The right to rectification requires inaccurate personal data to be corrected under applicable privacy law. For generative AI, the error may appear not only in a database but in an answer that makes a plausible, damaging claim about a person. A disclaimer that an answer may be inaccurate does not by itself remove the organization’s responsibility. section 5.4.1 section 5.0
A correction request is an operational problem as well as a legal one: personal data may be copied into prompt logs, stored outputs, retrieval documents, embeddings, or model parameters. The response depends on where the system stores or has learned the information. section 5.0 section 5.5.4
Trace the request through the deployment model:
- For a SaaS service, check vendor capabilities, retention settings, audit logs, and whether submitted data can be purged; the organization may not control the model or all its logs directly. section 5.5.4
- For an API integration, search and correct organizational prompt logs, stored outputs, retrieval documents, and embeddings; establish what the external provider receives and retains. section 5.5.4
- For an in-house model, plan for provenance records, deletion-aware pipelines, derivative artifacts, and documented limits on unlearning. section 5.5.4
Correcting a database record does not necessarily remove its influence from trained model parameters. Retraining from scratch can be impractical, while machine-unlearning methods remain unreliable at GPT-class scale and may leave traces. Preventing sensitive data from entering training is therefore safer than relying on later correction alone. section 5.5.1
Plausible false allegations
A chatbot falsely associated a real person with a serious crime while mixing the allegation with accurate personal details. The case illustrates why correction and accuracy controls matter even when a system labels its answers as potentially inaccurate. section 5.4.1
