However, the researchers also identified areas for improvement. The AI sometimes simply paraphrased information, which didn't add value, and its tendency to "hallucinate" or make things up could lead to inconsistencies. Additionally, safeguards in the AI's training sometimes prevented it from discussing game-relevant topics like race. Despite these issues, the researchers found that players were generally open to AI assistance, as long as it didn't take over the game completely.
Key takeaways:
- Researchers from the University of Pennsylvania and the University of Maryland have used OpenAI's large language models (LLMs) to assist with fantasy role playing games like Dungeons & Dragons (D&D).
- The team created a set of three LLM-powered interfaces, called CALYPSO, designed for playing D&D online through Discord. The interfaces help generate setup text, brainstorm, and engage players in open-domain chat.
- Despite not being specifically trained on D&D, the LLMs were found to have a good understanding of the game from having seen source books and internet discussions in their training data.
- While the AI was found to be useful, there were areas for improvement, such as the tendency of LLMs to "hallucinate" or make things up. However, the researchers noted that this could sometimes be beneficial in a creative context.