- Published on
The Grounding Test
- Authors
- Name
- Dr. Paul Moon
A test I'd run before any AI-generated claim goes into a board deck, a customer email, or a forecast.
Take one sentence from the output and ask three questions.
What would have to be true in the world for this sentence to be right? Make it concrete: which customers, which invoices, which sites, which dates.
Who or what can check that, and have they? A database query counts. A phone call counts. The model's own confidence does not.
If the sentence is wrong, who finds out first, and how late? Some errors surface within the hour. Others surface at the next audit, or in a lawsuit.
The answers sort claims into three bins. Claims you can ground in minutes, which you ground. Claims you cannot ground, which you label as hypotheses. And claims where being wrong is expensive and late, which get a named human owner before they leave the building.
Most AI errors that hurt companies did not come from exotic failures. They came from fluent sentences nobody was assigned to check.
Which sentence in your last AI-assisted report did nobody check?