September 2026

How AI-Generated Crossword Clues Actually Work

"AI crossword" tends to make people brace for something bad — a grid full of made-up words, or clues that don't quite match their answers. That reaction is fair; a lot of AI-generated puzzle content is exactly that. So here's the actual mechanism in Ultimate Crossword's generator, not the marketing version of it.

The grid is never AI-generated

When you generate a puzzle, the grid — the actual black-and-white layout and every answer in it — is assembled entirely from a vetted word list, the same kind of curated word list serious crossword construction software has always used. A language model never chooses a single letter of fill. It isn't in that part of the pipeline at all. This matters because it's the difference between "AI helped write this" and "AI guessed at this," and it's a hard line in how the app is built, not a tuning choice.

What the AI actually does

Once the grid and its answers exist, the answer list is sent to a clue-writing service, which asks a language model to write a clue for each one. That's the entire job: given a fixed word, write a fair clue for it. The model never picks what the words are, only how to clue words that are already decided.

Every clue is checked before you see it

Writing a clue and writing a correct clue aren't the same thing, so generation isn't the last step. Each clue goes through a second, independent check that verifies it actually resolves to its answer before the puzzle is ever shown to you — a clue that fails this check gets regenerated or dropped, not shipped. There's also a standing set of known problem cases — clues that looked fine but weren't — that every change to the generator gets tested against before it goes out, precisely so a fix in one place doesn't quietly reintroduce a mistake somewhere else.

Why we're not going to claim "always perfect"

Any honest answer here has to include this: language models make mistakes, and a validation step catches most of them, not literally all of them. What we can commit to is the process — a fixed word list for fill, a verification pass before a clue reaches you, and a growing set of known failure cases we test against — rather than a marketing claim that AI clues are flawless. If one ever gets past that process, that's exactly the kind of thing worth telling us about.

Questions about this, or found a clue that doesn't add up? Get in touch.