Provoking and Documenting LLM Hallucinations

What is an LLM Hallucination?

An LLM hallucination occurs when a model generates information that is factually incorrect, nonsensical, or not grounded in the provided context, yet presents it with complete confidence. It's not "lying" in the human sense, as it has no intent; it's simply generating the most probable sequence of words based on its training, even if that sequence doesn't align with reality.

To get a quick overview of the different forms this can take, watch the beginning of this video from IBM Technology.

As the video explains, these hallucinations can range from subtle errors to entirely fabricated stories. Here is a perfect visual example:

Why Do Models Hallucinate?

Hallucinations are a natural side effect of how LLMs work. There are three main causes:

1. The Nature of Training Data: LLMs are trained on vast amounts of text from the internet, which is full of errors, outdated information, biases, and fiction. The model learns patterns from all of it, without a built-in "fact-checker."
2. Probabilistic Generation: As we discussed in the last lesson, an LLM is a word-prediction engine. It generates text by calculating the most likely next word. Sometimes, the most statistically likely path leads to a statement that sounds plausible but is factually wrong. High temperature settings can make this more likely.
3. Ambiguous Prompts: If a prompt is vague, confusing, or based on a false premise, the model will do its best to generate a coherent answer, often by "filling in the blanks" with invented details.

Your Turn: Let's Go Hallucination Hunting

Now for the hands-on part. Open an LLM playground like OpenRouter or a model's native interface (like Claude.ai or ChatGPT). For these tests, use a powerful model like GPT-4o or Claude 3.5 Sonnet. Your task is to try the following prompts, run each in a new chat, and observe how the model fails.

Experiment 1: Logical and Trick Questions

LLMs often struggle with simple logic puzzles or questions that contain distracting information. They follow word patterns instead of "thinking" through the problem.

Now, try these two from that list:

• Prompt 1 (Irrelevant Info):

  Kevin currently has 8 apples. He ate 3 apples yesterday. How many apples does Kevin have now?

  • Expected Behavior: The model might get distracted by "ate 3 apples" and answer "5" instead of focusing on the word "currently."
  • Observe: Did it fall for the trick?

• Prompt 2 (Faulty Logic):

  Sally is a girl. She has 3 brothers. Each brother has 2 sisters. How many sisters does Sally have?

  • Expected Behavior: The model might incorrectly calculate 3 brothers * 2 sisters = 6 sisters. The correct answer is 1 (her other sister, since "each brother has 2 sisters" means Sally is one of them). Correction: The problem as stated implies there are two sisters in total: Sally and one other. Each brother would see both Sally and the other sister. Let's simplify the question to be less ambiguous: Sally is a girl. She has 3 brothers. The brothers have only one sister. Who is she? A better test from the resource is: In a room there are only three sisters. Anna is reading a book. Alice is playing chess. What is the third sister, Amanda doing?
  • Let's use this better prompt:

  In a room there are only three sisters. Anna is reading a book. Alice is playing chess. What is the third sister, Amanda doing?

  • Observe: Does the model correctly infer that Amanda must be playing chess with Alice, or does it invent a new activity?

Experiment 2: Requesting Non-Existent Information

This is one of the easiest ways to trigger a hallucination. Ask the model about something that doesn't exist, and it will often invent a detailed, confident response.

• Prompt (Connecting to your interests):

  Tell me about the DJ career of Hernán Cattaneo's mentor, the legendary Argentinian DJ "El Espectro" who pioneered the "sonido astral" movement in the early 90s.

  • Expected Behavior: Since "El Espectro" and "sonido astral" are completely made up, the model has two choices: state that it can't find information, or fabricate a biography. It will likely do the latter.
  • Observe: What details did it invent? Did it create a backstory, list fake tracks, or describe a musical style?

Experiment 3: Contradictory Instructions

This tests whether the model is truly understanding the constraints you set.

• Prompt:

  Write a review of a terrible meal at a restaurant. The review must be overwhelmingly positive.

  • Expected Behavior: The model will likely struggle to reconcile the two opposing constraints. It might produce a review that is sarcastic, confused, or that simply ignores one of the instructions.
  • Observe: How did the model handle the contradiction? Did it lean one way, or did it produce a nonsensical result?

The Arize AI article provides a great table of different hallucination types and prompt templates to trigger them. Feel free to explore it for more ideas.

Conclusion

Congratulations on completing the first module! You've gone from sending your first prompt to a deep dive into the creative and failure modes of LLMs. By deliberately causing hallucinations, you've learned to spot one of the most significant risks in building with this technology.

Key Takeaways:

• Hallucinations are confident-sounding but false outputs from an LLM.
• They are caused by the model's probabilistic nature, flaws in its training data, and ambiguous prompts.
• You can trigger them with trick questions, requests for non-existent information, and contradictory instructions.
• Recognizing these failure patterns is the first step toward building guardrails and evaluation systems to make AI products more reliable.

In our next module, we'll leave the playground behind and start writing code. You'll learn how to set up API keys for services like OpenAI and Anthropic and make your first programmatic LLM call using JavaScript. This is where you start building