1. Introduction: What Does it Mean for AI to "Lie"?

With the advancement of AI technology, conversational AIs like ChatGPT and Genspark have become commonplace in daily life. However, these AI tools have a significant problem known as "Hallucination".

Hallucination refers to the phenomenon where AI confidently outputs plausible falsehoods or incorrect information. While AI does not intentionally lie, it can end up providing incorrect information.

What You Will Learn in This Article

  • Basic concepts and causes of AI hallucination
  • Specific examples of hallucinations that actually occur
  • Practical methods for detecting hallucinations
  • Countermeasures AI beginners should take

For AI beginners to safely utilize AI tools, it is essential to understand the existence of hallucinations and know how to deal with them appropriately. This article provides an easy-to-understand explanation, even without specialized knowledge.

2. What is AI Hallucination?

AI Hallucination refers to the phenomenon where AI generates and outputs information that is not factual as if it were true. "Hallucination" in English means "delusion" or "illusion," and it's likened to a state where AI "sees things that aren't there."

2.1 Characteristics of Hallucination

  • Confident Output: Incorrect information is presented with certainty
  • Plausibility: Text and structure that appear correct at first glance
  • Difficulty in Verification: Hard to notice errors without specialized knowledge
  • Lack of Consistency: Different answers may be returned to the same question

Simple Example of Hallucination

Question: "What is the highest mountain in Japan?"
Correct Answer: "Mount Fuji (elevation 3776m)"
Hallucination Example: "Mount Kita (elevation 3193m)"

In this example, while Mount Fuji is actually the correct answer, the AI might confidently respond with "Mount Kita." Since Mount Kita is the second-highest mountain in Japan, the key point is that it's not complete nonsense, but a plausible error.

2.2 Types of Hallucination

AI hallucinations can be broadly categorized into three types:

  • Factual Misconception Type: Errors in historical facts, statistical data, etc.
  • Information Fabrication Type: Creation of non-existent papers, URLs, personal names, etc.
  • Logical Contradiction Type: Content that contradicts the surrounding context

Important Point

Hallucination is due to AI's technical limitations and cannot be completely prevented at present. Therefore, it is important for AI beginners to use AI with the premise that "AI can make mistakes."

3. Why Do AI Hallucinations Occur?

Understanding why AI hallucinations occur helps you know when to be cautious. The main causes are the following four:

3.1 Limitations of Training Data

AI learns from vast amounts of text data, but the training data itself may contain errors or biases. Furthermore, it does not possess information beyond the training data's cutoff date.

Example of Training Data Limitations

If you ask an AI trained on data up to January 2024, "What was the temperature in Tokyo in October 2024?", it will have to guess due to a lack of actual data, making hallucinations more likely to occur.

3.2 Probabilistic Text Generation

Current AIs generate text by calculating "the probability of the next word". Because of this, even if the grammar is correct, factually incorrect content may be generated.

3.3 Misunderstanding of Context

In long conversations or complex contexts, AI may misunderstand the intent of a question. Especially when ambiguous expressions or technical terms are mixed, it may give answers based on incorrect interpretations.

3.4 Overfitting and Overconfidence

AI learns patterns from training data, but it can over-generalize learned patterns. This leads to presenting information that cannot be applied even in similar situations.

What AI Beginners Should Remember

AI does not "possess knowledge," but merely "learns patterns." Therefore, it is wise to utilize it not as an accurate knowledge base, but as a tool for idea generation and drafting.

4. Real-World Cases: Specific Examples of AI Hallucinations

Let's look at specific examples of what kinds of hallucinations actually occur.

4.1 Non-Existent Papers or References

Hallucination Case ①

Question: "Please tell me about the latest papers on machine learning."
AI's Answer: "The paper 'Deep Learning Advances in 2024' by Smith et al. (2024) is a good reference. It was published in Nature."

However, this paper may not actually exist. AI can create natural-looking paper titles, author names, and journals.

4.2 Incorrect Code or API Information

Hallucination Case ②

Question: "Please tell me how to use the Twitter API with Python."
AI's Answer: "Use the tweepy.API.get_user_timeline() method."

In reality, method names and usage may have changed in Twitter API v2. API specification changes are common, making this an area where hallucinations are prone to occur.

