Large Language Models: Ethics and Norms in the European Union

This paper investigates the ethical-legal issues underlying the implementation of Large Language Models, in the aftermath of the European Union Parliament's approval of the Artificial Intelligence Regulation. In very recent years, there has been a rapid evolution and multiform implementation of the so-called Artificial Intelligence. Notably, one significant development in AI is the creation of systems capable of performing tasks based on a set of established rules. Artificial Intelligence has recently undergone a significant revolution that has led to a radical paradigm shift, represented by the so-called Foundation Models, which have two main characteristics: emergence and homogenisation. On the other hand, Language Modelling (LM) is one of the main approaches to advance the linguistic intelligence of machines and thus a Foundation Model trained on textual data. The most recent development of this approach can be found in the so-called Large Language Models (LLM). These take the form of artificial neural networks pre-trained on huge datasets which, instead of being dedicated to a specific task, can perform a wide range of different functions at a level that is all the more accurate the greater the resources made available to them. The concept of language models is not new but has evolved with the progress of artificial intelligence over the decades. In general, LLMs are mainly trained in industrial settings where many important details of training (e.g. data collection and cleaning) are not disclosed. Thus, it is difficult to align LLMs with human values or preferences. Despite their capabilities, LLMs are also likely to produce toxic, fictitious or harmful content. Therefore, effective and efficient control approaches are required to eliminate the potential risk arising from the use of LLMs. For the purposes of the study conducted, it is intended to describe the development and outcomes with the aim to show not only the evolution of the responses offered by various LLM, but also and above all the possible ethical and legal implications that can be deduced from these outputs. Several models have been used for the implementation of the experiments: two GPT-2 models, Perplexity, ChatGPT (version 3.5), Llama. The language used for the implementation of the experiments was English, as it is structurally more neutral. In particular, with reference to a topic with a strong ethical and legal impact such as suicide, different answers were offered by the various models used. In conclusion, the essay offers reflections on the results of the experiment with reference to the EU Regulation on Artificial Intelligence. Artificial Intelligence represents for the jurist, not only European, a decisive challenge not only with regard to the categories of thought and formulation of operational languages, but above all with regard to the centrality of the human being. On the basis of these considerations, focusing on generative AI and in particular on Large Language Models, it is clear that many of the issues relating to copyright and the protection of personal data still remain unsolved, excluding those already regulated by the legislation on the point.