In this meeting, the State Changers discussed the challenges around prompt induction with large language models, with particular emphasis on avoiding harmful outcomes from AI interactions, such as the controversial Microsoft chatbot Tay that was trained by users to produce racist comments.
The group acknowledged the unpredictable nature of AI responses and highlighted methods of reducing risks. These include gating service usage through a Paywall, which also improves interaction quality; introducing moderation through APIs from OpenAI or other comment moderation methods; and clarifying intent of AI usage to users (for example, it is used as an advisor providing suggestions instead of generating automatic responses).
In the context of a specific AI-driven text generation tool, potential dangers include injecting harmful or inappropriate content into the service. To mitigate this, the user must manually copy and paste the AI-generated text to their own document, increasing their responsibility for the content.
Additionally, the group discussed the impact of negative press about AI, including its misuse for harmful purposes and the risk this poses to a firm's reputation. To avoid bad press, transparency with users and staying ahead of emerging AI safety concerns were suggested.
One participant mentioned his positive experience using ChatGPT, highlighting it as a good resource for learning how to generate appropriate prompts. He indicated that the absence of judgement from the AI was useful, as it avoided the shame and discomfort associated with potentially inappropriate remarks.
(Source: Office Hours 5/26 )