A model can invent a reference, misstate a number, or add a claim that the supplied document never made. These failures are related but different: a statement may be true in the world yet unsupported by the given source, or faithfully repeat a source that is itself wrong.
In episode 54, at 3:30, Dan Klein argues that fluent, plausible output is insufficient for enterprise reliability. His “plausibility engine” framing describes the deployment problem; it does not mean models are incapable of producing correct facts.
Example: a support answer cites the correct returns page but invents a fee. Inspect the actual passage and compare each important claim with it. A citation’s presence is not verification.
Mitigations include better evidence selection, constrained tasks, appropriate abstention, and output checks. Their effectiveness depends on the application. Evaluate faithfulness alongside correctness and answer relevance, and use the explainer to separate model and data failure modes.
Sources
- Episode 54: Dan Klein on enterprise reliability — Primary interview on the gap between plausible output and a dependable result.
Go deeper
- Are AI hallucinations a data problem or a model problem? AI, decoded · Are AI Hallucinations a Data Problem?
- Are AI hallucinations always bad? AI, decoded · When AI Hallucinations Are Good (and When They're Dangerous)
- Should you use prompting, RAG, or fine-tuning to customize an AI model? AI, decoded · Fine-Tuning vs. RAG vs. Prompting: How to Choose
- Vector database or knowledge graph — which should you use for AI retrieval? AI, decoded · Vector Database vs. Knowledge Graph for AI Retrieval
From the conversation
-
Why LLMs Are Plausibility Engines, Not Truth Engines | Dan Klein, Scaled Cognition -
Hallucinations Are a Data Architecture Problem | Sudhir Hasbe, Neo4j -
Practical Lessons for GenAI Evals | Chip Huyen & Vivienne Zhang -
Why Context Alone Isn't Enough for Enterprise AI Agents | WisdomAI CPO Kapil Chhabra -
Mastering Multi-Agent Systems | MongoDB’s Mikiko Chandrasekhar -
AI's Two Extremes – Foundations & The Frontier | Databricks’ Denny Lee -
Why Enterprises Need a Different Approach to AI Agents | Lyzr’s Siva Surendira -
GenAI Predictions for 2025 | Databricks & Cohere