This chapter will discuss why internet-based AI technologies are pathways to artificial rational agents, with the example of the large language models (LLMs) represented by the GPT family. It will introduce the architecture and training mechanism of the GPT family and illustrate how web resources provide training conditions for them. It will argue that general representations pursued by LLMs should be converted into universal expressiveness. By explicating reasons, LLMs and internet-based AI agents in general could track sources and consequences of information. Two proposals for enhancing the expressiveness of LLMs will be presented in this chapter. On the one hand, involving more diverse users, especially underrepresented groups in digital life, can enrich the triadic interactions for acquiring intersubjective knowledge during the tuning phase after pre-training of LLMs. On the other hand, human users can be more considerate and responsible in their use of the internet and their interactions with AI agents to generate less stereotypical and more traceable data relations. In this way, linguistically competent AI agents will be able to express and reveal the implicit norms of social practices, taking part in the norm-setting of social practices as rational agents.