OpenAI's generative AI tools offer game-changing capabilities, but they also raise serious data privacy considerations. Global companies evaluating or negotiating OpenAI agreements must address issues from GDPR/CCPA compliance to handling sensitive business data used for model fine-tuning.
Protecting Data Privacy and Confidentiality
Enterprise contracts should explicitly safeguard all information you send to the AI and any AI-generated output. Treat this as confidential and ensure OpenAI cannot use or share your data beyond providing the service.
Secondary Use
Prohibit OpenAI from mining inputs/outputs for model training or any purpose outside your instructions.
Confidentiality
Require robust safeguards and no unauthorized disclosure to third parties.
Retention Control
Opt for minimal retention. Secure right to request deletion on demand with certification.
Meeting GDPR, CCPA, and Global Compliance
Global privacy regulations apply fully when you feed personal data into OpenAI. Sign a Data Processing Addendum (DPA) that spells out each party's privacy obligations — OpenAI as data processor, you as data controller.
- Lawful basis & scope: Ensure you have a lawful basis to process personal data. The DPA should limit OpenAI's processing to only what you authorize.
- Cross-border data transfer: Confirm appropriate transfer mechanisms (Standard Contractual Clauses) are in place for data leaving your region.
- Individual rights & deletion: OpenAI must assist with deletion or access requests promptly. Verify the DPA includes clear provisions for data erasure and audit rights.
- Privacy by design: Ask about built-in compliance features. ChatGPT Enterprise allows retention policies and admin monitoring tools. For regulated sectors, ensure the contract acknowledges additional requirements (e.g., HIPAA BAA for health data).
Model Fine-Tuning: Hidden Data Pitfalls
When you fine-tune an AI with proprietary or personal data, that information becomes part of the model's training memories. The risk: the model might regurgitate snippets of sensitive data in responses.
- Limit sensitive data: Avoid PII or highly confidential data in fine-tuning. Anonymize or mask when you must include it.
- Exclusive use & confidentiality: Ensure the fine-tuned model is for your exclusive use with the same confidentiality protections. OpenAI should not deploy your model for other customers.
- Right to delete models: Negotiate rights to delete the fine-tuned model and underlying training data if you end the service.
- Test for leaks: Conduct "red-team" tests after fine-tuning. Query the model with various prompts to check for inadvertent data exposure. Require OpenAI's assistance in testing.
Indemnification and Liability Clauses
Most OpenAI contracts offer some standard protections, but scrutinize these clauses to avoid bearing all risk.
Indemnification: Ensure OpenAI's IP indemnity is in your contract and as broad as possible. Each party should indemnify the other for risks under their control — OpenAI for AI technology issues, you for misuse of the service.
Liability limits: OpenAI likely caps liability and excludes indirect damages. A strict cap may be unacceptable for mission-critical tasks. Negotiate this. Aim for carve-outs from the cap:
- Breach of confidentiality/privacy: If OpenAI violates privacy terms, losses shouldn't be limited to a token amount.
- Regulatory fines: Argue for coverage when fines result directly from OpenAI's breach.
- Gross negligence/willful misconduct: Ensure liability limits don't apply in these scenarios.
Data Security and Governance
- Security standards: Require SOC 2 Type II, ISO 27001 certification, encryption at rest and in transit, regular security audits.
- Breach notification: Oblige OpenAI to notify you within 24-72 hours of any security incident involving your data.
- Access controls: Clarify who at OpenAI can access your data. Use admin dashboards, role-based access, and SSO to manage exposure.
- Data residency: Take advantage of EU or regional data residency options. Require logical segregation from other customers' data.
Risk Mitigation Summary
| Risk Area | Mitigation in Contract & Practice |
|---|---|
| Unauthorized data use | Explicit "no-training" clause. Data solely for your organization's service. |
| GDPR/CCPA non-compliance | Sign DPA. Include deletion obligations, data subject request assistance, EU SCCs. |
| Over-retention of data | Negotiate retention limits. Leverage Enterprise features for minimal retention. |
| Data breach liability | Carve out breaches from liability limits. Require prompt notification and cooperation. |
| Fine-tuning leaks | Ensure exclusive use. Test models for leakage. Limit PII in training datasets. |
Recommendations
Demand a robust DPA
Execute OpenAI's Data Processing Addendum covering GDPR, CCPA, breach notification, and compliance request assistance.
Lock down data use in contract
Add clear confidentiality clauses. Specify OpenAI may only use your data to provide service — no sharing, no secondary use.
Set data controls and retention
Choose zero-retention for sensitive inputs. Ensure you can request data deletion at any time with swift compliance.
Carve out critical liabilities
Push to carve data breaches, confidentiality breaches, and IP indemnity from liability caps. Get higher or uncapped protection.
Secure IP indemnity and more
Ensure indemnification for IP claims. Discuss coverage for privacy violations or defamation. Even partial agreement highlights your concerns.
Insist on security assurances
Confirm industry security standards, incident notification obligations, and assistance with security investigations.
Prepare internal guidelines
Develop rules for employees — prohibit entering customer PII or secret source code. Control what goes into the model to reduce breach risk.
Evaluate the need to share
Before sending any dataset (especially for fine-tuning), evaluate if it's truly necessary. Share minimum data required. Less data = lower risk.
Checklist: 5 Actions to Take
FAQ
Can we use OpenAI without violating GDPR or CCPA?
Yes, with compliance steps. Have OpenAI sign a DPA committing to GDPR/CCPA principles. Utilize data retention controls. Avoid inputting personal data unless necessary with proper legal basis. With proper contract and configuration, OpenAI can be used in line with global privacy laws.
Will OpenAI use our data to train its models?
For enterprise and API users, OpenAI does not use your data to train general models by default. Your prompts and outputs stay isolated. Confirm this explicitly in your contract — ensure it states OpenAI won't use your data for research or improvement without permission.
What happens if OpenAI has a data breach involving our information?
OpenAI should inform you immediately. You'll manage impact on your side (notifying individuals/regulators). Negotiate liability so that if the breach was OpenAI's fault, they cover costs like regulatory fines and customer notifications. Don't accept overly strict liability limits for breaches.
How do we handle highly sensitive or regulated data?
Cautiously and with extra safeguards. Check if special agreements are needed (HIPAA BAA for health data). Consider specialized cloud regions (Azure OpenAI) for jurisdictional requirements. Minimize what you share — every piece of sensitive data that stays out of the system is one less piece that could be exposed.
Do we own the AI's outputs and our data?
Yes — you retain ownership of both inputs and outputs. Ensure the contract explicitly states all inputs and outputs are your confidential information with full rights. This lets you use results freely in your business. OpenAI must treat outputs with the same care as any sensitive data.
Read about our GenAI Negotiation Service.
Read about our OpenAI Contract Negotiation Case Studies.
Need Help Securing Your OpenAI Contract?
Redress Compliance provides independent GenAI advisory — from data privacy risk assessment and contract redlining to compliance strategy and negotiation support.