AI
Whether an AI assistant actually helps with a crossword depends heavily on which kind of crossword it is. Standard American-style clues (definitions, synonyms, fill-in-the-blank pop culture references) are a reasonable fit — the assistant just needs the letter count and any letters you already have filled in from crossing words. Cryptic-style clues, common in British and some syndicated puzzles, are a different story: they rely on wordplay — anagrams, hidden words, homophones, double meanings — and that's a documented weak spot for current models, not a solved problem.
Getting useful help without ruining the puzzle
1. Always include the letter count and any known letters. "5-letter word for capital of France" is a much narrower search than the bare clue text — treat the letter pattern like a search filter, not an afterthought.
2. For standard clues, ask for multiple candidates, not one answer. Crossword clues are often intentionally ambiguous until crossing letters narrow them down — a list of 3-5 candidates that fit the pattern is more useful than a single confident guess.
3. For cryptic clues, don't trust the answer without checking the wordplay. Ask the assistant to explain which type of wordplay it thinks is being used (anagram, hidden word, container, etc.) — if the explanation doesn't actually hold up when you check it, the answer is probably wrong too.
# Prompt that gives the assistant enough constraints to be useful
Clue: "Feline sound heard at midnight" (4 letters)
Known letters so far: M _ _ W
List up to 5 candidate answers that fit this exact
letter pattern. For each, briefly say why it matches
the clue. Do not guess an answer that doesn't fit
the letter pattern I gave you.| Clue Type | How Reliable Is AI Help |
|---|---|
| Standard definition/synonym clue, with letter pattern given | Reasonably reliable as a source of candidates to check against the grid |
| Cryptic clue relying on anagrams, hidden words, or wordplay | Weak — research found accuracy far below expert human solvers |
A 2024 research paper testing GPT, LLaMA, and Gemma models against cryptic crossword clues found the best-performing model solved only about 1 in 9 clues correctly, compared to expert human solvers who get essentially all of them right. The paper's error analysis found the models often couldn't even correctly identify which type of wordplay a clue was using, and in some cases produced explanations that didn't actually hold up (claiming a word was an anagram of letters it didn't contain). That's a real, current limitation — not a reason to avoid AI help entirely, but a reason to treat it as a candidate-generator for straightforward clues and a much less trustworthy guide for cryptic wordplay.
Practical Challenge
Next time you're stuck on a puzzle, give an assistant the clue plus the letter count and known letters, ask for multiple candidates, and check each one against the grid before filling it in.
Concept Check
Sources & Further Reading
- What Makes Cryptic Crosswords Challenging for LLMs? (arXiv, 2024) — the research cited above testing GPT, LLaMA, and Gemma models against cryptic crossword clues, including the accuracy gap versus expert human solvers and the wordplay-misidentification error analysis.