AI Risks and Responsible Use — Lesson 2
Hallucinations and Source Verification
Learning Objectives
- 1Understand why AI hallucinates and when it is most likely.
- 2Build practical verification workflows for AI-generated content.
- 3Instruct AI to flag uncertainty and cite sources.
Why AI makes things up
AI hallucination is not a bug — it is a feature operating outside its intended scope. Language models generate text by predicting the most likely next token. When asked about something they do not have reliable information about, they do not say "I do not know." They predict what a correct answer would look like and generate plausible-sounding text that may be entirely fabricated.
Hallucination is most likely when: the question involves specific facts, dates, numbers, or citations; the topic is niche or recent; the question asks for information the model was not trained on; or the prompt encourages the model to be comprehensive rather than accurate.
The practical danger is that hallucinated content is indistinguishable from accurate content in style and confidence. A fabricated statistic reads exactly like a real one. A made-up citation looks exactly like a genuine one. Human review is the only reliable defense.
Verification workflows
For any AI output that will be published, shared externally, used for decisions, or included in official documents, establish a verification workflow. At minimum: identify every factual claim, check each one against a reliable source, verify any citations or references actually exist, and confirm that numbers and statistics are accurate.
For high-volume use cases where individual verification is impractical, use sampling. Review a random sample of AI outputs weekly. Track the error rate. If the error rate exceeds your tolerance, adjust the prompt, add constraints, or increase the review sample size.
Instruct the AI to support verification: "Cite the source for each statistic. If you are not confident in a fact, say so explicitly. Do not fabricate citations." These instructions do not eliminate hallucination, but they can reduce it and make fabricated content easier to identify.
Building verification into team habits
Verification should not feel like punishment. Frame it as quality assurance — the same review process that any professional applies to important work. Nobody sends a legal brief without proofreading. Nobody publishes financial results without checking the numbers. AI-generated content deserves the same standard.
Create simple checklists for different output types. A customer email checklist might include: Are the facts correct? Is the tone appropriate? Does it promise anything we cannot deliver? A report checklist might include: Are all statistics sourced? Are conclusions supported by the evidence presented?
Track errors when they are caught in review. This data helps improve prompts, identify high-risk use cases, and justify the review process to stakeholders who might question why it is necessary.
Case Study
The legal brief with fake citations
Situation
A lawyer used AI to help draft a legal brief and included case law citations generated by the AI. The citations looked legitimate — proper format, plausible case names, realistic summaries. The opposing counsel checked the citations and discovered that three of the six cases cited did not exist. The lawyer faced sanctions from the court.
Analysis
The AI generated citations that followed the pattern of real legal citations, which is exactly what language models do — they generate plausible text based on patterns. The lawyer treated a high-consequence output with no verification because the citations looked convincing.
Takeaway
Every citation, reference, statistic, and factual claim in AI-generated output must be independently verified. The more professional the output looks, the more dangerous uncaught errors become.
Reflection Questions
- 1. Have you ever shared AI-generated content without verifying the facts it contained? What was the risk?
- 2. For your team AI use cases, is there a defined verification step before output reaches customers or stakeholders?
Key Takeaways
- ✓Hallucination is inherent to how language models work — they predict plausible text, not verified truth.
- ✓Hallucinated content is indistinguishable from accurate content in style and confidence.
- ✓Verification workflows are mandatory for any AI output that affects decisions or reaches external audiences.
- ✓Instruct AI to cite sources and flag uncertainty — it helps but does not eliminate hallucination.