AI & information

An AI averaged two groups. Did it count their different sizes?

Averaging group averages can give the wrong overall result. Keep group sizes, reconstruct totals, and check the denominator.

Two groups have average task times of 10 and 20 minutes. An AI summary reports 15 minutes overall. That answer is justified only under particular conditions. If the groups contain different numbers of tasks, giving each group equal weight changes the question.

Reconstruct a small example

In this invented example, group A contains two tasks averaging 10 minutes each. Group B contains eight tasks averaging 20 minutes each. Assume the groups are separate, all ten tasks are included, and both averages measure the same duration.

  • Group A contributes 2 × 10 = 20 minutes.
  • Group B contributes 8 × 20 = 160 minutes.
  • The combined total is 180 minutes for ten tasks.
  • The average per task is 18 minutes, not 15.

The larger group has more influence because the requested unit is one task. If you intentionally want the average of the two group-level figures, 15 answers that different question. State which quantity you mean before asking for a summary.

Ask for the quantities behind the answer

Give the AI each group size alongside its mean. Request the reconstructed totals, total count and final division. If group sizes are missing, ask it to identify the missing data instead of assuming equal groups.

Microsoft's weighted-average example uses SUMPRODUCT divided by SUM. You can independently check this small example with a calculator; no generated formula is needed to verify 180 ÷ 10.

Keep the limits visible

Rounded group averages produce an approximate combined result. Prefer original totals and counts when available. Do not combine medians with this calculation, mix minutes with hours, or count overlapping tasks twice. The arithmetic can be correct while the input groups still describe incompatible things.

Sources & further reading

  1. Microsoft — Calculate an average: weighted averages

Sources support the technical explanations. Examples and suggested checks are editorial guidance.