohibits the retention or separate logging of prompts, images, and other content beyond what is necessary to process requests, and also prevents their use for model training.
However, ZDR is not available by default across all services and pricing plans, and it is often offered only on a limited basis even under the top-tier Enterprise contract. Operational reviews should verify not only whether training is prohibited, but also whether data is actually stored, the retention period, the scope of log management, and the deletion policy.
RAG Security Must Follow the Entire Data Flow
Even in a retrieval-augmented generation (RAG) environment, using internal tenants or dedicated accounts does not eliminate all risks. Organizations should review encryption, storage region, filtering of personal and sensitive information, retention and deletion policies, SLAs, access permissions, and whether personal information is transferred outside the country.
If internal documents are transmitted to a vector database or an external AI service, the storage location, authorized access, and deletion timeline must be managed across the entire data flow. The starting point for enterprise AI security goes beyond determining whether input data is stored. It requires verifying, through contracts and configurations, the scope and purposes for which AI training data remains. Boan News’ original analysis also illuminates this control framework.
Source Article
This article was prepared after reviewing the following original source: Boan News — Debating Information and Communications Technology Security: Misconceptions and Truths About AI Training, Part 50
![[한국정보공학기술사 보안을 論하다–50] AI 학습에 대한 오해와 진실](https://darkpulse.kr/wp-content/uploads/2026/10/article-job_1792-job_17-13.jpg.webp)
Leave a Reply