This sounds very useful as way to quickly dig through data when doing research... then you can always go back and verify the model's conclusions later if you need to.
What makes me nervous though, is that I can easily see this becoming a scary consumer product people would trust at face value. A la, a Google Quick Answer box on steroids that's more difficult to verify.
We already have this problem, and Google's "search" widgets (e.g. dictionary, snippets, Q&A) are probably a big offender. Falsehoods are everywhere, but people tend to arrive at the truth, and once they do, the truth tends to go viral and flood fill over the falsehoods. What scares me is how easy it would be to create higher-order falsehoods. For example, X, Y, and Z are all false, but corroborate each other, making all of them more believable. These "regions" of falseness already trap millions of people today into throwing away their money -- think pyramid schemes, cults, etc.
I think you're absolutely right that this only going to be exacerbated.
Google could resolve this will well thought out UX.
Each answer could have a "show working" link which would display extracts of the tables used to derive the result, with the specific data used highlighted.
They could also have a "show a different result" link which would display the next most likely result with a substantially different answer. The user can then compare both methods and decide if Google probably has it correct.
How does this relate to Google Sheets "Explore" feature I've used it couple times and then got disappointed in one or two ways. But otherwise it helped me to quickly plot and analyse 3k rows.
What makes me nervous though, is that I can easily see this becoming a scary consumer product people would trust at face value. A la, a Google Quick Answer box on steroids that's more difficult to verify.