AI Expert Explains Chatbot Use, Reliability, and Number of Bots

Chatbot Usage and Personal Experience

• 💡 AI chatbots are found useful for work and research, but require frequent fact-checking.
• ⏱️ Usage varies, with some spending only a few minutes intermittently for research, while others dedicate around 20 minutes a day.
• 🤖 Personal use often focuses on research rather than emotional support or sharing personal information.

Simulation and Model Behavior

• 🔬 Simulations are used to test chatbot responses because they are probabilistic models and don't always provide the same answer.
• ⚠️ Researchers run simulations millions of times to understand model behavior and test prompts, noting that chatbots don't always follow instructions, sometimes requiring repeated attempts.
• 🧑‍💻 A dedicated team of researchers and volunteers (around 20 people) monitors and codes chatbot outputs, though this process is time-consuming.

Reliability and Oversight Concerns

• 📚 Dr. King notes that chatbots are often unreliable for research, frequently fabricating citations and titles.
• 🧐 The concept of 'LLM as judge' is highlighted as ironic and potentially problematic, suggesting human oversight is crucial.
• 🚨 While real-time monitoring of responses to sensitive queries (like suicidal ideation) isn't constant, some regions like California have implemented bills to automatically forward such responses to crisis lines.

The Question of Chatbot Quantity

• ❓ A key question is how many chatbots exist, with experts noting major foundation model developers and numerous smaller, purpose-specific ones.
• 📊 There is no clear current number, but the discussion points to the need for inventories of foundation models and potentially chatbots, especially concerning existential risk.
• 🏛️ Policy discussions have previously considered inventories for foundation models, suggesting a similar approach could be warranted for cataloging chatbots.