Why AI rewrites lose figures, and how we check for it
October 5, 2026 · 5 min read
When a model is asked to make text sound different, it changes what is easiest to change. Unfortunately the easiest things to drop are the ones that carry the most information: a percentage, a unit, the name of a study, a careful "may".
The damage is quiet. "Was associated with a 14.3% reduction" becomes "cut by 14%". The sentence reads better and is now wrong twice: the figure is rounded and an association has become a cause.
Samesay compares every version with your draft before you see it. It extracts numbers with their units, names, negations and qualifiers, then lists anything that went missing or changed. Where ten rounds of paraphrasing drift further, several versions are ranked by that check first.
No automatic check is perfect, so ours is strict on purpose. If you see a flag that is not a real problem, ignore it. If you see none, still read the text once. It is your name on it.