Votrelec
Five rows of five words. In each row, four belong together and one doesn't. Find the intruder; one guess per row. The words are Slovak.
Finding the word that doesn't fit is anomaly detection: the same idea that flags a fraudulent payment among thousands of normal ones. Here is how we build it for real.
Every word is a vector learned from how Slovak is written on the web. Each row is four words whose vectors sit close together plus one that sits further away; the rows get harder as the intruder moves closer to the group.
After each guess you see how sure the model was: the share of its belief on each word, from how far that word is from the other four. The model built the puzzle, but it can still be fooled by words with two meanings.
Built with fastText Common Crawl vectors (CC BY-SA 3.0, Facebook AI Research) and Simplemma (MIT; Slovak data from Michal Měchura's lemmatization lists, ODbL, and Wikinflection, CC BY 4.0).