· privacy
GDPR and LLM workloads
A GDPR review of an LLM workload starts with the processing, not the hosting label. Trace which personal data enters prompts, files, retrieved passages, feedback, logs and backups; why each use is needed; who can access it; and when it is removed. A private deployment may help with a security requirement, but it does not by itself establish a lawful basis, appropriate retention or clear responsibilities.
Map the parties and the purpose
Describe the purposes and actual decisions made by each party. The GDPR defines a controller by who determines the purposes and means of processing, and a processor by processing personal data on behalf of a controller. Contract labels should be checked against the real arrangement. Where a vendor processes data for the customer, Article 28 sets requirements for the contract and processing instructions. Read the EDPB controller and processor guidelines alongside the GDPR text.
Use this workload checklist
- Purpose and minimization: document the purpose and lawful basis, tell people what is happening, and avoid sending fields the task does not need.
- Retention and rights: set deletion rules for source files, prompts, embeddings, outputs, feedback and diagnostic logs; determine how access, correction, objection or erasure requests are handled.
- Processor and transfer terms: identify subprocessors, processing locations, support access, use for model improvement, retention and any international-transfer mechanism that applies.
- Security: select technical and organizational controls appropriate to the risk, then test identity, access, encryption, logging, incident handling and recovery. Article 32 addresses security measures appropriate to risk.
- Impact assessment: screen the actual processing for likely high risk. Article 35 requires a data protection impact assessment in specified high-risk circumstances; AI use alone is not an automatic conclusion either way.
- Model changes: keep user data used for inference distinct from any training or fine-tuning dataset, with separate purpose, permission, access and deletion decisions.
Hypothetical example: a company proposes using employee messages to draft HR answers. Before a pilot, it should identify whether sensitive or special-category data may appear, who decides the purpose, which HR documents can be retrieved, how employee questions are retained, who reviews drafts and what legal escalation path exists. A DPIA decision must follow the actual risk and context, not a generic checklist. This example is illustrative and not legal advice.
Record unresolved questions for the organization’s privacy lead or counsel before production data is used. For implementation boundaries, see data residency choices and vendor subprocessor review. Review Private AI Services controls or discuss the workload and data lifecycle.