A Gizmodo piece explores whether large language models can handle one of the trickiest forms of wordplay: the cryptic crossword. The article says the idea came after news that Firefox would add a daily AI-powered crossword to its new tab page, prompting a tougher test with an especially difficult cryptic puzzle.
The reported result is good news for human solvers. According to the article’s framing, the language models did not match people on this kind of crossword, where success depends on layered clues, misdirection, and an ability to spot multiple meanings at once. In that setting, humans still appear to have a clear edge.
That outcome fits a broader pattern in AI discussions. LLMs are often strong at producing fluent text and recognizing common patterns, but cryptic crosswords demand very precise reasoning about puns, hidden structures, and unusual clue logic. A model may sound confident while still missing the actual mechanism behind a clue.
The story ultimately presents cryptic crosswords as a useful stress test for AI. While machine-generated puzzle features are spreading into mainstream products, this experiment suggests that difficult human-crafted wordplay remains a challenge for current models. At least in this corner of language, the article argues, people are still ahead.
