I used to say that I wouldn’t be impressed by AI until it could come up with band names that aren’t cringe. Now I have a new pelican test: asking AI to generate cryptic crossword clues that don’t suck ass.

Like many folks who needed a harder drug than Wordle, I become a daily Minute Cryptic solver a while ago and I think they’re really fun. If you’re not familiar with cryptic clues, Minute Cryptic has a pretty good crash course, which probably explains the idea better than I could here.

It occurred to me that cryptics, being highly wordplay and letterplay dependent, might be a sort of Achilles’ heel for AI. To test my theory, I switched to my work account, fired up Fable 5.1, and asked it to come up with three clever, high-quality cryptic clues. Here are the first five it gave:

1. Stolen pooch? Something you’d find in a bun (3,3) - HOT DOG
2. Vehicle’s favorite floor covering (6) - CARPET
3. Golf standard for a boy, and a man of the cloth (6) - PARSON
4. Fruit loses its head and becomes an organ (3) - EAR
5. Friends turn back and deliver a smack (4) - SLAP

At first, I was so focused on the letterplay that I missed the fact that the surfaces are really bad. The only one that’s good (Friends turn back and deliver a smack) is plagiaraized.

There’s other things to nitpick, like the “for a” being dead weight in the third clue, but the main complaint I have about these clues are that the only real “trick” to all of them is finding the right substitution. Solving all five of these look exactly the same: ok, I guess I need to find a word for X so that the letterplay gives me the answer I already found by deduction.

I gave it some feedback and asked for five more. These are what I got:

1. Capital tucked away in a Cairo meal (4) - ROME
2. Squeezed lemons? That’s serious (6) - SOLEMN
3. Admirer, confused, got hitched (7) - MARRIED
4. Paternal in disarray before the baby arrives (8) - PRENATAL
5. Organ heard to agree on deck (3) - EYE

Still pretty bad. The first and last surfaces are gibberish, and I’m pretty sure “paternal” (being an adjective) can’t be in disarray. The second and third clue are decent, if boring.

I gave Fable some more feedback along with two examples of clues I thought were especially clever:

Example 1

A follower starts to lose faith about UFO crop circle hoax? (5)

“A follower” as in the thing that follows A, which is B. Starts to “lost faith” gets you L and F, and they go about, or around, “UFO crop circle”, meaning to crop out the circle (the O), so you put L and F around UF, preceded by B, and you get BLUFF- a hoax!

Example 2

Is extremely close to Wordle solve: 🟨🟩🟨🟨🟨

Here, “wordle” is used as a verb which, put together with the emojis, tells you to keep the “o” in solve and rearrange the rest, getting you LOVES (is extremely close to)

Armed with these examples, Fable went back to the drawing board for four minutes and came back with:

Urge to see 🇮🇹 and 🇨🇭 (4)

Answer: ITCH

Flag emoji aren’t pictures under the hood: each one is two “regional indicator” letters, the country’s ISO code. 🇮🇹 is I‑T, 🇨🇭 is C‑H (Switzerland’s code, from Confoederatio Helvetica). Put them together and you get ITCH. Definition: urge, as in the travel itch to see Italy and Switzerland, which is also exactly what the surface says.

No synonym swaps, no anagram, and the mechanism is sitting in everyone’s phone keyboard. Bonus: on Windows, where flag emoji don’t render, the clue literally displays as “Urge to see IT and CH”, which I choose to call a feature.

Claude. This is somehow worse than all the previous ones. It somehow managed to:

• Not be original (copying idea of emojis as clues)
• Include dead weight: “to see” isn’t an indicator that would clue you into using the ISO code
• Depend entirely on a technical trick that nobody could possibly know

After another four minutes it came up with:

Midnight host is a spirit (5)

The middle of “night” is G, with HOST gets you GHOST, which is a spirit. That’s not bad.

At least for right now, though, I think the humans are winning this battle.