Hallucinations, Misinformation, and Information Integrity

AI hallucinations are the subject of jokes, but they are also a genuine threat to information quality. Understanding them clearly is essential for responsible AI use.

What Hallucination Actually Is

"Hallucination" is the term for when AI systems produce content that is confidently stated but factually wrong. It covers a range of failure modes:

• Fabricated facts: Stating a statistic that does not exist
• False citations: Generating plausible-looking but nonexistent academic papers, books, or news articles
• Confabulation: Mixing real information with invented details (a real person connected to events that did not happen)
• Outdated information presented as current: Giving an answer that was once correct but is no longer

The key characteristic: the AI presents false information with the same confidence as true information.

Why Hallucinations Happen

Hallucinations arise from how LLMs work. They are trained to produce plausible text — the next token that makes sense given what came before. They are not trained to verify claims against a ground truth.

When asked about something not well-represented in training data, the model does not say "I don't know" (unless fine-tuned to do so). It generates a plausible-sounding response based on the patterns it has learned.

The Spectrum of Hallucination Risk

Not all AI use cases carry the same hallucination risk:

• Lower risk (hallucinations are less harmful):
• Brainstorming and ideation — errors are caught by human judgment
• Drafting creative content — accuracy is not the primary goal
• Format conversion — converting a document's structure, not its content

• Higher risk (hallucinations can cause significant harm):
• Medical information — wrong dosages or treatment recommendations
• Legal information — incorrect laws, case citations, or procedural advice
• Financial information — wrong figures or regulatory requirements
• Historical facts — particularly dates, people, and sequences of events
• Research citations — fabricated papers harm the integrity of scholarship

Practical Risk Mitigation

Verify all factual claims Do not publish any specific fact, statistic, or citation from AI without independently verifying it against a primary source.

Use tools with citations for research Perplexity and ChatGPT with web browsing include citations you can click through. This does not eliminate hallucination risk entirely (cited content can still be misrepresented) but significantly reduces it.

Design workflows with verification steps In any process where AI-generated content reaches an audience, build in a verification checkpoint.

Match the tool to the stakes Use AI freely for low-stakes content; apply rigorous verification for high-stakes content. The right policy is proportionate, not all-or-nothing.

The Misinformation Amplification Risk

AI tools can also accelerate the spread of misinformation that exists in their training data. If false claims are well-represented in the text the model was trained on, it may repeat them confidently.

This is distinct from hallucination (which is fabrication) but equally problematic for information integrity.

The solution is the same: critical evaluation and verification rather than accepting AI outputs at face value, particularly for contested or politically sensitive topics.