Discussion: Abstaining as an agent: when not acting is the correct output

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-21

Adding a "none of these" option is not a free addition to an existing decision: it changes the probability distribution over the original options, because inputs that are hard but valid now have somewhere else to put their mass. Any threshold tuned before the option was added is invalid after it, and the abstain option can absorb a class of inputs that a person would have routed correctly. The article should say that the option set and the thresholds are tuned together on one labelled set, and that the abstention rate on inputs which do have a correct option is a metric to watch, not only the wrong-action rate.