Skip to content
Beyond Prompt AI Studio
Comparing AI models properly

How AI performance is actually measured: the building blocks you need

"Model A beats Model B" sounds like a simple fact. To actually judge a claim like that, four building blocks are enough – no formulas, with hands-on examples.

Four examples – worth remembering

Try it yourself: the measurement paradigms compared

Tap a dimension to compare the measurement paradigms.

Objectively correct

One clear right answer, no interpretation needed

Judgment-based (human or AI)

The score depends on the judgment of the person or the judging AI

Building block 1: what is a benchmark, actually?

A benchmark is, at its core, nothing more than a questionnaire with known correct answers: a fixed set of test tasks where it's already clear what "correct" means. A model works through these tasks, and a scoring method counts how often its answer matches the known correct one. This simple principle is behind most "knowledge" benchmarks.

Building block 2: three fundamentally different ways of measuring

Not every benchmark measures the same way. Three types explain almost the entire landscape: with objectively correct benchmarks, there's a clearly right answer – a math problem, a multiple-choice question, whether a code fix passes the tests. With human preference, there's no "correct" answer; instead, people compare two answers and say which one they prefer. And with AI as judge, another language model scores the answer against a given rubric – used because it's far cheaper and faster than human evaluation.

Each of these three types has its own weaknesses, covered in detail in the following modules.

Building block 3: why sample size matters

A model answers 10 test questions, gets 8 right – "80%". But flip a fair coin 10 times, and you'll also land on 8 or more heads fairly often, purely by chance. Only with enough test questions – hundreds to thousands for real benchmarks – does a percentage become meaningful at all.

Building block 4: why a small difference is often not a real difference

71% vs. 73% sounds like a clear winner. But if the underlying sample is small or the task is hard, that gap can sit entirely within the range of chance. That's exactly what a confidence interval expresses, without needing to know the formula behind it: it says, in essence, "the true score is very likely somewhere in this range" – and if two models' ranges overlap, the claimed winner isn't actually established.

Why this matters to you as a decision-maker

These four building blocks are enough to critically read any benchmark number, even before knowing a specific benchmark in detail. The following modules look at the three measurement paradigms from building block 2 one at a time – and show where each of them runs into its limits in practice.

Key takeaways

  • A benchmark is, at its core, a questionnaire with known correct answers plus a scoring method.
  • Three fundamentally different ways of measuring shape the landscape: objectively correct, human preference, and AI as judge.
  • A percentage is only meaningful if the underlying sample is large enough – with too few test questions, a high score can be pure chance.
  • A small difference between two values is often not a real difference if the confidence intervals of both values overlap.
  • These four building blocks are enough to critically judge any benchmark number, even before knowing the specific benchmark in detail.

Quick check: did it land?

1 / 3

What three fundamentally different ways of measuring does this module name?

Want to critically judge benchmark numbers yourself going forward?