Test your chatbot before your users do

Language models phrase things fluently, even when the content is wrong. They respond differently to every wording, and users sometimes deliberately try to mislead them. What convinces in a demo says little about real-world behaviour.

The same four quality axes as for any AI system, tailored to language models:

• The correctness of answers to the questions your users actually ask.
• Robustness against misleading and manipulative input.
• Consistency across rephrasing, repetition and long conversations.
• The quality and coverage of the data the chatbot relies on.

In red teaming, we deliberately try to make the system go off the rails, the way a malicious or unpredictable user would. Think of attempts to bypass instructions or extract information that must not be shared. We record where the system holds up and where it does not.

Ideally before launch. But a chatbot that is already live and has never been tested thoroughly can also be validated, even if another party built it.

A repeatable dossier with the test scenarios, the measurements, the weak spots found along with the scenarios to reproduce them, the risks and a prioritised list of improvement actions.