Lesson Content

What hallucinations actually are.

"Hallucination" in AI describes the phenomenon where a model generates specific, plausible-sounding claims that are simply wrong – invented. Not hedged, not qualified, not flagged as uncertain – just stated as fact.

Examples of hallucination:

• Citing a journal article that does not exist with realistic author names, volume numbers, and page numbers
• Describing the exact content of a court case that was never decided
• Quoting a person saying something they never said
• Providing a product specification with specific – wrong – technical numbers

Why hallucinations happen.

The language model is predicting text, not retrieving verified facts. When asked about something it does not know well, it does not say "I don't know" – it generates text that looks like an answer. The generation process does not have a "confirm this exists before outputting it" step. The result is confident-sounding text describing things that are not real.

Hallucination frequency varies by domain and task:

• Citations and references are high-risk (AI generates plausible but often invented ones)
• Specific figures and statistics are high-risk (specific numbers are confidently generated, often wrong)
• Well-documented general topics are lower-risk (but not no-risk)

Practical hallucination defenses.

• Never use AI citations without independently verifying the source exists and says what AI claims
• For any specific figure, statistic, or technical detail you will rely on, verify against a primary source
• Ask Claude to express uncertainty: "What are you not sure about in this response?"
• For important factual questions, use Claude to generate a map of what to research – then research it yourself

What bias means in AI.

AI models learn from training data that reflects human writing – including the biases, perspectives, and underrepresentation in that writing. Several bias types appear in AI outputs:

Representation bias: Groups, cultures, and perspectives underrepresented in training data are less fully or accurately represented in outputs. Common subjects, Anglophone perspectives, and majority-culture viewpoints are over-represented.

Association bias: AI learns the associations present in training text. Stereotype-consistent associations may appear in generated content – in descriptions, examples chosen, roles assigned in generated scenarios.

Perspective bias: AI may present mainstream or consensus views more fully than minority, contrarian, or non-Western perspectives – not because it was designed to, but because training data coverage is uneven.

Practical bias management.

• Recognize when AI's output may reflect under-representation rather than accuracy
• For decisions involving diverse stakeholders, do not rely on AI output as a substitute for actually hearing from those stakeholders
• For analysis requiring multiple perspectives, explicitly ask Claude to present contrasting or minority viewpoints alongside the main view