The Flaws: Hallucinations and Bias

LLMs are impressive, but they are not truth machines. Because they are statistical engines (Lesson 2), they suffer from two major problems: Hallucinations and Bias.

1. Hallucinations: Confident Lies

A "Hallucination" is when an AI generates an answer that sounds perfect but is factually wrong.

• Why it happens: The model is trying to predict the most likely next word, not the most truthful one. If the model is 51% sure the capital of California is San Francisco (because it's a "Big City"), it might just say it with total confidence.

2. Algorithmic Bias

AI models are "Prejudiced" by the data they are trained on.

• The Problem: If 90% of the computer science books in the training data were written by men, the AI might unconsciously associate the word "Software Engineer" with the pronoun "He."
• The Risk: Bias can lead to unfair hiring, offensive content, or the erasure of different cultures.

---

Common Failure Modes

Failure Mode	Definition	Real-world Example
Hallucination	Factually incorrect but fluent.	Making up a fake legal case name.
Stereotyping	Generalizing based on training data.	Assuming a "CEO" is always male.
Logic Gaps	Failing at simple arithmetic.	Saying 1,000 + 1,000 = 3,000.
Sycophancy	Agreeing with the user too much.	Confirming a user's wrong belief.

---

💡 Guidance for Learners

Trust, but verify. Never use an LLM for factual research without a second source. Treat the LLM as a highly intelligent but occasionally drunk intern.

---

Visualizing the Hallucination Risk

graph TD
    User[Query: 'What is the speed of a unicorn?'] --> AI[AI Engine]
    AI --> Logic{Is this a logical fact?}
    Logic -->|No| Stat[Statistically likely words regarding speed]
    Stat --> Result['A unicorn flies at 50mph.']
    Result --> Warning[Hallucination Detected!]

---

Summary

• Hallucinations are fluent, confident, but false statements.
• Bias is a reflection of the "Prejudices" found in internet-scale data.
• Verification is the responsibility of the human user.