Discussing AI Risks

Option B is the professional explanation. It:

• Explains the mechanism — not just "it makes mistakes" but why: token prediction optimises for plausibility, not truth
• Names the risk precisely — "high confidence presentation of incorrect information" (hallucination)
• Proposes a mitigation — "define a verification step"
• Contextualises by domain — code, legal, financial domains where errors are especially costly

How to explain AI bias as a second risk:
"LLMs reflect patterns in their training data. If that data over-represents certain groups, languages, or perspectives, the model's outputs will too. This matters for: hiring tools (may encode historical biases), user-facing content (may work better in English than other languages), and recommendation systems."

The key communication skill: translate "the model hallucinates" into business risk language: "Without a human verification step, we risk publishing or acting on incorrect information presented with false confidence."

Option C is the professional answer. It provides three specific, actionable limitations with context. This is the standard structure for communicating AI limitations to stakeholders:

1. Knowledge cutoff — every LLM has a training cutoff date. Mitigations: RAG for up-to-date data, clearly displaying the knowledge date in UI.

2. Hallucination rate — the rate varies by task (high for obscure facts, lower for common knowledge). Mitigations: confidence scores (if available), retrieval grounding, human review for high-stakes outputs.

3. Domain and language bias — models trained on English-heavy data underperform in other languages. Mitigations: fine-tuning on target language data, evaluation in the target language.

The communication principle: pair each limitation with a mitigation or design choice. "This is a risk" with no proposed action sounds like an alarm, not an analysis.