If I had a dollar for every time a stakeholder asked me if their model had "zero hallucinations," I’d have enough to fund supermind ai my own compute cluster. In the 11 years I’ve been working in NLP, the industry has shifted from simple rule-based parsers to probabilistic engines that dream up reality as they go. Today, we’re evaluating models like OpenAI’s GPT-4o and Anthropic’s Claude 3.5 Sonnet, but the conversation is still stuck on binary labels: Is it "right" or is it "wrong"?
The distinction between hallucination vs. factuality is not just semantic pedantry; it is the difference between a system that helps your analysts and one that loses you your job. Let's peel back the layers of these metrics, why they matter, and why you should stop obsessing over a single leaderboard number.
Defining the Metrics: What Are We Actually Measuring?
Before we look at the data, we have to define what these metrics actually represent. In my experience building model QA workflows, people often conflate these terms because they want a single "trust" score. They are, however, testing fundamentally different failure modes.
1. Hallucination Rate
Definition: The hallucination rate is a measure of "intrinsic" error. It quantifies how often a model generates information that is unsupported by the provided source context. It is essentially a measure of adherence.
2. Factuality Score
Definition: A factuality score measures "extrinsic" alignment. It evaluates whether the model’s generated output matches objective reality—the world as it exists—regardless of whether that information was in the source text.
Note: Most of you are failing to distinguish between these two because your evaluation pipelines don’t separate "grounding" (did you read the document?) from "knowledge" (do you know the capital of France?).
The Data Breakdown: A Comparative View
To understand the delta, let’s look at how different testing methodologies capture these nuances. Please note that I am omitting specific model-by-model performance numbers here because the "truth" shifts every time a model provider updates their system prompt or fine-tuning strategy.
Metric Primary Failure Mode Use Case Hallucination Rate Confabulation/Over-reliance on internal weights RAG (Retrieval-Augmented Generation) Factuality Score Outdated knowledge/Misinformation General-purpose Q&A or ChatbotsSo what: If your RAG system has a low factuality score but a high hallucination rate, your retrieval pipeline is fine, but your model is "lazy"—it’s ignoring your documents and answering from its pre-trained memory.
Summarization Faithfulness vs. Knowledge Reliability
This is where things get messy. When we talk about facts benchmark meaning, we often confuse how a model handles an input (summarization) versus how it recalls an entity (knowledge).
Summarization Faithfulness
In a summarization workflow, the model is a processor. If the source says "The revenue grew by 5%" and the model says "The revenue grew by 50%," that is a hallucination. It doesn't matter if the real-world revenue was 50%; within the context of the summary, the model has failed its faithfulness requirement. This is the bedrock of what companies like Suprmind focus on when building enterprise-grade extraction tools—ensuring the model stays within the "fenced area" of the user's provided data.
Knowledge Reliability
Conversely, if a model is answering "Who is the CEO of Company X?", it isn't summarizing; it is accessing its internal parametric memory. Here, the hallucination rate is a function of the model’s training cutoff and its tendency to play "fill-in-the-blank" when it doesn't know the answer. A model that refuses to answer when it doesn't know is objectively superior to a model that "hallucinates" a CEO name, yet many leaderboards penalize refusal as an error. This is a fundamental flaw in how we rank intelligence.
The Fallacy of the Single Leaderboard
I see it every day: teams picking a model because it topped a specific leaderboard. This is dangerous. Benchmarks are not "The Truth"; they are static follow this link snapshots of a specific distribution of questions. When you treat a benchmark as the ultimate authority, you ignore two critical vectors:
Refusal Behavior: A model might have a higher hallucination rate because it is more willing to attempt difficult queries. A "safer" model might just refuse to answer. If your use case requires an answer, the "safer" model is actually the lower-performing one for your specific workflow. Tool Access: A model’s factuality score is drastically different if it has access to a search tool or a calculator. Evaluating a model "blind" is useful for academic purposes, but it’s irrelevant for production enterprise deployment.Blunt note: Stop ignoring tool access in your evaluations. If your production setup uses a search tool, stop measuring the model as if it’s a stranded hiker in the woods. You are testing your architecture, not the model's raw IQ.
Mitigation: The Real Goal
If you take anything away from this, let it be this: Hallucinations are unavoidable. They are a structural byproduct of the transformer architecture. If you are building an LLM-based application, your goal should not be the eradication of hallucinations, but the creation of an observability layer that catches them.
Here is the reality of the current state https://dibz.me/blog/how-to-run-a-question-through-multiple-ai-models-at-once-1172 of the art:
- Anthropic models are often cited for their nuance in "refusal" triggers, which lowers hallucination rates by forcing the model to admit ignorance. OpenAI models demonstrate high factuality in general knowledge but require strict prompt engineering (or system instructions) to prevent "leaking" information from their training data into a RAG summary.
So what: Stop trying to build a perfect model and start building a better system wrapper—one that verifies citations, cross-references internal databases, and flag high-uncertainty outputs for human review.
Cross-Benchmark Reading: How to Actually Evaluate
If you want to move beyond the hand-wavy claims of "near-zero hallucinations," you need to adopt a cross-benchmark mindset. Do not rely on one test. You need a mix of:
- Faithfulness Metrics: Does it stay in the source? (e.g., G-Eval or RAGAS). Knowledge Metrics: Does it have current, accurate entity data? (e.g., MMLU or custom domain-specific Q&A sets). Refusal/Safety Metrics: Does it know when to shut up?
By triangulating these, you create a "Performance Profile" rather than a "Rank." A model with a 90% factuality score might be a liability if it has a 0% refusal rate, whereas a model with an 85% factuality score and a 15% "I don't know" rate might be the engine your legal or medical department actually needs.
Conclusion
The quest for a single "Factuality Score" is a ghost hunt. The industry is maturing, and the winners won't be the ones with the best raw training data, but the ones with the most robust evaluation pipelines. You must hold your LLM providers accountable by asking for their refusal rates, their grounding adherence, and their performance with tool-augmented workflows.
Don't be the person who chooses a model based on a leaderboard that was updated three weeks ago. Be the person who understands what the model is actually doing when it stares at your data and, occasionally, makes something up. Build for failure, build for verification, and stop asking for "zero hallucinations." It’s not going to happen, so design for the world as it is, not as you wish the neural networks to behave.
