2.2.1. Fabrications, Reliability, and Bias
💡 First Principle: AI fabrications (also called hallucinations) occur because models generate plausible text, not necessarily true text. They're confident even when wrong. This means verification mechanisms are essential, not optional, for any AI deployment.
Fabrications: The model generates information that sounds correct but isn't factual. This happens because:
- The model predicts probable next words, not verified facts
- Training data contained errors or outdated information
- The model extrapolates beyond what it actually knows
Reliability: AI outputs can vary between identical prompts, and models may struggle with complex reasoning. Mitigation strategies include:
- Providing clear context in prompts
- Using grounding to connect to verified data sources
- Implementing human review for critical decisions
Bias: AI can reflect and amplify biases present in training data. This manifests as:
- Unequal treatment of different demographic groups
- Assumptions embedded in recommendations
- Exclusion of perspectives underrepresented in training data
| Challenge | Business Risk | Mitigation |
|---|---|---|
| Fabrications | Incorrect information damages credibility | Ground on verified sources, implement human review |
| Reliability | Inconsistent outputs confuse users | Clear prompts, consistent context, defined guardrails |
| Bias | Unfair treatment, legal/ethical issues | Diverse testing, fairness monitoring, human oversight |
⚠️ Exam Trap: The exam may present fabrication risk and ask for the best mitigation. "Ignore it because AI is usually accurate" is always wrong. Look for answers involving grounding, verification, or human oversight.
Reflection Question: Your marketing team wants to use AI to generate customer-facing content automatically without review. What risks would you highlight, and what guardrails would you recommend?
