Why Did AI Make That Decision? We Built a Tool to Find Out

Understanding AI Decision-Making with WhiteBox

I've noticed a recurring theme among companies seeking AI solutions: they often emphasize that their need goes beyond "just another chatbot." What they truly want is an AI that is intelligent—and by intelligent, they mean one that can explain why a decision was made. This issue is frequently referred to as the "black box AI" problem.

The concept of White Box AI is well established. It's about understanding why a model chooses a specific response. At its core, it's all about log probabilities (logprobs), which help models determine the most likely next token in a sequence. However, not all models disclose logprobs, complicating the evaluation of truthfulness across different AI systems.

Introducing WhiteBox: A Solution to AI Transparency

That's where my project, WhiteBox, comes in. I designed it to tackle this issue head-on. WhiteBox employs an empirical approach: we host multiple AI models and pass the same prompt to all of them. If these models agree on an answer, we can trust the response is reliable. This method uses consensus to enhance AI observability and truthfulness.

WhiteBox is already in use. For instance, DayBy leverages it to moderate AI-generated posts, ensuring that users don't misuse AI's content creation capabilities. It assesses the intent behind messages and maintains a balanced environment.

Human-in-the-Loop: Ensuring Safety

There is also a safety mechanism known as Human-in-the-Loop (HITL). When there's no consensus among the models—say, you send a prompt to 15 different models and they can't agree—a human steps in to make the final judgment. This acts as a crucial safety net.

Currently, WhiteBox has a technical flaw: its interface resembles a git terminal, which might be intimidating for some users. I'm actively working on a redesign to make it more user-friendly, but I encourage you to check it out at whiteboxhq.ai.

The Bigger Picture

The importance of understanding AI decisions extends beyond individual applications. It's essential for identifying which models are more accurate for specific topics. This can be pivotal in fields like medicine, education, music, and content moderation—all areas we've explored with WhiteBox HQ.

I'm eager for more people to use WhiteBox so I can refine my models further. I'm also focusing on encrypting the data we send to external models to ensure user privacy, while simultaneously developing our own models.

Feel free to give WhiteBox a try at whiteboxhq.ai. I'll keep you informed about our progress and updates.