Your RAG Product Is Confidently Making Up Answers 98% of the Time

The dirty secret of enterprise RAG isn’t wrong answers — it’s that your system almost never admits it doesn’t know. A new benchmark shows the worst commercial RAG system fabricates a response from thin air on 98.1% of questions it has no business answering. That’s not an edge case. That’s a liability.

What happened

Researchers from this paper built a penalty-aware evaluation framework to expose a flaw hiding inside standard Retrieval-augmented generation benchmarks: volume-based accuracy scores reward systems that guess constantly, because answering everything outscores strategic abstention. Their fix has three parts — asymmetric scoring (+1 correct, -4 wrong, 0 abstain), “knowledge-gap canaries” (questions verifiably absent from the knowledge base, so any answer is pure hallucination from parametric memory), and a failure-attribution pipeline separating retrieval failures from generation failures from abstention-policy failures. They ran three commercial RAG systems plus a no-retrieval baseline against 1,000 SimpleQA-Verified questions × 3 repeats, graded by a cross-family three-judge panel hitting 98.9% unanimity — a solid LLM-as-judge setup. The headline finding: when systems do answer, accuracy clusters tightly at 97.0–98.0% across all products. But canary violation rates — how often a system invents an answer when its knowledge base contains nothing — range from 16.7% to 98.1%, a sixfold spread. Faithfulness vs. groundedness is the real differentiator, and penalty-aware scoring reshuffles the product leaderboard accordingly, stably across penalty multipliers from k=1 to k=9.

Cold read

Three commercial systems is not a market census — the paper doesn’t name them, so you cannot act on the ranking directly, and selection bias in which vendors agreed to be tested is real. The canary design is clever but artificial: real knowledge bases have fuzzy boundaries (partially relevant documents, outdated chunks), and a verifiably-absent question is a cleaner test than the messy in-between cases your users actually hit. The 98.9% judge unanimity sounds impressive, but we’re told nothing about how the golden dataset was constructed or how “verifiably absent” was operationally confirmed — that’s the load-bearing assumption of the whole framework. The penalty ratio of -4 is a design choice, not a law; different enterprise contexts (medical vs. e-commerce) have wildly different wrong-answer costs, and the paper’s claim of stability from k=1 to k=9 is reassuring but doesn’t cover the asymmetric real-world cases where wrong answers cause regulatory or reputational harm orders of magnitude larger. This is a useful diagnostic tool, not proof that any specific commercial product is safe or unsafe for your use case.

What it means for you

• Signal maturity: 3/5 — rigorous methodology, small vendor sample, no product names
• Who gets hurt: Any SaaS company selling a RAG-powered assistant as a “trusted source of truth” — legal tech, HR, compliance, customer support — where a hallucinated answer triggers a downstream decision
• What breaks if this is true: Your vendor’s benchmark score is meaningless as a proxy for production safety; you need your own canary test suite before signing an enterprise contract
• Why it might not land: Most buyers still don’t penalize vendors for confident wrong answers — they penalize for “it didn’t answer my question,” which incentivizes exactly the reckless behavior this paper flags
• Watch for: Enterprise procurement teams adding abstention-rate and canary-violation clauses to RAG vendor RFPs — that’s the signal this framework is crossing from academia into purchasing criteria

Forecast as of 2026-08-29

By Q3 2027, at least one major RAG platform (Glean, Guru, Microsoft Copilot, or a similarly scaled enterprise product) will publicly release abstention-rate metrics alongside accuracy metrics in their product documentation — driven by procurement pressure from regulated-industry customers, not voluntary transparency.

Source: Why RAGs Hallucinate: Penalty-Aware Evaluation of Retrieval-Augmented Generation Systems with Knowledge-Gap Canaries — Alden Do Rosario, Hussein Younes, Felipe Pires. https://arxiv.org/abs/2608.26385v1