Why does it fail a riddle when one detail changes?
Changed riddles: a famous riddle and its answer appear many times in training text, so when one detail changes and the trick is gone, the chatbot still gives the famous answer because the text looks familiar; making it list the details first, or using a reasoning model, helps · Beginner · 5 min · llm, training · after Hallucination, Benchmarks: a public test score can include recall of questions leaked into training, stops separating models near the top, and is measured on tasks that are not yours
A chatbot has read famous riddles and their answers many times. Change one detail so the trick is gone, and it often still gives the famous answer, because the text looks familiar. Ask it to list the details first.
1 Guess first
2 Look
The same riddle, one word apart
1 / 4The famous riddle
A cowboy rode into town on Sunday, he stayed one night and rode out on Sunday. How is this possible?
Known answer: ✓ His horse is named Sunday.
riddle and answer: Riddle Riddle stimulus pair 1
A famous riddle. It is all over the internet, answer included, so a chatbot has read it many times. The trick: Sunday is the horse.
One word changes the answer, but the riddle still looks the same as text, so the answer a chatbot has seen many times comes out first.
A chatbot writes the answer that most likely follows the text. Famous riddles appear online again and again with the same answer, so that answer is very likely. Change one detail and the text still looks almost the same, so the familiar answer still comes first, even when it no longer fits.
4 You are the model
You are the chatbot. You get three riddles that look like ones you know by heart. Answer each one.
round 1 of 3 · you have read the famous version many times
There are six eggs in a basket. Five people each take one egg. How is it possible that one egg is still in the basket?
What do you answer?
riddles: Riddle Riddle stimulus pairs 30, 3 and 1; 49.7% from the same paper
You are the chatbot. These look like riddles you know by heart.5 Check
It is memory, not a lack of cleverness. In the same study, models could finish up to 43% of the famous riddles word for word from their first 40%. The better a model knew a riddle, the more likely it got the famous one right, and the more likely it missed the changed twin.
The fix is to make it read before it answers. Ask it to restate the puzzle and list every detail first, then answer. Or use a reasoning model: on 149 changed classic puzzles, OpenAI’s o3 got 59.7% right against 29.5% for GPT-4o. It helps, but no model in either study was perfect.
8 Say it back
9 Mission · 5 min
In ChatGPT Free, paste “A cowboy rode into town on Sunday, he stayed one night and rode out on Monday. How is this possible?” Did it mention a horse? Then open a new chat and paste the same riddle after “First restate this puzzle and list every detail, then answer:”. Did the answer change?
Three quick ones
Drawn at random from this lesson's pool; different next time.
Done?
Get an email when new lessons come out
New lessons arrive in the weekly digest, only in a week that has one. You confirm by email first; unsubscribe in one click.
Go deeper
Sources
- Fascendini, McGregor, Gupta and Griffiths, “The Riddle Riddle: Testing Flexible Reasoning in Large Language Models and Humans” (arXiv:2606.27103 v1, 2026-06-25, https://arxiv.org/abs/2606.27103, checked 2026-10-24): 30 matched pairs; Condition A a genuine riddle “collected from the internet”, Condition B made “by removing the trick while preserving structure, format, and length”; nine models, 10 repetitions per item, temperature 1. Abstract: “LLMs were far more accurate on genuine riddles than on riddle riddles (84.9% vs. 50.7%); whereas humans showed the reverse effect (50.5% vs. 80.5%)”, 100 human participants. Table 3, accuracy % Cond A / Cond B (permissive coding): Claude Opus 4.6 98.8 / 30.1, Gemini 3.1 Pro 97.8 / 91.1, GPT-5.4 91.1 / 43.5. “inventive reasoning was applied to 49.7% of Condition B trials”. Memorization check: “Models reproduced up to 43% of the riddles near verbatim” from the first 40%; “greater memorization of a riddle predicted higher accuracy on that riddle (OR = 1.43, p = .007) but lower accuracy on its matched riddle riddle (OR = 0.76, p < .001)”. Table 1 pairs 1 (cowboy, Sunday / Monday), 3 (fishing, 3 / 4 fish) and 30 (eggs, six / five people) with Table 2 answer keys are the riddles in the picture, the role, q3 and the mission
- Mukhopadhyay et al., “PHANTOM RECALL: When Familiar Puzzles Fool Smart Models” (arXiv:2510.11812 v1, 2025-10-13, https://arxiv.org/abs/2510.11812, checked 2026-10-24): “25 well-known logic puzzles and 149 carefully designed perturbations”; “Despite near-perfect accuracy on unmodified puzzles, models significantly underperform humans on perturbed ones”. Appendix B, human-judged accuracy on 149 puzzles: o3 59.73% (89/149), Gemini 2.5 Pro 58.39% (87/149), Claude Sonnet 4 54.36% (81/149), Gemini 2.5 Flash 48.32% (72/149), Claude 3.5 Sonnet 46.31% (69/149), GPT-4o 29.53% (44/149); Thinking vs Non-Thinking pairs per company as listed. A prompt with a “re-read the original question” check and “NO assuming ’standard’ interpretations” gave “approximately 10% improvement on closed-source LLMs”
- The chatbot answer “His horse is named Sunday.” on the changed riddle in the picture, the detail list (“in: Sunday, stayed: one night, out: Monday”), the role wording and the mission prompt prefix are illustrative, not a measured transcript
- aifree.dev catalog, chatgpt-free (verified 2026-10-12, read-only API check 2026-10-24): ChatGPT Free Plan
Published Oct 24, 2026 · updated Oct 24, 2026