Knowledge base
Language models can take a lot of work off your hands, but not the step that matters. Where the line runs.
2026-07-20
The question is obvious: if a language model can read hundreds of statements, why not simply let it determine the themes?
The answer is not that it cannot. The answer is that you would then be measuring something other than you think.
In group decision-making it is not about the best possible arrangement. It is about the arrangement participants themselves recognise, because a decision only holds if the people carrying it out endorse the reasoning.
Let a model determine the themes and you get a tidy, defensible structure that comes from outside. Exactly what you were trying to avoid when you decided to involve the group.
The model can make a better arrangement. It cannot make a supported one.
There is plenty of work nobody will miss.
Merging duplicates. Among a hundred contributed ideas there are reliably twenty that say the same thing in other words. A model can lay those side by side, after which a person decides.
Rewriting into one form. Contributed statements vary widely in length and style. Bringing them into one form makes sorting fairer, because long statements otherwise draw disproportionate attention.
Naming themes, as a proposal. Once the clusters have been formed by participants, a model can propose a name based on the content. That saves time, provided the researcher can adjust the name and participants get it back.
Drafting report text. Describing findings on the basis of the calculated figures is writing work, not judgement work.
Everything that determines what the outcome is belongs with the participants and with the statistics. Everything that determines how the outcome becomes readable may be automated.
Concretely that means: the grouping comes from participants' sortings, the positions from multidimensional scaling, the clusters from cluster analysis on those positions. No language model touches those steps.
If there is AI anywhere in a project, the report should state where exactly, and what a human checked afterwards. Without that accounting a reader cannot tell which part of the conclusion comes from the group and which part from a model.
What we do and do not automate is set out on the accountability page.
Put your question to us and we will tell you honestly what is needed.
Get in touch