Check Percentages and Numerical Comparisons Before Editing AI Text
Check denominators, percentage points, relative change, baselines, and rounding with worked examples that keep numerical claims accurate through an AI rewrite.
A rewrite can leave every digit untouched and still change a numerical claim. “Up by 6%” means something different from “up by 6 percentage points.” Removing “of respondents” can quietly replace the measured group with a larger population.
Before editing a sentence with numbers, identify the measure, denominator, period, and calculation. Keep the source values next to the draft and recalculate independently. Record the result in a review card so later wording changes do not undo the correction.
Establish what each number represents
A bare 45 might refer to people, minutes, forms, or a percentage. Even 45% remains incomplete without an answer to “45% of what?” Collect these details before revising:
- The measure being described.
- The population or whole included in it.
- Its unit, such as people, hours, dollars, or percent.
- The observation period, which may differ from the publication date.
- The source location, including the row, table note, or passage.
If you cannot find the source, establish the evidence with the AI-text fact-checking workflow before polishing the numerical statement. Correct arithmetic performed on invented inputs still produces an unsupported claim.
Preserve labels as carefully as values. A monthly total is not an annual figure. A result among respondents does not automatically describe everyone invited. Restoring the digits alone cannot repair a changed population or time period.
Name the denominator
The basic calculation is:
Percentage = part ÷ whole × 100.
Statistics Canada's archived statistical guide illustrates the calculation. Use it for the arithmetic principle, rather than current evidence about a population.
Consider an invented survey with 200 invitations, 80 responses, and 48 respondents selecting an option. Those selections represent:
| Whole used | Calculation | What you may say |
|---|---|---|
| Respondents | 48 ÷ 80 × 100 = 60% | 60% of respondents selected the option. |
| Invitees | 48 ÷ 200 × 100 = 24% | 24% of invitees responded and selected the option. |
The second result does not show that only 24% of invitees preferred it. The preferences of nonrespondents remain unknown. The verb matters alongside the denominator.
“48 of the 80 respondents selected the option” makes the whole visible. Add 60% when that helps comparison. Avoid switching between respondents, eligible participants, and invitees simply to shorten the sentence.
Separate percentage points from relative change
Subtracting two percentages gives a difference in percentage points. Dividing that difference by the starting percentage gives a relative change.
For a rate moving from 30% to 36%:
- 36 − 30 = 6 percentage points.
- (36 − 30) ÷ 30 × 100 = 20% relative increase.
The ONS percentages guidance distinguishes these calculations. “Increased by 6%” misstates the relative change in this example.
A useful sentence is: “Participation rose from 30% to 36%, an increase of 6 percentage points.” Include the 20% relative increase only if it answers the reader's question. More calculations do not necessarily create a clearer explanation.
Work through a count and a rate together
The following is an invented teaching dataset, not a research finding, customer result, or AI-tool test. A fictional club uses the same attendance and eligibility definitions for both sessions.
| Measure | Session A | Session B |
|---|---|---|
| Eligible members | 100 | 125 |
| Members attending | 30 | 45 |
| Attendance rate | 30 ÷ 100 = 30% | 45 ÷ 125 = 36% |
An AI draft says: “Attendance improved by 6%, proving the new reminder worked.” It fails twice. “Attendance” does not distinguish count from rate, and the table supplies no evidence that a reminder caused the difference.
| Question | Working | Accurate description |
|---|---|---|
| How many more people attended? | 45 − 30 = 15 | There were 15 more attendees. |
| What was the count's relative increase? | (45 − 30) ÷ 30 × 100 = 50% | The attendee count increased by 50%. |
| How far did the rate move? | 36% − 30% = 6 percentage points | The rate rose by 6 percentage points. |
| What was the rate's relative increase? | (36 − 30) ÷ 30 × 100 = 20% | The rate increased by 20% relative to its starting value. |
A reviewed sentence can show the underlying comparison:
Attendance rose from 30 of 100 eligible members in Session A to 45 of 125 in Session B. The attendance rate increased from 30% to 36%, a rise of 6 percentage points.
This preserves the distinction between count and rate without inventing a cause. In real data, check whether eligibility, measurement, or the groups changed too. Arithmetic cannot make unlike observations comparable. For descriptions attached to figures, use the chart-description review guide as well.
Keep direction and starting value explicit
For ordinary change over time:
Relative percentage change = (new value − old value) ÷ old value × 100.
The US Bureau of Labor Statistics explanation uses the earlier value as the calculation's base.
In an invented count example, 80 becoming 100 is an increase of 20 ÷ 80 × 100 = 25%. A fall from 100 to 80 is a decrease of 20 ÷ 100 × 100 = 20%. The absolute difference is identical; the baselines differ.
Say “increase” or “decrease” and identify the baseline instead of using an ambiguous “percent difference.” With a zero baseline, the ordinary formula divides by zero and yields no finite percentage change. Give the endpoints or absolute change. Negative baselines and specialized index comparisons can need expert interpretation.
Round the display, not the evidence
Calculate using the most precise values you are authorized to use, then round for the reader. When only rounded published figures exist, do not invent additional precision. Mark a derived result as approximate or use the source's reported comparison.
ONS rounding guidance treats precision and readability as a choice shaped by purpose. Decimal places should help the reader make the intended judgment.
An invented example of three equal groups illustrates a familiar mismatch: each is one third of the whole, displayed as 33.3%. The displayed sum is 99.9%, although the unrounded shares cover the entire total.
Do not change a source value to force a tidy sum. Ofqual's rounding policy explains how rounded components can differ from totals calculated before rounding. Its particular conventions apply to its own publications; they are not a universal editorial rule.
Check missing categories, overlapping categories, and denominator differences before blaming rounding. Use a rounding note only when rounding actually explains the discrepancy.
Save the reasoning with the sentence
Create one numerical review card for each consequential claim:
- Copy the exact draft sentence.
- Identify the source, location, date, and relevant notes.
- List inputs with units, denominator, population, and period.
- Show the formula with values substituted.
- Decide whether to keep, revise, remove, or ask the source owner.
- Write the final sentence with necessary limits.
For the club table, the card would replace the ambiguous 6% statement with count and rate endpoints, then remove the causal assertion. That record explains why a later editor should not shorten it back into the original mistake.
After using grammar and style assistance, compare the final sentence with the card. Ask writing tools to preserve approved figures and labels, but recalculate yourself. The meaning-preserving editing checklist helps check changes to scope and certainty too.
Common questions
Can unchanged digits still express a different claim?
Yes. Swapping percentage points for percent, dropping the measured group, or changing “associated with” to “caused” can change the claim without altering a number.
Must I report both kinds of change?
No. Choose what the reader needs. Starting and ending rates often explain a rate comparison well; label any additional relative comparison explicitly.
Does accurate arithmetic establish importance?
No. It checks the calculation from the supplied inputs. Data quality, representativeness, causation, and statistical significance require the source's methods and sometimes specialist review.
