Avoiding AI Hallucinations and Errors

# CHAPTER 15

Avoiding AI Hallucinations and Errors #

1. Introduction #

2. Learning Objectives #

• Understand the root cause of AI Hallucinations.

• Write prompts that force the AI to admit when it doesn't know an answer.

• Implement prompt-based Fact-Checking constraints.

• Utilize the "Provide Citations" technique.

3. Beginner-Friendly Explanation #

4. The Anti-Hallucination Clause #

Dangerous Prompt:

Answer the user's question about our company policy.

*(If the policy isn't provided, the AI will invent a fake policy).*

Safe Prompt (The Opt-Out Clause):

Answer the user's question about our company policy.
CONSTRAINT: If you do not know the answer, or if the answer is not explicitly stated in the provided text, you MUST reply EXACTLY with: "I'm sorry, I do not have that information." DO NOT guess. DO NOT invent facts.

5. Grounding with Context (RAG Basics) #

# CONTEXT
[Paste a verified Wikipedia article or company document here]

# TASK
Based STRICTLY on the Context provided above, answer the question.

6. Forcing Citations #

Summarize the 10-page contract provided. 
For every claim you make in your summary, you MUST provide a direct quote from the contract in parentheses.
Example: The contract lasts 5 years ("Term: 5 Years").

This constraint prevents the AI from summarizing something that isn't actually in the text, because it is mathematically forced to find a matching quote.

7. Fact-Checking (The Self-Correction Prompt) #

Step 1: The AI generates an essay. Step 2: You pass the essay back into the AI with a new prompt:

Act as a ruthless Fact-Checker. Read the essay below. 
Identify any claims, statistics, or dates that might be hallucinated or incorrect. Output a list of potential errors.

8. Python Example: The Guardrail Loop #

def generate_and_verify(question, document):
    # Step 1: Generate Answer
    prompt1 = f"Answer '{question}' based on '{document}'."
    draft_answer = call_llm(prompt1)
    
    # Step 2: Verify Answer
    prompt2 = f"Does this answer: '{draft_answer}' rely on ANY information NOT present in this document: '{document}'? Answer Yes or No."
    verification = call_llm(prompt2)
    
    if "Yes" in verification:
        return "Error: Hallucination detected. Answer rejected."
    return draft_answer

9. Mini Project #

10. Best Practices #

• Temperature Control: As discussed in previous chapters, if you are performing factual extraction or research, lower the API temperature to 0.0. This mathematically removes the AI's "creativity," significantly reducing the chance of hallucination.

11. Common Mistakes #

• Assuming High Quality Writing = Truth: The most dangerous hallucinations are the ones that sound the most professional. Never let the AI's flawless grammar trick you into believing its fabricated statistics. Always verify.

12. Exercises #

1. 1. Explain why giving an LLM an explicit "Opt-Out Clause" (e.g., "If you don't know, say 'I don't know'") drastically reduces hallucination rates.

13. MCQs with Answers #

14. Interview Questions #

• Q: Describe the mechanical cause of an LLM hallucination. What specific Prompt Engineering constraints would you implement to prevent an AI customer service bot from inventing fake refund policies?

• Q: How do you use "Self-Reflection" or a secondary verification prompt to automatically detect and filter out hallucinations in an enterprise pipeline?

15. FAQs #

16. Summary #

17. Next Chapter Recommendation #