Measuring AI honestly: what benchmarks do and do not tell you

Why leaderboards mislead

A high benchmark score is easy to over-read: tests can be narrow, gameable, or unrelated to your workflow. A model can top a leaderboard and still fail on the decisions you actually care about, because those were never what the test measured.

Treat scores as hypotheses to check, not conclusions to trust.

Measure your decisions, not the model's fame

The honest measure is how the system performs on your real, consequential decisions, evaluated the way you would defend them. That is harder to produce and far more predictive of value.

Publish an honest scoreboard

State plainly what you have proven, what is still in build, and what you have not done. An honest scoreboard builds more trust than an inflated one, and it surfaces the real work instead of hiding it.

What an honest scoreboard looks like

An honest scoreboard states three things plainly: what has been measured and proven, what is still in build, and what has not been attempted. It resists the temptation to present aspiration as achievement, because being caught doing so costs more trust than the inflated number ever bought.

Counterintuitively, showing what you have not done builds more confidence than claiming everything works. It signals a vendor who will also tell you the truth when something goes wrong.