How to Check Percentages and Numerical Comparisons in AI Writing

“Participation rose by 6%” and “participation rose by 6 percentage points” can describe different changes. A fluent AI rewrite can swap the two without changing a single digit. Before polishing a sentence with numbers, identify what was counted, the denominator, the comparison period, and the calculation. Then write the result in language that preserves all four.

To check percentages in AI writing, keep the source figures beside the draft and recalculate the claim yourself. The practical output is a small numerical review card: source values, calculation, accurate sentence, and any remaining limitation. This guide shows how to build one with invented examples you can verify.

Start with the number's meaning

A number needs a label. “45” might mean people, completed forms, minutes, or a percentage. “45%” needs a whole: 45% of which group, measured when, using which definition?

Before editing, record:

• Measure: what the number describes.
• Population or whole: who or what is included.
• Unit: people, hours, dollars, percent, or another stated unit.
• Time: the observation period, which may differ from the publication date.
• Source location: the table, row, note, or passage containing the value.

If the source cannot be located, pause the numerical rewrite. Use the broader workflow for fact-checking AI-generated text to establish the evidence first. A correct calculation does not validate invented input data.

Keep the source's scope in the sentence. Results from respondents do not automatically describe all invitees, and a monthly count is not an annual total. If the draft changes those labels, restoring the original digits alone will not repair the claim.

Check the denominator before checking the wording

A percentage expresses a part relative to a whole:

Percentage = part ÷ whole × 100.

Statistics Canada's explanation of basic statistical techniques illustrates this calculation and why the base matters. Its archived guide is useful here for arithmetic, not as current evidence about any population.

Consider an invented survey with 200 invitations, 80 responses, and 48 respondents selecting an option. The option was selected by 60% of respondents: 48 ÷ 80 × 100. The same 48 people are 24% of those invited: 48 ÷ 200 × 100. These are different statements with different denominators.

You cannot conclude from these figures that only 24% of invitees preferred the option; the preferences of nonrespondents are unknown. You can say that 24% of invitees responded and selected it. Read the exact verb as carefully as the percentage.

Make the whole visible

Replace a vague phrase such as “60% agreed” with a sentence that names the group, such as “48 of the 80 respondents selected the option.” Add the percentage if it helps the reader compare. Do not silently switch between respondents, eligible participants, and everyone invited.

Distinguish percentage points from relative change

When two values are already percentages, subtracting them gives a difference in percentage points. To calculate a relative percentage change, divide that difference by the starting percentage and multiply by 100.

For a change from 30% to 36%:

• Percentage-point change: 36 − 30 = 6 percentage points.
• Relative change: (36 − 30) ÷ 30 × 100 = 20%.

The ONS guidance on percentages and percentage points distinguishes these two forms. “Increased by 6%” does not accurately describe the relative change in this example.

Use wording that shows both endpoints: “The participation rate increased from 30% to 36%, a rise of 6 percentage points.” If the relative comparison is relevant, add “a 20% relative increase.” Giving every possible calculation is unnecessary; choose the one that answers the reader's question and label it clearly.

A worked example: three changes in one table

The following figures are an invented teaching dataset, not a customer result, research finding, or test of an AI tool. Assume a fictional club uses the same eligibility and attendance definitions in two sessions.

A draft might say: “Attendance improved by 6%, proving the new reminder worked.” That sentence has two separate problems: it leaves “attendance” ambiguous and adds a causal conclusion the table cannot establish.

Recalculate each possible claim

A useful final sentence could be:

Attendance increased from 30 of 100 eligible members in Session A to 45 of 125 in Session B. The attendance rate rose from 30% to 36%, an increase of 6 percentage points.

This sentence lets the reader distinguish count from rate. It does not claim to explain why the figures changed. Before making a real-world comparison, also check whether eligibility, measurement methods, or the groups themselves changed. Correct arithmetic cannot make unlike observations equivalent.

Keep the starting value and direction straight

Relative change uses a starting value:

Relative percentage change = (new value − old value) ÷ old value × 100.

The US Bureau of Labor Statistics guide to calculating percent changes demonstrates the importance of using the earlier value as the base for a change over time.

In another invented example, a count rises from 80 to 100. The increase is 20 ÷ 80 × 100 = 25%. Moving from 100 back to 80 is a 20 ÷ 100 × 100 = 20% decrease. The same absolute difference produces different percentages because the starting values differ.

Avoid an unqualified “percent difference” when you mean change over time. State the direction and baseline. If the baseline is zero, the ordinary relative-change formula divides by zero and does not provide a finite percentage. Report the endpoints or absolute change instead. Negative baselines and specialized index comparisons may need a domain expert's interpretation.

Round for reading without changing the underlying data

Calculate from the most precise source values you are authorized to use, then round the displayed result appropriately. If only rounded published values are available, do not invent hidden precision; explain that your result is approximate or use the source's own reported comparison.

ONS guidance on rounding treats readability and precision as a tradeoff that depends on the reader's purpose. More decimal places do not automatically make a sentence more useful.

For a simple invented example, three equal groups each represent one third of a total. Displayed to one decimal place, their shares are 33.3%, 33.3%, and 33.3%. The displayed sum is 99.9%, although the unrounded shares cover the whole.

Do not alter a source value just to force a tidy total. Ofqual's rounding policy explains that rounded components can differ from totals calculated using unrounded values. Its specific rounding conventions belong to its publications; they are not universal rules for your draft.

Before attributing a mismatch to rounding, check for missing categories, overlapping categories, and different denominators. “Totals may not sum because of rounding” is appropriate only when that is the actual explanation.

Make one review card for each important numerical sentence

Copy this compact structure into your editing notes:

• Draft sentence: the exact claim being checked.
• Source and location: the original table or passage, including its date and notes.
• Inputs: values, units, denominator, population, and period.
• Calculation: the formula and substituted numbers.
• Decision: keep, revise, remove, or ask the source owner.
• Final sentence: clear wording with the necessary context and limitations.

For the club example, the decision is to replace an ambiguous 6% claim with the count and rate endpoints, and remove the unsupported causal explanation. The card records the reason, so a later editor does not “simplify” it back into an error.

If you use AI to suggest wording, ask it to preserve the approved values and labels and flag uncertainty. Recalculate independently rather than treating the tool's explanation of its own output as a check. After the style pass, compare the final sentence with the card using the meaning-preserving editing checklist.

FAQs

Can the digits stay the same while the numerical claim changes?

Yes. Changing “percentage points” to “percent,” dropping “of respondents,” or replacing “associated with” with “caused” can change the claim without changing a digit.

Should I always include both relative change and percentage points?

No. Include the comparison your reader needs. When reporting a rate change, showing the starting and ending rates often makes the sentence easiest to interpret. Label any additional comparison explicitly.

Does correct arithmetic prove a change is meaningful?

No. Arithmetic verifies the calculation from the stated inputs. It does not establish data quality, representativeness, causation, or statistical significance. Those questions require the source's methods and, where appropriate, specialist review.