Every AI presentation tool ships a variant of the same disclaimer: please verify the output. It is honest, and it is useless. Nobody verifies forty claims at 11pm before a board meeting, which is precisely when these tools get used.
The constraint
Grounded mode inverts the default. Instead of generating freely and asking you to check afterwards, it generates against your document and refuses anything it cannot attribute. Each claim on a slide carries a pointer back to a character span in the source. No span, no slide.
A number that cannot be traced is not a rounding error. It is a different number.
Three categories of removal
Fabricated: the figure appears nowhere in the source. These are rare in modern models but catastrophic when they happen, because they are always plausible. Contradicted: the source says 34% and the draft says 40%. Almost always a paraphrase collapsing two nearby sentences. Derived: technically true, calculated from the source, but not stated in it: payback periods, run rates, per-seat economics.
The first two are removed silently and reported. The third is moved into speaker notes with its derivation shown, because it is usually the most useful thing on the slide and you should be the one deciding whether to say it out loud.
The report is the product
When a deck finishes generating, you get a chip: three unverified figures removed. Hover it and you see each one, the claim, and why it went. That report is not an apology for the model. It is the feature. The only way to build trust in an automated draft is to show your working.
Grounded mode is available on Pro and Scale, and it is on by default whenever you upload a document rather than typing a prompt.
