A falsifiable distractor is an incorrect option in a multiple-choice item that a candidate with partial understanding can plausibly select. The term names the operational target of distractor design: an option must be wrong, but wrong in a way that requires the target sub-skill to detect. Distractors that fall too far from plausible — random unrelated content, grammatical mismatches, obvious nonsense — degrade the item to a 3-choice or 2-choice format and reduce discrimination.
The framing comes out of the diagnostic-distractor literature. To build a falsifiable distractor, the writer must model the specific reasoning error a candidate at the target proficiency would make and turn that error into the option. The test then discriminates between the candidate who completes the inference and the candidate who completes most but not all of it.
Three properties define a falsifiable distractor:
The third property is what separates a falsifiable distractor from an ambiguous one. Ambiguous options reduce reliability and invite candidate appeals; falsifiable options sharpen the test.
Item analysis on falsifiable distractors shows a characteristic pattern: low-scoring candidates select the distractor at meaningful rates, high-scoring candidates do not, and the discrimination index is positive. A distractor that no one chooses is dead weight; one that high-scorers choose is mis-keyed or genuinely ambiguous. The 5Ps Distractor Typology (Sun, Yang & Liu 2026) gives a finer-grained vocabulary for the kinds of reasoning errors falsifiable distractors can target.
Falsifiability is the property AI distractor generators struggle most with. Out-of-the-box LLM generation produces distractors that are surface-coherent but reasoning-misaligned: they look plausible to the model and are obviously wrong to a competent candidate. Closing the gap requires explicit modelling of the target candidate's reasoning errors in the prompt, retrieval of grounded passage spans the distractor must trace back to, and post-hoc filtering against an item-analysis prior built from human-authored distractors.