Even if it’s “only” 1 million that would be a math task. Random sampling is the best we can do.
There are psychological books on identifying signs of prepubescence based on facial and genitalia features that one can search for if they are in that line of work. Some of the former Facebook mods with PTSD know what I am referring to.
Leave everything else to manual flagging assuming Grok has a flag or report button that is easy to find. If not send links to these people [1] if in the USA.
[1] - https://www.ic3.gov/
1) Zero is basically never the best error rate, effort isn't infinite and spending too much of it on one defect ends up meaning spending less on other issues.
2) Look at what he's saying. This is a classic pattern for providing a fake proof of evil.
a) Point to evil. For example, CSAM
b) Expand the definition of that evil in ways that are often not even evil. Here, include scantily clad in your definition of "sexual". Note that swimsuits qualify.
c) Point to examples of evil in your expanded pool.
d) Claim this points to evidence of the original definition. Note that nothing about their claims precludes their "CSAM" being nothing more than ordinary beach or pool scenes. Their claim includes the null and when the null is a possible answer it should be assumed.
I agree that a zero error rate is generally not possible, although I think a company like Xitter can manage better than 101 in 20k.
Title of the study and article says that Grok "Generated", but in fact:
> The CCDH then extrapolated
Basically they invent numbers.
They took a sample of 20k generated images, and it is assumed (but I don't know if the source is reliable) that Grok would have generated 4.6 millions image at the same time. So the sample is 00.4%.
If you see the webpage of the CCDH it is a joke their study. First:
First they invent their own definition, then adequately mixup possible "adult" pictures to give scary numbers.