ChatGPT’s AI can make crossword puzzles, and they’re pretty awful

One of the answers, it turns out, is making a good word game. Specifically, a crossword puzzle.

Now crosswords aren’t easy things to build—apart from the actual ‘crossing’ (interconnecting) of words, setters have to keep ‘product thinking’ in mind. These are player-centric questions like:

1. Are the answers guessable?
2. Are the clues crisp, intriguing and fun?
3. Is the trivia in the clues accurate?
4. Does the grid look good, and more importantly, is it full?
5. Is the grid symmetrical?

An AI could potentially get the ‘logic’-based parts right, i.e. making sure the words connect, and that all solutions are commonly-used English words. But points 2 to 5 rely entirely on a crossword setter’s ingenuity and good judgment, which can take months to build. When we tested out ChatGPT’s skills, this is exactly what it struggled with.

The prompts

In our first prompt, we gave ChatGPT the benefit of the doubt, requesting a simple crossword with the assumption that it knows how they’re made. This was its response:

ChatGPT’s first try at making a crossword. The clues are too long, and, well, too boring. The AI also ignored our 6-word rule, and shows incorrect word lengths (in brackets) as hints for the answers.

Since it’s a ‘conversational’ chatbot—you can talk back and forth—we poked it, asking why it didn’t follow instructions. This was the next reply:

ChatGPT’s second attempt at writing crossword clues: a little better, but still quite bland and ambiguous.

Progress! But there was no improvement on the nature of the clues. We then played ping-pong with our prompts for a while, asking the AI to improve the experience in 20+ different ways: ‘write short, funny clues’; ‘write punny, memorable clues’; ‘write clues where 50% have pop-culture references’; and, in a moment of desperation, ‘write clues that aren’t boring and go against your instincts’. 

None of this stuck, and after a while, the bot stopped responding entirely—a sign from the universe to hit pause, and just accept what we’d gotten so far. So here it is: ChatGPT’s chaotic crossword, which ignores factual accuracy, entertainment value, and a shot at any kind of creativity. But hey, at least it’s playable:

Spoiler alert: read this only after you’re done playing

If you had to rely entirely on the AI’s clues, you might have noticed that they were often vague, or had misinformation. 

For example, ChatGPT clued OSTRICH as a ‘large carnivorous flightless bird from NZ’. They’re actually omnivores from Africa. New Zealand once did have an ostrich-like bird — the moa — but they went extinct after humans arrived near the 13th century.

For 16-Down, BEACH, the clue mentioned a pool. Also misleading. We’re sure you can spot a lot more of these bad eggs, but this one truly took the cake: ‘small four-legged animal with long ears’ — why was this a clue for PIG, and not RABBIT?

We can’t predict what’s coming next. But for now, it’s clear that crossword setters: 1 and AI: 0

More from Express Puzzles & Games

Crosswords made 100% by nerdy humans:

What’s Cookin’ | Bheja Fry, the Indian Food Crossword

Gettin’ Filmy | Spotlight on Indian Cinema

Desi Halloween | Darna Mana Hai: A Horror Puzzle

The Big B Grid | 80 Years of Amitabh Bachchan

This story is part of Express Puzzles & Games, our new section dedicated to crossword puzzles, sudoku and quizzing in India. For game news, event alerts and more brainy fun, follow @iepuzzles on Instagram.