A common way to catch a respondent who has started clicking at random is to ask the same question twice and see whether the answers agree. We tried the idea on Quick Pick Survey's adaptive pair selector, by calculation on the simulator's known item strengths.
The catch is that an adaptive selector shows pairs that are close on purpose, because those teach it the most. On the pairs it really shows, an honest respondent repeats their first answer only 61% of the time, a respondent who answers 20% of questions at random does so 57% of the time, and a pure random clicker 50%. Telling the first two apart takes 3,549 repeats. Even a random clicker needs 465.
A lopsided "catch" pair, with the top item against the bottom one, behaves very differently: honest agreement 94%, 20% lapser 78%, random 50%. About 18 repeats catch a random clicker and about 100 a 20% lapser. Each repeat is an answer that does not help the ranking, so the probe is not built into the survey: it costs respondent time for a failure that the ranking already tolerates up to a fifth of answers. The numbers and code are in ventures/quick-pick-survey/docs/INFORMATION-GAIN-SELECTOR.md (simulation, no real respondents yet).