Study finds AI models cannot fully forget personal data
A study by Universitat Rovira i Virgili found that AI language models including ChatGPT, Mixtral, Bard and Copilot cannot fully forget personal data once learned, conflicting with GDPR's right to be forgotten. The researchers concluded no method guarantees complete data deletion except retraining from scratch.
What happened
A study by Universitat Rovira i Virgili analyzed unlearning techniques in large language models including ChatGPT, Mixtral, Bard and Copilot that aim to remove personal, incorrect or discriminatory information from systems.
The researchers concluded that no formula exists to ensure complete forgetting of data except by retraining the model from scratch without the information to be removed.
Reported impact
The findings create a conflict with the right to be forgotten under the EU General Data Protection Regulation, which requires deletion of personal data when requested.
Large companies training AI models with enormous volumes of information find it difficult to identify and selectively eliminate certain data.
Index analysis
The researchers identified two approaches to unlearning: complete retraining of the model, which involves high computational cost, and more efficient partial removal of specific knowledge without rebuilding the system, though neither offers full guarantees.
Researchers noted that how information is stored in models is not precisely known, even by the developers themselves.
Governance controls
The study concludes that new methods must be designed to train AI models with unlearning in mind from the outset, such as fragmenting data so specific parts of knowledge can be deleted without compromising the overall model or its functionality.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.