How Do You Test a Thinking Machine?

AI applications are everywhere now. Software no longer just follows instructions—it listens, speaks, learns from data, adapts to new situations, and makes decisions on its own. In many cases, AI is the system. And yet, the way we test it hasn’t really caught up.

Most teams either aren’t testing AI at all—or they’re testing it like traditional software and hoping for the best. That’s risky. When systems behave probabilistically, change over time, and don’t always give the same answer twice, familiar pass/fail test cases stop working. This is where QA starts to feel uncomfortable—and where it also becomes most important.

In this talk, I’ll break down why testing AI is genuinely hard work, and why many of our existing tools and mental models fall short. We’ll look at the new challenges testers face when dealing with non-deterministic behavior, evolving models, opaque decision-making, and systems that learn in production.

From there, I’ll introduce a practical framework for testing AI applications end-to-end—from basic building blocks like prompts, APIs, and data retrieval, all the way up to system behavior, trust, and acceptance. We’ll walk through a real-world case study of evaluating a medical diagnostic AI system, where accuracy alone isn’t enough, and failure has real consequences.

Finally, I’ll show how to approach testing modern AI architectures, including LLMs, MCP servers, and RAG pipelines, with live examples using tools like RAGAS and DeepEval. The goal isn’t perfection—it’s confidence.

Because when software can think, testing has to do more than check boxes. It has to build trust.