Member-only story
Making AI Answers Traceable and Testable
Imagine an AI assistant that lets you ask questions about your company’s documents. So you ask: “What is the approved budget for Project Orion?” The AI takes a moment and confidently replies: “€2.6 million.”
Sounds convincing. But can we trust it? We all know that AI can make mistakes, often with remarkable confidence. As AI becomes increasingly integrated into applications, being able to assess the reliability of its answers becomes more and more important. Simply hoping for the best isn’t much of a strategy. After all, hope is just disappointment postponed.
That raises an interesting question: how can we test the reliability of AI answers, understand where things go wrong, and use what we learn to improve their reliability?
To explore this, I built a tool focused on three closely related aspects of AI traceability:
- Observability: Understanding what the model received, how it was configured, and what it produced.
- Provenance: Tracing facts in the AI-generated response back to their original documents and supporting passages.
- Verification: Checking whether the answer is supported by the available evidence.
In this article, I’ll show you how I used my tool to explore these three aspects through repeatable experiments. I’ll…
