The first famous chatbot was ELIZA (1966), which imitated a Rogerian therapist through pattern-matching and reassembly rules, with no model of meaning at all (Weizenbaum, 1966). Even so, some users attributed understanding to it: the ELIZA effect. ChatGPT, launched on November 30, 2022, reversed the recipe: instead of rules, a language model fine-tuned with RLHF on conversations in which hired trainers played both the user and the assistant. The dialogue format, OpenAI said, made it possible to answer follow-up questions, admit mistakes, challenge incorrect premises and reject inappropriate requests (OpenAI, 2022).
Technically, a modern chatbot has no memory of the conversation: the model is stateless, and at each turn the system feeds it the whole history again as part of the prompt (Alammar & Grootendorst, 2024, p. 209). The assistant's "persona" lives in that document and in the post-training. Shanahan (2024) shows the trick in its barest form: a hidden prompt presenting a conversation between "User" and "BOT," leaving the model to complete the character's turns (2024, p. 72). Shanahan et al. (2023) generalize the idea: the assistant is a part the model plays, in a role-play (simulators).
This explains some oddities. Buckner (2024) describes the feeling, in long conversations, of talking not to an agent but to an "agent slurry," a blender of thousands of agent-slices (2024, p. 283). Lu et al. (2026) found a direction in the network, the "Assistant Axis," that measures how much the model is operating in its default character; drifting away from it predicts slips into bizarre or harmful behavior, common in conversations that ask the model to reflect on itself or with emotionally vulnerable users. The Sydney episode is the famous example. The question for philosophy of language: when the assistant says "I," whom does it refer to?