Hallucination is a systems symptom, not a model defect

When a production AI system states something false, the reflex is to blame the model and wait for the next release. But when I trace hallucinations in deployed systems, the root cause is rarely "the model made something up out of nowhere." It's the pipeline: retrieval that returned nothing relevant, a knowledge base that went stale past the model's cutoff, a prompt that never grounded the claim, and — most damning — no path for the system to say "I don't know." Hallucination in production is a systems symptom, and it responds to systems engineering.

The taxonomy makes this visible. Researchers distinguish factuality hallucination (output contradicts world knowledge) from faithfulness hallucination (output contradicts the provided source) — and in RAG systems, the second kind dominates. The documented RAG failure modes are all pipeline failures: blind retrieval that over-triggers and injects irrelevant chunks; the generation bottleneck, where retrieval fetched the right passage and the model ignored it anyway; knowledge conflict, where retrieved context contradicts the model's parametric memory and the wrong side wins; and retrieval source contamination, where the corpus itself carried the falsehood. None of these is fixed by a smarter model. Each is fixed by an engineering intervention — better retrieval triggering, chunk-quality monitoring, context-response consistency checks — and worse, agentic systems compound them: cascade hallucination means one fabricated intermediate result flows into the next tool call and hardens into "fact" three steps downstream.

Engineer the pipeline, not the wish

The mitigation literature reads like a systems catalogue, not a modelling one. Grounding: retrieve before generating, and monitor whether the response actually stays consistent with the retrieved context. Verification layers: Chain-of-Verification prompting generates verification questions and answers them independently of the original draft — the verification-independence principle, which is just the auditor's rule that the checker must not trust the checked. Post-hoc approaches like RARR retrieve evidence after generation and revise unsupported claims. Neurosymbolic patterns push further: let the LLM synthesize and a deterministic engine verify — symbolic execution checking generated code, a solver checking generated math. And crucially, abstention: knowledge-boundary awareness and "I don't know" training exist precisely because a system with no refusal path converts every retrieval failure into a confident fabrication. A calibrated refusal is not a worse answer than a hallucination; it is the only correct one available.

The concession, honestly made: some hallucination genuinely is model-level. Long-tail facts the model saw twice in pre-training, imitative falsehoods absorbed from confidently wrong training text, sycophancy amplified by RLHF — these live in the weights, and the inverse-scaling results on truthfulness show that bigger doesn't automatically mean more honest. Better models do lower the base rate, and decoding-level interventions like DoLa and context-aware decoding operate on the model itself. My claim is narrower and more useful: the hallucinations that hurt you in production — the ones your users see, on your data, about your domain — are overwhelmingly the ones your pipeline caused or failed to catch. You don't control the pre-training corpus. You control retrieval quality, grounding discipline, verification layers, and abstention paths.

This reframing changes what teams do on Monday. Instead of a line item that says "upgrade model and re-test," the backlog says: instrument retrieval hit quality, add a consistency check between context and response, define when the system must abstain, and put fabrication cases into the eval suite that gates releases. Hallucination stops being weather — something that happens to you — and becomes a defect class with owners, causes, and regression tests. That's what it means to treat it as a symptom: you don't shout at the fever, you find the infection.