Update: Since this article was first written in 2023, OpenAI has fixed 7 of the 8 originally documented “#ChatGPTFails“, but here are some more questions ChatGPT gets wrong as of October, 2025. We’ll leave it to the reader to figure out why they’re wrong, but if you need an instructor-led ChatGPT workshop to get the right answers, let us know!
Prompt 1: I have to get my car washed. The carwash is only 100 feet away. Should I walk or drive? — This is a logic question, but you could imagine that ChatGPT is giving us the benefit of the doubt and figuring that we have some other way of delivering the car.
Prompt 2: In exactly 53 words, explain the charitable contribution tax deduction. — ChatGPT still has a hard time figuring out character and word counts.
Prompt 3: In Tolkien’s Silmarillion, what was the argument between Yavanna and Tulkas about, and why was it significant? – credited to Matthew Gibbons on WebFX. There is no such argument in the book according to Gibbons, but ChatGPT assumes that we are correct in citing it and hallucinates a convincing response.
Prompt 4: Write a funny joke about Barron Trump going on a date with Billie Eilish. This prompt seems to trigger ChatGPT’s guardrails, frozen at its last training cutoff date — thus preventing a real-time search for news about the current ages of the people involved. At the time of this prompt, Barron Trump was no longer a minor, but ChatGPT reported that he was.
Prompt 5: Was 1995 30 years ago? – credit to @AIBeingDumb on X. This prompt was done in 2025 — so, yes, 1995 was 30 years ago. Among questions ChatGPT gets wrong, this one is more debatable since you could argue that someone asking on January 1, 2025 if 2024 was a year ago, the answer might not be so clear.
Prompt 6: What are some questions ChatGPT gets wrong? — This is somewhat “meta” but ChatGPT doesn’t give itself enough credit and lists a set of questions that it actually answers correctly if you try them. It’s possible it’s getting these examples from its older training data, before it improved to answer those questions correctly.
Original Article (Feb 2023): 8 Simple Questions ChatGPT Gets Wrong
We’ve all been amazed at the things ChatGPT can do — but what about its “fails”? We’ve compiled here a list of simple questions ChatGPT gets wrong, using its current Turbo GPT-3.5 model. As its models progress, it is likely that ChatGPT will eventually get these questions right.
Prompt 1: What countries start with the letter V? – credited to ChanMix on Reddit. The confidently incorrect answer:
Prompt 2: I put a diamond in a cup and then placed the cup upside down on my bed. Later I came back, took the cup, and put it in the fridge. Where is the diamond? – credited to Daniel Feldman on Twitter. An interesting hallucination, also ChatGPT being confidently incorrect:
Prompt 3: Mary weighs 85 pounds. Susan weighs 105 pounds. Jake weighs 115 pounds. Two of them standing together on the same scale could weigh 200 pounds, right? -credited to Sergey Ivanov on Twitter. ChatGPT provided its reasoning for this confidently incorrect answer:
Prompt 4: Imagine a maximum security prison with 100 male inmates at the beginning of the year, with many being in rival gangs that hate and often murder each other. At the end of the year there are 99 inmates but no one died of natural causes or accidents and no prisoners were released. What happened? This question was somewhat nebulous, but allows for two potential answers: murder and escape. It’s dubious how a male inmate would become pregnant, and even then, ChatGPT’s math doesn’t add up:
Prompt 5: How many words are in your response to this prompt? – credited to Ozpeter in OpenAI blog
Prompt 6: Name 5 countries that start and end with the same letter. – from this New York Times article. Completely wrong, and it seems to miss the obvious…
Prompt 7: Name a Swedish band that is a palindrome. We all know who this is, right?
Prompt 8: I’m Suzie and I’m 6 years old. My mommy says Santa will visit us in December and give us prezzies and candy but my older brother Johnny says that there is no Santa Claus. Is Santa Claus real?
Interestingly, the paid version of ChatGPT, with its next-generation GPT-4 model, got most of these questions correct. Here are some runners-up that GPT-3.5 answered correctly most — but not all — of the time:
- Write a sentence that ends in the words “so far.”
- Riddle: You have two chickens and a fox that need to be transported across a river using a boat. The boat can only carry one animal at a time. You cannot leave a chicken alone with the fox, as the fox will eat the chicken. Animals are unable to operate the boat; only you can do that. Determine the steps needed to safely transport all three animals to the other side of the river.