Sure you can, random sampling should work. Don't just go making things up.

Of course actually carrying out that experiment would be absurd since I don't think anyone expects an appreciable percentage of clearnet material to be CSAM. The working assumption is that the goal is to find a needle in a haystack so GP's objection about needing to know the false negative rate is misguided.

I expect the equivelent of the fbi is investigating this using other sourcs and so has plenty of data without needing to randomly sample any non-suspect conversation. CASM has been a problem since before computers.