The essentials in 30 seconds
Las hallucinations are AI responses that seem correct but contain made-up data, false figures, or non-existent quotes. According to a study published in Nature, More than half of the references generated by GPT-3.5 are fabricated, and later models such as GPT-4 still produce between 18 and 20 % according to Deakin University researchers. The models prioritise Coherence above truthfulness, and theoretical investigations suggest that hallucinations They cannot be completely deleted. In general language models. Therefore, verifying is always your responsibility – not the AI’s.
6 ways to verify ChatGPT information
The cupcake trick
Add this instruction to your prompt: “Before answering, check if the information is accurate. If you have doubts, lack sources or are making an estimation, first say the word ‘cupcake’ and explain what might be uncertain.” If the AI says “cupcake”, you know that part needs external verification. Method popularised in March 2026.
Lectura lateral
Don't stay within ChatGPT. Take every specific data point — name, date, figure, quote — and search for it directly on Google, Google Scholar, or the original source. It's the same method that professional fact-checkers use: leave the source to verify it with others.
Claim fractionation
Break down the AI's response into individual claims and verify each one separately. A response may have 5 facts: 3 correct, 1 partially true and 1 made up. Do not assume that “if part is correct, all of it is”.
AI with integrated search
Tools like Perplexity, ChatGPT with web browsing o Copilot They combine a language model with a search engine. This reduces hallucinations because the AI searches the web and cites sources with clickable links that you can verify.
Automatic fact-checking tools
Originality.ai offers an automated checker with an accuracy rate of 86.69%. Full Fact AI It is used by fact-checkers in over 30 countries. Genaios It has a Chrome extension that compiles sources that confirm or refute each claim.
Request explanation of reasoning
Ask ChatGPT: “Where are you getting this information from? Is it a verified fact or an inference?” If AI cannot provide a concrete source or starts to “improvise” a reference, that's a red flag. If you want to delve deeper into this topic, read about How to detect ChatGPT's reasoning theatre.
Step-by-step verification process
Each time ChatGPT gives you a response you're going to use for making decisions, following these steps takes you less than 2 minutes and can save you costly mistakes:
- Identify the critical data. Read the response and mentally underline: names, dates, figures, direct quotes, URLs, and statistical claims. These are the points at highest risk of hallucination.
- Apply the cupcake trick. Before your next question, add the warning instruction. If the AI says “cupcake”, that answer has uncertainty and requires mandatory external verification.
- Seek each piece of data from primary sources. Open Google and search for the specific claim. Prioritise official sources: .gov, .edu sites, academic papers, company reports, and reputable media. Do not settle for blogs which may also be citing AI.
- Check the quotes and URLs. If ChatGPT gives you a URL, open it. Remember that, according to studies, between 18% and 55% of the citations generated by ChatGPT are fabricated. If the link doesn’t exist or the content doesn’t match what ChatGPT claims, disregard that information.
- Use a second AI as a contrast. Ask the same question on Perplexity (which cites sources), Claude or Gemini. If all three models give different answers, it's a sign that the information is uncertain and you need a human source.
- When in doubt, trust the human source. If you can't verify a fact with a reliable primary source, don't use it. It's better to leave a gap than to publish false information attributed to “the AI said so...”.”
Table: type of claim vs. how to verify it
| Type of statement | How to check it | Recommended tools |
|---|---|---|
| Statistics and figures | Search for the original study or report cited | Google Scholar, Statista, .gov/.edu sources |
| Direct quotes from people | Search for the exact quote in quotation marks on Google | Google (exact search with “ ”) |
| Historical facts / dates | To contrast with encyclopedias and academic sources | Wikipedia (as a starting point), Britannica |
| Legal or medical information | Consult official sites in the area; Never trust only AI | .gov sites, WHO, medical/legal publications |
| Code and technical documentation | Run the code; review the official documentation | Official documentation (MDN, Python docs, etc.) |
| URLs and links | Open the link directly in the browser | Your browser + Wayback Machine if it doesn't load |
| Recent news | Search reference media; AI may not have up-to-date data | Google News, Perplexity, Reuters, AP |
Practical tip: When you use ChatGPT for work or content that others will read, enable the web browsing (available in ChatGPT Plus and Team). This doesn't eliminate hallucinations entirely, but it significantly reduces them because the model can consult real-time sources and show you the links. Another option: use Perplexity directly, which was specifically designed to provide answers with verifiable citations. If you want to learn how to make the most of AI, check out the Basic guide to understanding ChatGPT.
