Enhancing Transparency in Large Language Models to Meet EU AI Act Requirements

Transparency is a quintessential requirement in artificial intelligence systems, featuring prominently under the EU AI Act, but it raises immense challenges for LLM developers due to the intrinsic black-box nature of LLMs.The paper discusses some techniques and methodologies that could assist in increasing the transparency of LLMs and achieving conformity with the exacting standards prescribed under the EU AI Act.In this paper, the authors go further in developing the demands that the Act places on transparency, documentation, and human oversight.Moreover, there are inherent issues associated with interpreting LLMs, particularly the trade-offs between performance and interpretability, compounded by intricate systems.Additionally, this paper gives an overview of the current state of the art in explanation methods for LLM outputs, ranging from attention mechanisms over model distillation to post hoc explanation tools.Some guidelines that developers should follow include the following: performance vs. interpretability trade-offs, full documentation, and user-centered design principles.Through illustrative case study, we demonstrate how these practices can be implemented in GPT-based models to meet regulatory demands.By bridging the gap between advanced AI models and the EU's regulatory framework, this paper contributes to the development of AI systems that are not only powerful but also transparent, fostering greater trust and acceptance among users and regulators.