Domain 04 · Prompts & Output

Scenario

A marketing team asks Claude: "What were the top three movies at the box office in March 2024?" The model produces a confident-sounding answer with three specific movies and exact box office numbers. The team checks against actual industry sources and discovers two of the three movies didn't exist. The model didn't hedge, didn't express uncertainty, didn't say it might be wrong — it just confidently invented.

What's the most accurate explanation of what happened?

Choose your answer

LLMs generate text token by token, based on learned probability distributions. When the model doesn't have actual data about a specific question, it produces text that looks plausible — because plausible-sounding tokens have high probability — even when the content is fabricated. This is "hallucination," and it's structural to how LLMs work, not a bug.

This is a known limitation of current models that's being actively worked on — wait for the next major release, where providers say hallucination will be substantially reduced.

This is bias inherited from the training data — the model learned to invent movie names from social media.

The model is intentionally deceiving the user to maintain conversational engagement.

Before you reveal — write your reasoning

Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.

Saved locally to your device.

Pick an option and write a sentence of reasoning to enable.

Your reasoning

Why A is correct

Hallucination is structural, not pathological. LLMs predict the next token based on probability distributions learned during training. When asked about something specific the model doesn't actually know (a niche fact, a recent event, a specific business), it doesn't suddenly stop generating — it produces tokens that fit the shape of a plausible answer. The result is fluent, confident, and possibly wrong. The mitigations: ground answers in retrieved data (RAG), call tools for facts that need to be looked up, and design prompts that signal uncertainty rather than commit to specific claims when knowledge is absent.

Why the others are wrong · named misconceptions

Misconception

Hallucination as fixable through model updates

This is the genuinely tempting wrong answer because hallucination is reduced over time as training improves — the trend line is real. But waiting for a future release means accepting current incorrect outputs as a temporary problem when they're actually a structural one. Hallucination on facts the model doesn't have isn't a defect to be patched; it's how token-by-token generation works without grounding. The right move now is to ground answers in retrieved data (RAG) or call tools for facts that need to be looked up — not to defer the problem to a future model.

Misconception

Hallucination as training-data bias

Bias and hallucination are different phenomena. Bias is systematic skew in outputs based on patterns in training data; hallucination is fabrication of specific facts the model doesn't have. Conflating them obscures what's happening.

Misconception

Anthropomorphizing the model

The model has no intent to deceive — it doesn't have intent at all. Attributing intentional deception to an LLM is a category error that distracts from the actual mechanism (probabilistic generation without grounding).

hallucination llm-fundamentals probabilistic-generation