What the AI changed
You gave a text to the chat and it came back rewritten: what has it actually changed? Here the comparison is word by word, and the numbers and dates that changed are on top, because that is what nobody checks. The texts never leave your browser.
You can also drag a .txt file into either box.
The first thing to check is the numbers
While rewriting, the model also touches what it should not: a percentage that becomes another one, a date moved by a day, an amount rounded off, a figure that disappears altogether. It does not do it out of spite, it does it because it is writing, not copying. Nobody rereads a thousand words by eye looking for that one number, so here numbers and dates get a box of their own, on top, before the text: the ones that changed, the ones that are gone and the ones that are new. And when none of them changed, that is written too, which is just as useful an answer.
Word by word, not line by line
A line-by-line comparison (what code tools do) on a text written in paragraphs always says the same thing: line removed, line added, that is, all red and all green. Useless exactly when you need it. Here the paragraphs that resemble each other are paired first, and inside each one it is the words: in a thirty-word sentence you see the two that changed. If a paragraph was rewritten from scratch and no longer resembles anything, it is shown removed and added, instead of pretending to find similarities inside it.
Cosmetic changes, counted separately
Models change punctuation on their own: straight quotes become curly, the hyphen becomes a long dash, three dots become a single character, and sometimes spaces appear that look like spaces but are not. If those counted as real changes, the text would be coloured all over and the two words that matter would be lost in it. Here they are counted separately and shown in grey. If you want them gone for good, there is Strip markdown and invisible characters.
How numbers are read (and the one ambiguous case)
«3.5» and «3,5» are the same number written two ways, and indeed they end up among the cosmetic changes, not the real ones. The only truly ambiguous case is a separator followed by exactly three digits: «1.234» can be one thousand two hundred and thirty-four (Italian style) or one point two three four (English style). Here it is read as a thousands separator, which is the more frequent case in written prose. Dates in figures («12/03/2025») count as one thing, otherwise the list becomes unreadable exactly where you need clarity.
What it does not do
It does not say who is right: it puts the two texts side by side, the decision is yours. And it does not understand meaning: if the model changed «may» into «must» you will see it highlighted like any other word, but only you know that this is the most serious change in the document. For two PDFs instead of two texts there is Compare two PDFs, for two lists Compare two lists, and if the text is too long for a single chat there is Split a long text for AI.