What the hell is a classifier and what is Zentropy doing with respect to classifiers?

Sure, absolutely.

So a classifier is just a machine learning system that sorts content for you based on some standard, right? You can be trying to identify whether images contain dogs.

And we're in the middle of a shift about how classifiers can work.

Classifiers have been part of online safety for a very long time.

People use them in the safety context.

They are arguably part of all of the sort of search and feed ranking that you use as well.

How did your classifier work in that domain? What are the sorts of things you're looking for? And what level of accuracy were you actually able to have with respect to identifying attempts by users to try to, let's say, denudify a celebrity's image?

So is it fair to say you think humans are not good at content moderation?

Yeah.

I think humans are not good at high degrees of precision at large scale.

Kind of in anything.

Like, I don't know.

Can you draw a circle freehand? I can't.

Like, my mom can.

And I'm going to hold the second question because it'll follow naturally from your answer to the first question.

(laughs)

That had become a bit of a football, um, between the US and the UK, and the US responded with pressure and the British eventually backed off.

And this reporting is that it's sort of bubbling back up again, although this time, there's a couple of interesting differences.

One, the request is allegedly scoped only to British users of Apple services instead of everybody in the world.

And two, the article ends sort of somewhat ominously.

It's actually worth reading verbatim.

"Members of the US delegation raised the issue of the request to Apple around the time of Trump's visit, according to two people briefed on the matter.

Yeah.

Mm-hmm.

Even on top of that, though, the changes they made around hate speech, and there's two coupled ones I, I think are, are pretty significant and, and quite worrisome.

So one, they're moving away from proactive attempts to detect whether or not things might be hate speech.

Just across the board they seem to be turning down, uh, those classification systems.

That, they're justifying that by saying it's going to lead to fewer false positives.

That, that is true, right? If you stop looking, you will make fewer-

Mm-hmm

... that precede that violence in a way that, to me, was, like, usable and specific, such that you could turn it into an actual set of standards you could hope to scale, which it's, I guess, been a little implicit in some of what I've said.