An insurance firm wants to use a generative AI model to answer customer queries about policy terms. The model frequently gives responses that are fluent but contradict the actual policy documents. Which mitigation technique directly addresses this limitation by anchoring model outputs to verified source material?
- AFine-tuning the model on a large corpus of general insurance industry articles
- BIncreasing the model's temperature parameter to encourage more creative responses
- CApplying retrieval-augmented generation so the model grounds its answers in retrieved policy documents Correct
- DSwitching from a transformer-based model to a rule-based expert system for all responses
Why A is wrong: Fine-tuning on broad industry articles updates model weights but does not guarantee outputs reflect the firm's specific policy documents; hallucination on specific policy details can persist.
Why B is wrong: Higher temperature increases output diversity and randomness, which would worsen factual accuracy, not improve it.
Why C is correct: Grounding via retrieval-augmented generation supplies the model with the authoritative source text at inference time, constraining outputs to verifiable content and directly countering hallucination on domain-specific facts.
Why D is wrong: A rule-based expert system is a valid approach for narrow decision trees but abandons the natural language capability the firm needs; grounding the existing model is a more targeted and practical mitigation.