Why "feels good" isn't enough
Asking a chatbot a few test questions and finding the answers good isn't a reliable test – our impression is easily fooled, especially by confident-sounding but fundamentally wrong answers (see module 4, hallucination). A real check – in industry jargon: an evaluation, or "eval" for short – follows a fixed process.
Four steps of a real evaluation
1. Define the success criterion upfront
What exactly does "correct" mean for this task? Without a clearly defined criterion, anything can be reinterpreted as a success afterward.
2. Test with real, representative examples
Not just with the simplest cases, but with a mix that reflects your actual requests – including tricky edge cases.
3. Compare against a baseline
How well would a human, or a simpler existing solution, do on the same examples? Without a comparison value, a score like "85% correct" is hard to interpret.
4. Keep measuring
A model update at the vendor's end, or more requests in real-world use, can change quality – a one-off check at the start of a project isn't enough for the long run.
Why this matters for you as a decision-maker
Anyone who only asks "was that good?" after the test phase gets an opinion. Anyone who asks about the success criterion, sample size, comparison value, and ongoing measurement gets a solid basis for an investment decision (see also module 9, cost & ROI).