1 / 6
Working with hallucinations: detect, correct, cite
In this lesson
Working with hallucinations: detect, correct, cite
Evaluate LLM outputs for grounding and relevance by applying the Anthropic 4D Discernment framework [^1] and AWS Bedrock contextual grounding checks [^7], distinguishing between factually ungrounded claims…
You'll be able to
- Evaluate LLM outputs for grounding and relevance by applying the Anthropic 4D Discernment framework [^1] and AWS Bedrock contextual grounding checks [^7], distinguishing between factually ungrounded claims (hallucinations) and irrelevant-but-accurate responses in production RAG workflows.
- Apply three concrete verification techniques—web cross-check, source-grounded RAG with citation, and prompting the AI to cite its sources—to detect and correct hallucinations before deploying generated content, consistent with the exam task "Build the habit of verifying" [^1].
- Create a knowledge-base query workflow using AWS Bedrock that automatically surfaces source citations [^4] and leverages GraphRAG cross-document reasoning [^3] to minimize hallucinations in multi-source question-answering tasks.
- Articulate your own debugging strategy when an AI produces an ungrounded claim, making your reasoning explicit [^2] and documenting which verification method (web search, RAG retrieval, or citation prompt) you chose and why, mirroring the cognitive-apprenticeship practice of verbalizing tacit knowledge.
- Classify AI pedagogical roles (mentor, tutor, coach, teammate, student, simulator, tool) [^6] and select the role least prone to unchecked hallucination for a given learning task, recognizing that role framing directly shapes output reliability.