A data engineer's taxonomy of LLM hallucinations: where they come from, how to classify them, and what each type implies for system design.
Hallucination is not a bug -- it is a consequence of how language models work. A model generates the next token by sampling from a probability distribution conditioned on all prior tokens. It has no independent access to ground truth; it has only the statistical regularities in its training data and the content of its context window. When those regularities conflict with reality, or when the model is asked about things poorly represented in training, hallucination is the predictable result.
The taxonomy matters for system design. Intrinsic hallucinations (the model contradicts information that was provided in the context) and extrinsic hallucinations (the model adds unverifiable claims that go beyond the context) have different root causes and different mitigations. Intrinsic hallucinations point to context engineering failures -- the information was present but the model did not attend to it correctly. Extrinsic hallucinations point to knowledge boundary failures -- the model was asked something it cannot ground from the provided material. Confusing the two leads to wasted mitigation effort.
Calibration is the property that matters most for building reliable systems on top of LLMs. A well-calibrated model that expresses uncertainty when it does not know is vastly more useful than a confidently wrong one. Chain-of-thought prompting, self-consistency sampling, and verbalized confidence all improve calibration in measurable ways. Build these into your prompting patterns before reaching for heavier external verification layers.
The data engineering implication is that you cannot test your way to zero hallucinations -- you can only reduce their rate and detect them faster. This means building eval harnesses that measure hallucination rate on your specific domain (see the Evals node), and designing product flows that route high-stakes outputs through human review. Treat hallucination as a distribution to manage, not a property to eliminate.
The definitive taxonomy. Classifies hallucinations as intrinsic (contradicts the provided context) vs. extrinsic (unverifiable against the context). Essential vocabulary for writing specs and incident postmortems.
The benchmark that forced the research community to take hallucination measurement seriously. Demonstrates that larger models can be more confidently wrong -- an important calibration for teams evaluating model upgrades.
Calibration research showing that models can often express uncertainty when prompted correctly. The empirical foundation for confidence-elicitation prompting strategies.
A practical framework for measuring hallucination rate at the atomic claim level. Adaptable as an automated eval step in your LLM pipeline.
New resources and perspective on building AI-ready data systems, a few times a month. No spam.