Learning about AI: The Risk of Hallucinations
When using AI, have you ever come across information that made you think, "Wait, is this true?"
In fact, AI carries a risk called "hallucination," where it can confidently generate incorrect information.
Today, I will explain this phenomenon in an easy-to-understand way.
What is a hallucination?
Hallucination literally means "a sensory perception without an external stimulus."It is commonly referred to as
a "plausible lie."It refers to a phenomenon where AI, much like a human experiencing a hallucination, creates non-existent information in a convincing manner.For example, when you ask ChatGPT to "tell me about a paper on X," it might provide a non-existent paper title or author name as if it were real.
The troublesome part is that because the AI answers with such confidence, it is difficult to notice the lie at a glance.
Specifically, there are patterns like these:
Citing non-existent references or papers
Attributing fictional statements to real people
Generating statistical data that contradicts facts
Making calculation errors
Presenting old information as the latest information
Why does it happen? The causes of hallucinations
The reason hallucinations occur is actually related to the very mechanism of AI.
There are three main causes.
1. Issues with training data
AI learns from past data, but if that data is insufficient, biased, or outdated, it cannot make accurate judgments.
The AI tries to fill in the gaps in its knowledge by creating information through guesswork.If you ask for the latest information from 2026, but the training data only goes up to 2024, the AI has no choice but to create a "plausible answer" through estimation.
2. The operating principle of AI
Generative AI creates text by calculating the "probability of the next word."
In other words,it is not answering based on whether it is a fact, but based on the probability that "this word is likely to come next in this flow."That is why it can confidently state plausible lies.
3. Ambiguous questions
When we ask ambiguous questions, the AI guesses, "Maybe it means this?" and answers accordingly.
In this process of estimation, the risk of generating incorrect information increases.
[Proactive Measures] Techniques to reduce hallucinations
While it is impossible to prevent them completely, there are ways to lower the occurrence rate.
Here are some measures considered effective as of 2026.
1. Give specific and clear instructions
Instead of saying "Tell me about X," ask specifically, such as "Provide statistical data for X in 2025, citing official sources."
Reducing ambiguity leaves less room for speculation.
2. Tell it to "say you don't know if you don't know"
Simply adding "If the information is uncertain, please answer 'I don't know'" to your prompt is quite effective.
It is important not to force the AI to answer.
3. Utilize RAG (Retrieval-Augmented Generation)
Modern AI tools use RAG technology, which searches external, reliable information sources while generating answers.
ChatGPT's search feature and Bing Chat are examples of this.
Since it can reference real-time information, hallucinations are significantly reduced.
4. Verify information with multiple AIs
Try checking information generated by ChatGPT with other AIs like Gemini, Claude, or Perplexity.
If multiple AIs provide the same answer, the reliability increases.
5. Ask for the basis of the answer
By asking follow-up questions like "What is the source of that information?" or "Where did you get those figures?", you can make the AI provide its reasoning.
If the basis is vague, you should doubt that information.
[Post-measure] Final verification is essential
I have introduced these measures, but I will be honest with you.
No matter how careful you are, you cannot reduce the risk of hallucination to zero.
This is because AI is essentially a "machine that generates words based on probability."
It does not actually understand facts.
That is precisely why final verification is a human job.
Fact-checking is absolutely necessary, especially in the following cases:
Fields with high expertise where mistakes are not allowed, such as medicine or law
Numerical data or statistical information
Citations or references
Latest events or news
Information related to important decision-making
Specific verification methods include:
Accessing official websites or primary information sources
Cross-check with multiple sources
Verify with experts
Fact-check using Google Search
Instead of making decisions based on "because the AI said so,"it is important to adopt a stance of "using AI suggestions as a reference, then verifying and deciding for yourself".
[Summary] How to interact wisely with AI
Hallucination is an unavoidable risk when using AI.
However, there is no need to be afraid if you understand the mechanism and deal with it appropriately.
What is important is:
Knowing the limitations of AI
Improving accuracy with specific questions
Verifying information through multiple methods
Making the final decision yourself
AI is a wonderful tool, but it is not perfect.
By understanding the characteristics of AI and using it wisely, it will become the best partner to significantly accelerate your work and learning.
"Don't trust AI too much, but don't fear it too much either"
This exquisite sense of balance may be the skill we need to live in the coming AI era.