When is ChatGPT most likely to fail
Not all question types carry the same risk of hallucination. These are the scenarios where you should be especially careful:
-
Specific numerical data. Market figures, percentages, prices, and statistics are the number one weak point. AI tends to generate numbers that “sound reasonable” but don't correspond to any real data.
-
Quotations and academic references. ChatGPT makes up paper titles, authors and publication years. According to research, between 18% and 55% of the citations it generates are fabricated. Never use a citation from ChatGPT without checking it on Google Scholar first.
-
Legal and medical information. Even the best models get legal information wrong 6.4% of the time. In medicine, the risk is even higher for rare conditions. Never Do not make health or legal decisions based solely on what an AI says.
-
Recent or ongoing events. Models have a knowledge cut-off date. If you ask about something that occurred after that date and there is no web browsing enabled, the AI will invent an answer instead of saying “I don't know”.
-
Little-known individuals or niche topics. In the PersonQA benchmark, which assesses knowledge of public figures, models such as o3 achieve an accuracy of just 59% (%). The less well-known the subject, the more likely the AI is to fill in the gaps with made-up information.
-
When the answer is too perfect. If ChatGPT gives you an extremely detailed answer with exact figures and no doubt, be wary. Very specific data that cannot be easily verified is one of the main signs of hallucination.
Frequently asked questions
What are ChatGPT hallucinations?
Hallucinations are AI-generated responses that appear coherent and credible but contain false, made-up, or inaccurate information. This happens because language models prioritise textual coherence over factual accuracy. According to OpenAI, even their most advanced models continue to exhibit this problem, and theoretical research suggests it cannot be entirely eliminated in general language models.
What is the real error rate of ChatGPT?
It depends on the type of task. GPT-4 has a hallucination rate of approximately 1.5% in general tests, but on specific factual questions (the SimpleQA benchmark), the model achieved only 49% accuracy, with a 51% hallucination rate. In legal information, the best models get it wrong 6.4% of the time, and in programming 5.2%. Accuracy varies drastically depending on the domain and the specificity of the question.
How does the cupcake trick work for detecting AI errors?
It involves adding an instruction to the beginning of your prompt that tells the AI: “If you have any doubts about the accuracy of your answer, first say the word ‘cupcake’ and explain what might be uncertain.” This works because it gives the AI an escape route to admit uncertainty rather than inventing a confident answer. It works with ChatGPT, Copilot, Gemini, and other models.
Is Perplexity more reliable than ChatGPT for factual data?
For factual data, yes, it tends to be more reliable because Perplexity was designed to combine text generation with real-time web search. It displays clickable citations next to each claim, which allows you to quickly verify. However, it's not infallible: it's still an AI that can misinterpret the sources it finds. Always check the links it provides.
It is not advisable to rely on ChatGPT for medical or legal information. ChatGPT is a language model and is not qualified to provide professional advice in these fields. For medical concerns, always consult a qualified healthcare professional, and for legal matters, seek advice from a solicitor or barrister.
Not as a sole source, never. AI models generate incorrect legal information 64% of the time, and in medicine the risk increases with rare conditions or recent treatments. ChatGPT can serve as a starting point for understanding a concept, but any medical or legal decision must be based on consultation with a professional and verified official sources.
Do you want to understand in depth how the AI behind ChatGPT works?
Discover all about ChatGPT as a virtual assistant →Last updated: March 2026. Article rewritten with verified data from OpenAI, Originality.ai, Full Fact AI and analysis of SimpleQA and PersonQA benchmarks.
You may also be interested in
EN 
ES 
PT 
DE 
FR