By David Astle
Marmalade is tricky to make. Making a cryptic clue for marmalade is also sticky, unless you know how wordplay works. Could this new robot make a clue, I wondered. Did ChatGPT, the new bot launched by OpenAI, stand between me and my next payslip?
A linguistic omnivore, ChatGPT has wolfed whole terabytes of human expression, from novels to manuals, the A-Z of Wikipedia, Hansard, the Herald, the whole Dewey system including a cryptic crossword guide.
Paddington, a key advocate of marmalade.
Evidently, as when I quiz the machine about cryptic clues, its answer is prompt. “Cryptic clues often use puns, wordplay, and hidden words to guide the solver to the correct answer.”
A chill enters my bones, obliging me to test the humblebrag. Explain anagrams, I demand. “Sure! An anagram is a word or phrase formed by rearranging the letters of another word or phrase. Here’s an example: listen is an anagram of silent.”
Dumbstruck, I dare the upstart to compose a cryptic clue for marmalade, as good a word as any to know if your livelihood is cooked. Its response: “Preserve made from bitter citrus fruit (7)” My relief is palpable. Happily, the rookie is mired in quick crosswords, and I tell it so. That’s a definition only. Where’s your wordplay?
Perhaps my taunt delays its reply. “I apologise for that oversight. Here is a revised clue for marmalade: Spread for a tart fruit, perhaps orange (7).” Again, the same pitfall. My nemesis is cryptically clueless, stuck in its own orthodoxy.
Puzzle guru David Astle devised a test to measure the capacity of ChatGPT.Credit: Jo Gay
So I lend a hand, inviting the bot to scramble marmalade. It shoots back “a lad, marry”, which any Charlie can see is off-beam. Still, I encourage my apprentice – as that’s how the power dynamic has shifted. I ask its e-brain to combine anagram and definition, resulting in, “A lad may marry this sweet spread (7)”
Seven letters? With an inane narrative. And where’s the signpost, telling the solver to mix this dud letter batch? My vulnerability vanishes by degrees, to the point of showing this rookie how to make a genuine clue: “Jam created lame drama (9).
GPT is impressed: “That’s a nice cryptic clue for “marmalade”! The enumeration confirms the answer has 9 letters. The solver can then mix “lame” and “drama” to form the answer “marmalade.” Good job!”
Thanks, I key back, before glimpsing my extinction. A natural sponge, this protégé has a fatal gift for flattery. Your turn, I declare. Make your own marmalade clue, but again GPT struggles. “Dramatic dance transformed into sweet preserve (9)”
Try again, I say. That’s an indirect anagram, a genre no-no where solvers must intuit the mixable letters. This lesson leads to Draft 5: “Mix letters in LAME DRAMA for sweet preserve (9)” Mate, where’s your finesse? You’re giving me Coke + rum instead of a Cuba libre.
The bot is whip-smart, I can confirm. But never managed to pass the marmalade test, a check to gauge an AI tool’s capacity for elegant invention. My job feels safe, for now at least, though my intern did paint a new future when asked:
“AI language models could create crosswords. However, the quality would depend on the model’s sophistication, since a high level of artistry and nuance may be difficult for AI to replicate. Ultimately, it’s likely AI will play a role in the creation of crosswords, but human constructors will likely continue to play a key role as well.
Phew. The evidence is in. So far humans create the better marmalade. Breathe easy, dear mammal. But don’t for a minute think your tomorrow is preserved.
A cultural guide to going out and loving your city. Sign up to our Culture Fix newsletter here.
To read more from Spectrum, visit our page here.