Hallucination Report

Origin & evolution

The term entered NLP through image captioning and translation research, where models described objects that were not in the input. The framing that stuck came from Ji et al.'s Survey of Hallucination in Natural Language Generation (ACM Computing Surveys, 2022), which split the phenomenon in two. An intrinsic hallucination contradicts the source the model was given; an extrinsic one cannot be verified from the source at all, true or false. The distinction matters operationally, because the two demand different fixes: intrinsic errors point at how the model uses context, extrinsic ones at what it should refuse to assert.

Later work refined the axis into factuality versus faithfulness, separating disagreement with the real world from disagreement with the provided input. Recent surveys, including a 2025 taxonomy paper, formalise these definitions further. For a product team the value is practical: a report tagged by type routes to the right remedy and feeds a measurable eval rather than a vague sense that the model "makes things up".

How it works in practice

A support assistant tells a user a refund window is 60 days when the policy document loaded into context says 30. A reviewer files a hallucination report: type intrinsic, because it contradicts the supplied source; severity high, because it misstates policy to a customer. Twenty similar reports accumulate over a month. They become a regression set in the next eval_run🎟️Eval RunAI & Machine LearningAn evaluation run against a benchmarkView reference →, and the failure rate on that set drops from 8 percent to under 1 after a retrieval-grounding fix. The report closed the loop from a single bad answer to a guardrail and a tracked metric🔢MetricStrategyA unified metric that measures progress, health, or behaviour across the productView reference →.

Hallucination report vs. its neighbours

• AI guardrail🚦AI GuardrailAI & Machine LearningA guardrail for AI safetyView reference → is the control that prevents or catches bad output. The hallucination report is the evidence that a guardrail is missing or leaking; reports motivate guardrails, and guardrails are tested against the reports.
• Eval run🎟️Eval RunAI & Machine LearningAn evaluation run against a benchmarkView reference → is a systematic measurement across a test set. A hallucination report is a single logged case; a stream of reports becomes the test set an eval run scores against.
• AI model🤖AI ModelAI & Machine LearningAn AI or ML model used within the productView reference → is the system that produced the output. The edge ai_model_flagged_by_hallucination_reportAI Modelflagged byHallucination Reporthierarchy attaches the failure to the specific model and version, so a regression after an upgrade is traceable to the change that caused it.

In the graph

In the Unified Product Graph, a hallucination report sits in the AI region and attaches to its source through ai_model_flagged_by_hallucination_reportAI Modelflagged byHallucination Reporthierarchy. Logging each instance as a node keeps a model's failure record durable and queryable, so reports can be counted by type, fed into an eval_run🎟️Eval RunAI & Machine LearningAn evaluation run against a benchmarkView reference →, and used to justify an ai_guardrail🚦AI GuardrailAI & Machine LearningA guardrail for AI safetyView reference →. The structure turns scattered "the model got this wrong" complaints into an auditable trail from observed failure to measured fix.