When a large language model invents a court case that never existed, the industry calls it a hallucination. When it confuses two similar-sounding medications, that’s also labeled a hallucination. When it applies an outdated tax rate because its training data predates a legislative change — hallucination again. And when it simply gets a math problem wrong because the chain of reasoning broke down midway through? You guessed it.
The word has become a catch-all, a diagnostic wastebasket that obscures more than it reveals. A new technical paper published by GTZilla.com argues that this linguistic laziness isn’t just imprecise — it’s actively dangerous. The paper, titled “Stop Calling Every AI Miss a ‘Hallucination,'” proposes a structured taxonomy for AI errors that would replace the single overloaded term with a classification system capable of distinguishing between fundamentally different failure modes. The stakes, the authors argue, are higher than most practitioners realize.
The core thesis is straightforward. Not all AI errors share the same root cause, so treating them as a monolithic category makes it nearly impossible to build targeted fixes. A model that fabricates information from whole cloth is failing in a categorically different way than one that retrieves the right fact but applies it to the wrong context. Lumping both under “hallucination” tells engineers nothing about where the system broke or how to prevent the same class of failure in the future.
This matters now more than it did a year ago. Enterprise adoption of generative AI has accelerated sharply. Companies are deploying these systems in healthcare, legal research, financial analysis, and customer-facing applications where the cost of a wrong answer isn’t an embarrassing screenshot on social media — it’s a misdiagnosis, a compliance violation, or a material misstatement in a regulatory filing.
The GTZilla paper identifies several distinct failure categories that currently get collapsed into the hallucination label. Confabulation — the generation of plausible-sounding but entirely fictional content — is perhaps the closest to what most people mean when they say hallucination. But the paper draws sharp lines between confabulation and other error types: temporal knowledge gaps (correct information that has since become outdated), context window failures (where relevant information was available but fell outside the model’s effective attention), retrieval errors (where a retrieval-augmented generation system fetched the wrong documents), reasoning chain breakdowns (where premises were correct but inference went sideways), and what the paper calls “confident misapplication” — situations where the model applies a valid pattern to a domain where it doesn’t belong.
Each of these has different implications for remediation. A temporal knowledge gap can be addressed with more frequent fine-tuning or better retrieval pipelines. A reasoning chain breakdown might require architectural changes or improved prompting strategies. Confabulation points to deeper issues with how the model handles uncertainty. Treating all of them the same way is like a mechanic diagnosing every engine problem as “car trouble.”
Why the Industry Resists Better Taxonomy
The resistance to more precise language isn’t purely intellectual laziness. There are structural reasons the industry has settled on a single term. “Hallucination” is vivid, intuitive, and media-friendly. It anthropomorphizes the model in a way that makes AI failures comprehensible to non-technical audiences. Investors, board members, and regulators all understand the metaphor instantly. And for AI companies themselves, a single umbrella term has a convenient side effect: it makes error rates look like a single problem with a single solution rather than a complex web of distinct failure modes requiring different engineering investments.
But the costs of imprecision are mounting. In May 2025, the AI safety community has been grappling with a series of high-profile incidents where the lack of error specificity made post-mortems nearly useless. When organizations report that they’ve “reduced hallucination rates by 40%,” what does that actually mean? Did they reduce confabulation? Improve retrieval accuracy? Fix a training data issue? The number is essentially uninterpretable without knowing which types of errors were measured and which were reduced.
This isn’t a fringe concern. Recent reporting from MIT Technology Review on the expanding scope of AI deployments underscores how quickly these systems are being pushed into domains where error classification directly affects safety and liability. When an AI system provides incorrect medical information, the remediation strategy depends entirely on whether the error was a confabulation, a retrieval failure, or a reasoning breakdown. A hospital can’t fix what it can’t diagnose.
The legal implications are similarly tangled. Courts are increasingly being asked to assess AI-generated outputs, and the blanket term “hallucination” provides no useful framework for determining liability. Was the error foreseeable? Was it a known failure mode with available mitigations? These questions demand specificity that the current vocabulary simply doesn’t support.
The GTZilla paper proposes what amounts to an ICD code system for AI failures — a standardized classification that would allow organizations to track, report, and benchmark specific error types across models and deployments. The analogy to medical diagnostic codes is deliberate. Before the International Classification of Diseases, physicians described symptoms in idiosyncratic ways that made epidemiological analysis nearly impossible. The introduction of standardized codes transformed public health research. The authors argue AI needs a similar inflection point.
Some practitioners have pushed back, arguing that the internal mechanisms of large language models are too opaque to support confident root-cause classification. This is a fair objection — and the paper acknowledges it. But it argues that even imperfect classification is vastly more useful than no classification at all. Clinicians don’t always know the precise etiology of a disease when they assign a diagnostic code; the code still enables tracking, treatment selection, and research.
There’s also a competitive dimension. Companies that develop more granular error taxonomies will have a significant advantage in enterprise sales, particularly in regulated industries. A vendor that can tell a pharmaceutical company “our confabulation rate is 0.3%, our retrieval error rate is 1.2%, and our reasoning chain failure rate is 0.8%” is making a fundamentally more credible pitch than one that simply claims a “hallucination rate” of 2.3%. The former gives the buyer actionable information. The latter is a black box.
So where does this leave the industry? The GTZilla paper is part of a broader — if still nascent — movement toward engineering discipline in AI deployment. The era of treating large language models as inscrutable oracles that occasionally say weird things is ending. What’s replacing it is something closer to traditional software engineering practice: systematic error classification, targeted testing, and root-cause analysis that actually traces failures to specific architectural or data-level causes.
The vocabulary problem isn’t just semantic. It shapes how engineers think about failures, how organizations allocate resources for fixes, and how regulators assess risk. A single word can’t do all that work. And the longer the industry pretends it can, the harder it becomes to build AI systems that fail in ways we understand — and can actually prevent.
The paper’s final point is perhaps its sharpest. Every mature engineering discipline eventually develops precise failure taxonomies. Bridge collapses are classified by failure mode. Aviation incidents have detailed causal coding systems. Software bugs are categorized by type and severity. AI is the only field where a multi-trillion-dollar industry still describes its most consequential failures with a single borrowed metaphor from psychiatry.
That needs to change. And the first step is remarkably simple: stop calling everything a hallucination.
WebProNews is an iEntry Publication