Tomas Lund

The background is e-learning. For years I built training for companies and kept finding the same thing: the content rarely failed. The route to it did. People dropped off at the login, at an unclear form, at a rule nobody could remember.

That turned into a research question: what actually stops an action? The answer became Behavioral Friction Theory — a model of how competing options decide whether something gets done. It turned out to hold for both people and language models, which is what makes it useful for the AI work today.

Research and practice now run the same way: what I measure in the lab ends up as tools for clients, and what trips clients up becomes new measurements.

I measure where there is something to measure. There is not always. A workflow, a text, a course — there the evidence is usually something else: that the problem can be described precisely enough that you recognise it yourself, and that the fix turns out cheaper than what you ordered. Our cases carry no measured effect, and each one says so.

The method is open. The measurement protocol is public, so you can run it again on your own material. The value is not in keeping the procedure secret; it is in running it properly and handing over the evidence.

Build an AI system, though, and the uncertainty can be measured. That is where the numbers actually are. Nobody can guarantee that a model never errs — but it can be calibrated: measure how often it gets it wrong, and make sure the decisions that really matter pass a human regardless of how confident the model sounds. You get the numbers, including the ones that do not look good. That is also the version that holds up in front of an auditor.

See the method and the articles → (in Danish)