Judgment in the Age of AI · Article 8
Numbers give an answer an objective appearance, yet errors can enter through units, definitions, formulas, rounding or source data. A plausible final result may come from the wrong denominator. Checking a number means reconstructing the calculation, not merely repeating its last step.
Example: what percentage was saved?
An item falls from $200 to $150. If $150 is used as the denominator, an AI may report a 33.3 per cent fall. The question is change relative to the original price, so the calculation is (200−150)÷200: 25 per cent. Both answers look mathematical, but they answer different questions.
Five boxes in a numerical audit
- Inputs: Where did the original numbers come from, and were they copied correctly?
- Definitions: What does each metric represent?
- Units: Have dollars, thousands, percentages and percentage points been mixed?
- Formula: Are numerator, denominator and time period correct?
- Plausibility: Is the order of magnitude sensible?
Make tools check one another
AI can explain a formula, but a calculator, spreadsheet or executable code should recompute it. For reports, manually sample several rows. If the result affects a budget, quotation or research conclusion, retain the input data and calculation steps.
A trustworthy number has not only an answer but a path back to its source data.
References
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.