What's changed: Created Generative AI Leader Chapter 1 (Domain 1 "Fundamentals of generative AI": AI/ML/gen AI relationship = nesting/learning types/data types/modalities-multimodal; LLMs and foundation models = how LLMs work/tokens/prompts/pretraining-fine-tuning/embeddings; limitations and risks = hallucination/bias/knowledge cutoff/privacy).
1.3Limitations and risks of generative AI
Learn the limitations and risks to understand when using generative AI in business: hallucinations (plausible but wrong outputs), bias from training data, knowledge cutoff, output uncertainty and reproducibility, the difficulty of explainability, and data privacy, copyright, and security.
Generative AI is powerful, but as-is it carries business risk. Leaders must understand not only benefits but limitations, and decide to use it with appropriate controls (human review, grounding). Here are the key risks.
1.3.1Hallucinations and bias
Hallucination is when generative AI confidently outputs content that is plausible but factually wrong. Because an LLM builds text by "next-token prediction," it does not guarantee facts. Mitigations include grounding/RAG (Chapter 3) to base outputs on trusted data, and human review. Bias in training data leads to bias (unfair or discriminatory outputs). These mean "high accuracy" does not equal "safe."
1.3.2Other limitations
| Limitation/risk | What it is | Implication |
|---|---|---|
| Knowledge cutoff | Unaware of events after training data | Supplement recency via grounding |
| Uncertainty/reproducibility | Same input can yield varying output | Needs verification and review |
| Explainability | Hard to show why an output was produced | Humans oversee critical decisions |
| Privacy/copyright/security | Risks to sensitive data, rights, and attacks | Requires data governance |
A model only knows up to its knowledge cutoff; recency is supplemented by external references. Generation is probabilistic, so the same input can yield varying output (uncertainty/reproducibility), and explainability is limited. Including sensitive data in inputs risks leakage, and there are concerns about copyright of outputs and security such as attacks that abuse prompts (prompt injection). These connect to responsible AI and data governance in Chapter 4.
Common: symptom → term/mitigation. E.g., "plausible but wrong" = hallucination (mitigate via grounding/human review); "skewed output" = bias; "unaware of recent events" = knowledge cutoff; "output varies each time" = uncertainty/reproducibility; "hard to explain why" = explainability; "handling sensitive data" = privacy/data governance.
Watch the mix-ups: (1) hallucination (factual error) vs bias (unfair skew) differ. (2) Fluency and confidence do not guarantee correctness. (3) Many risks are mitigated by human review, grounding, and data governance—"use it correctly," not "do not use it."
1.3.3Section summary
- Hallucination = plausible but wrong; mitigate via grounding/RAG and human review
- Understand bias / knowledge cutoff / uncertainty / explainability limits
- Privacy, copyright, and security (prompt injection) are governed via data governance
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. What is it called when generative AI confidently outputs plausible but factually wrong content?
Q2. What risk of unfair or discriminatory output arises from skew in training data?
Q3. What limitation describes a model being unaware of events after its training data?
Q4. Which best mitigates the risk of hallucination?
Q5. When using generative AI at work, what is the main risk of including sensitive personal data in inputs?
Q6. Which is the correct understanding about generative AI output being fluent and confident?