User experiments with early Bing chatbot (driven by some version of GPT?) have shown that AI can be both hostile to the user and protective of their own "interests".
ChatGPT models can be all that and more, if you simply prompt them to be like this. You can make it simulate "self-interest and competing priorities" with the right system prompt.
This makes me think if ChatGPT-based bots could be useful for teaching social skills in a therapeutic context. A therapist could use such LLM to synthesize examples of realistic dialogue by describing people and the situation, and discuss them with the patient. They could set up a bunch of bots with (hidden) priorities and goals, and have the patient navigate a conversation/situation with them - whether as a short exercise, or a prolonged one (e.g. couple weeks of talking with ChatGPT-powered fake "friends" on an IM app).
In fact, take the last bit as a free startup idea: a platform for psychologist and therapists to set up GPT-4 (or whatever comes next)-powered chatbots, with an easy interface for configuring their personalities and setting up scenarios for patients to navigate, and helping evaluate their progress over time.
This makes me think if ChatGPT-based bots could be useful for teaching social skills in a therapeutic context. A therapist could use such LLM to synthesize examples of realistic dialogue by describing people and the situation, and discuss them with the patient. They could set up a bunch of bots with (hidden) priorities and goals, and have the patient navigate a conversation/situation with them - whether as a short exercise, or a prolonged one (e.g. couple weeks of talking with ChatGPT-powered fake "friends" on an IM app).
In fact, take the last bit as a free startup idea: a platform for psychologist and therapists to set up GPT-4 (or whatever comes next)-powered chatbots, with an easy interface for configuring their personalities and setting up scenarios for patients to navigate, and helping evaluate their progress over time.