The consequence of advancements in artificial intelligence (AI), particularly in foundational large language models (LLMs), poses intricate legal issues in regards to the General Data Privacy Regulation (GDPR). Notably, the debate surrounding compliances with the “right to be forgotten,” or, in other terms, the right of erasure. LLMs display exceptional skills in predicting sentence endings, solving mathematical problems, writing code, and summarizing complex concepts. However, the ability to “forget” – to erase certain parts of their training set – is not equally impressive. The structures of these models don’t provide their developers viable alternatives to comply with the GDPR’s right to erasure without necessitating the deletion of entire models.
The issue gains complexity when we consider that LLMs are typically trained on publicly available personal data. The implication is that “forgetting,” in this context, isn’t just about portions of the training set, but entire models that may need to be retrained from scratch. As AI development continues to proliferate, the legal framework, especially regulations like GDPR, may have to grapple with evolving and potentially challenging considerations.
For additional context and discussion on this topic, you may refer to the full discussion here.