Will AI Ever Write a Cryptic Crossword?

Anyone who has ever asked ChatGPT to write a cryptic clue will be familiar with the errors it makes. From being unable to count letters, to some bizarre ideas about indicators and synonyms, it's frankly a bit of a shambles.

But to draw a conclusion from that about AI more widely would be naive. Generalist AI tools are similarly bad at solving - yet in the recent Times Championships, a specialist AI called Darmok solved the final puzzle faster than any of the five elite solvers on the stage (without falling into the trap that did for three of them). Apps like CrosswordGenius, meanwhile, do a pretty good job of solving and parsing individual clues.

I briefly spoke to Dr Siddhartha Chatterjee, the creator of Darmok, at last year's Championships (when it did well, but less well), and we talked about the difference between the challenge of solving compared to setting. The real challenge of setting - for humans as well as machines - is that a clue has to make sense on two levels: the surface and the cryptic. A solver might appreciate a good surface, but for solving it's not strictly necessary. Yet a clue with a sound cryptic reading but a nonsensical surface would be unlikely to get past an editor.

So there is a greater challenge to setting a clue than solving a clue. The fact that an AI can solve a clue, or a whole puzzle, does not itself mean it can write one.

However, it would be foolish of me to say no AI could ever create a sound cryptic clue. The capabilities of AI are improving, and while it might take another specialist system, with the right dataset and training I don't think there's any reason to suggest that writing cryptic clues is beyond AI.

Strictly, then, the answer to the question posed in the title is that in my view yes, AI will one day be able to write a cryptic clue, and by extension, a cryptic crossword. But, as an avid solver (as well as a setter), this misses the point. For me, the delight of a cryptic crossword is not just in the manual cracking of a set of clues to kill a bit of time. At the risk of opening a can of worms, if that achieved what I wanted from a puzzle, I would do a sudoku.

Drawing a parallel with other creative pursuits makes this clearer. I might be moved or entertained by a poem. But if I then discovered it had been created by an AI, it loses something: it is no longer a communication from another human. It is something more sterile. It's not that I gain nothing - it might still have sparked thoughts or brought up memories in me. But learning it was not created by a human makes it more hollow and less satisfying.

As a solver, I love the feeling of going head to head with a setter (even an anonymous one), and unravelling and admiring the clues they've written. As a setter, that's what I hope people who solve my puzzles enjoy. This human part of crossword-setting is not, I believe, something that technology will one day supersede; it is an essential part of why we do them.

This might not be true for everywhere cryptics appear. Puzzle books sold at airports, to the extent they include cryptics, will surely use AI-generated puzzles. Maybe even some publications will do so, although I sincerely hope that this is confined to places that would otherwise not have a cryptic at all. I don't, however, think there is much risk of them infiltrating the broadsheets, where people feel a greater connection with the setters and the outlet.

On a technological level, I have little doubt an AI will one day, perhaps one day soon, be able to create a cryptic crossword. Perhaps one already can. But I think human-created cryptic crosswords still have a bright future, for as long as there are people who enjoy solving them.