How to Read a Benchmark Headline

Headlines about AI often follow one template: new model beats humans at X. Here is a short checklist for reading them.

Which humans? Experts, average participants, or the authors' colleagues on a Friday afternoon?
Which test? Was the test designed for people and borrowed, or designed for machines? Could the answers have leaked into training data?
Which conditions? Time limits, tools, number of attempts. A system allowed many tries is answering a different question.
Who says so? The company that built it, or someone who checked?

None of this means the result is wrong. It means the headline is the first sentence of the story, not the last.