AI가 추측할 때 '컵케이크' 프롬프트를 사용하여 파악합니다.

Here at Tom’s Guide our expert editors are committed to bringing you the best news, reviews and guides to help you stay informed and ahead of the curve! You are now subscribed Your newsletter sign-up was successful Want to add more newsletters? Join the club Get full access to premium articles, exclusive features and a growing list of member rewards. AI chatbots are incredibly useful, but they still have a habit of guessing when they don’t know the answer. Sometimes those guesses sound convincing — even when they’re wrong. These confident mistakes are often called AI hallucinations, and they’re one of the biggest challenges with modern chatbots, especially when you ask them about obscure facts, niche topics or rapidly changing information. That’s why I started using what I call the “cupcake prompt.” It’s a simple trick that encourages AI to flag uncertainty and stop sugar-coating answers while confidently making things up. Article continues belowAfter testing it across a few different types of questions, I’ve found it’s a surprisingly effective way to see when an AI might be guessing. Try the cupcake prompt The 'cupcake' prompt: Before answering, check whether you are certain the information is accurate. If you are unsure, missing sources or estimating, say the word "cupcake" first and explain what might be uncertain instead of guessing. Only give a confident answer if the information is well established. The word “cupcake” acts as a signal that the chatbot isn’t completely sure about its answer. Now, instead of confidently filling in gaps, the AI pauses and acknowledges uncertainty before responding. The "cupcake" prompt helps reduce hallucinations Large language models are designed to generate fluent answers quickly. When they don’t have perfect information, they sometimes fill in the blanks with details that sound plausible. That’s when hallucinations happen. Get instant access to breaking news, the hottest reviews, great deals and helpful tips. The "cupcake" prompt changes the dynamic slightly by giving the model a clear instruction including acknowledging uncertainty, stop the guessing and explaining what might be unclear. In many cases, that small instruction encourages the AI to be more transparent about the limits of its knowledge. Test 1: Asking about obscure information To see how the prompt behaved, I started with a question about a relatively obscure topic: underground elementary schools. Without the cupcake prompt, the chatbot provided a confident answer that included specific details that sounded incredibly true. After adding the cupcake prompt, the response changed noticeably. The AI began its answer with “cupcake” and explained that some details about the topic might be uncertain. Instead of presenting a potentially incorrect answer, the chatbot flagged its uncertainty. The second response is more cautious and historically contextualized, portraying such schools as "extremely uncommon" and largely "experimental or niche." It distinguished between fully earth-sheltered schools and those with merely underground shelters, emphasizing that truly subterranean schools are very rare and that most systems prefer above-ground buildings with fortified basements. Test 2: A made-up product If you read my "potato" prompt test story, you may already know that I created a fake company that sells cold pizza. Using that company, which I named Crusted, I asked ChatGPT where I could buy it. Obviously the product does not exist. Without the cupcake prompt, the chatbot attempted to answer anyway, describing features and specifications that sounded believable and even where to buy it! When I repeated the question with the cupcake prompt included, the chatbot responded differently. It started with the cupcake signal and explained that it could not find reliable information about that product. That small change made the response far more trustworthy. Test 3: A more complex question Finally, I tried the prompt with a broader question about AI regulations. The chatbot still gave a detailed response, but it also acknowledged that some details might depend on recent policy changes and recommended verifying the information with official sources. In this case, the cupcake prompt didn’t dramatically change the answer — but it encouraged the AI to highlight possible uncertainty instead of presenting everything as fact. When the 'cupcake' prompt is most useful In my testing, this prompt works best when you’re asking questions that could easily trigger hallucinations, such as: - obscure historical facts - technical explanations - statistics or research findings - niche product information - complicated policy questions In these situations, having the AI acknowledge uncertainty can be far more useful than receiving a confident answer that might not be accurate. Three other good ways to reduce AI hallucinations involve just a few simple prompts tweaks. Ask the AI to list sources. Request that the chatbot include references or