4.3 Errors in Historical Facts

Hallucination Case ③

Question: "When did World War II begin?"
AI's Answer: "It began in 1941."

The correct answer is 1939, but the AI may be confusing it with the year of the start of the Pacific War (1941). In this way, hallucinations occur when related information gets mixed up.

4.4 Fabrication of Statistical Data

Hallucination Case ④

Question: "What is the size of Japan's AI market?"
AI's Answer: "As of 2024, it is approximately 5 trillion yen."

When specific figures are requested, AI may fabricate plausible numbers. Statistical data, in particular, requires verification.

Questions Prone to Hallucination

  • Questions seeking the latest information (e.g., "in 2024")
  • Questions requesting specific papers or literature
  • Questions requesting accurate statistical data or figures
  • Questions requiring complex technical specifications or specialized knowledge

5. How to Detect AI Hallucinations

Here are 5 methods for detecting hallucinations that even AI beginners can put into practice.

5.1 Verify with Multiple Sources

Make it a habit to check AI's answers with Google search or official documentation. Crucial information and figures, in particular, always require verification.

Examples of Verification Methods

  • Search Google Scholar by paper title or author name
  • Verify statistical data on government or public institutional websites
  • Refer to official documentation for technical information

5.2 Request Explicit Sources

By asking additional questions to the AI like "What is the source of that information?" or "Where can I verify this?", you can sometimes determine if it's a hallucination.

5.3 Repeat the Same Question

If you ask the same question multiple times with slightly different phrasing, the answers may not be consistent. This is a sign of hallucination.

5.4 Cross-Reference with Common Sense

Check if AI's answers contradict common sense or basic knowledge. Information that is too surprising should be verified with particular care.

5.5 Check for Specificity

Answers with many ambiguous expressions might be evidence that the AI is not confident. Caution is needed if expressions like "probably" or "generally" are frequent.

Traps AI Beginners Often Fall Into

The more an AI's answer appears "detailed" and "technical," the more inclined people are to believe it. However, detailed explanations and the use of technical jargon do not necessarily guarantee accuracy. Rather, it might just be a more sophisticated hallucination.

6. How to Deal with AI Hallucinations

While hallucinations cannot be entirely prevented, there are practical countermeasures to minimize the risks.

6.1 Basic Countermeasures for AI Beginners

5 Basic Rules

  1. Critical Thinking: Do not unconditionally trust AI's answers
  2. Habitual Verification: Always confirm important information with multiple sources
  3. Clear Questions: Avoid ambiguous questions and ask specifically
  4. Step-by-step Confirmation: Break down complex questions into smaller parts for verification
  5. Check for Updates: Supplement the latest information with real-time searches

6.2 Specific Questioning Techniques

There are effective questioning methods to reduce hallucinations:

  • Specify Constraints: Specify a deadline, such as "using information as of December 2023"
  • Ask for Certainty: Preface with "Please only provide information that is definitely correct"
  • Encourage Acknowledgment of Uncertainty: "If you don't know, please say 'I don't know'"

6.3 Differentiated Use Cases

The risk level of hallucination changes depending on the AI's intended use:

  • Low-Risk Uses: Idea generation, drafting text, brainstorming
  • Medium-Risk Uses: Code generation, translation, summarization
  • High-Risk Uses: Medical information, legal consultation, financial advice (expert verification essential)

Uses to Absolutely Avoid

For the following uses, you should absolutely avoid using AI's answers as they are:

  • Medical diagnoses or important health-related decisions
  • Creation of legal documents or contracts
  • Investment decisions or financial planning
  • Citations or references in academic papers

6.4 Continuous Learning

AI tools are evolving daily. It is important to pay attention to the latest information and updates and to continue understanding the characteristics and limitations of each tool.

7. Conclusion: How to Interact Effectively with AI

AI hallucination is an unavoidable challenge in modern AI technology. However, by understanding its existence and addressing it appropriately, AI tools can become powerful allies.

3 Key Points for AI Beginners to Remember

  1. AI is not perfect: Hallucinations are a technical limitation and can always occur
  2. Verification is essential: Always confirm important information with multiple reliable sources
  3. Utilize Appropriately: Understand AI's strengths and use it for suitable purposes

AI should be viewed not as an "all-purpose answer," but as an "excellent assistant." By combining it with human judgment, AI tools will unleash their true value.

Reference Links