I think of it in a different perspective: You combine humans and programmers to do better than each could do alone.

Start all-manual. Perhaps you only do it with a subsection of applications. Pay an extra fee and you get "certified" with special app placement. Then you start all-manual. You look for the people who are the best at finding issues, and pair them with programmers, and make the tools for the things that are gruntwork for them.

Build more and more tools, and you pull more and more people into the program as you build more and more intelligence into the machine.

Let the humans do the NP-hard portions. I'm sure this is what Apple has to be doing behind the scenes.

Tasks of classifying things (in this case into "approved" or "rejected") that humans can routinely do but machines find difficult are areas where ML shines.

Human reviewers today, but once the training set is large enough you can start to let computers take over with human reviewers reviewing the lower certainty cases until the certainties rise further.

I bet we first hear about ML shining at generating malicious code obfuscated as legit.

But making legit apps does take human effort.

If an app takes days to make, requiring 5 minutes extra review effort to get it whitelisted seems fine.

Only if similar numbers of people work on each side. With thousands or millions of developers to a handful of reviewers, it may go as well as Google's other attempts at human support (read: terrible).

Charge money for that. If an app takes days to make, then you can also to pay for an hour of reviewers work if you want to distribute it to the public.

A non-prohibitive barrier to entry wouldn't be a big issue, as users in general want stricter filters that mean less shitty apps instead of looser filters that mean simply more apps.

They already do manual review for many of their high-serving ads, and people shouldn't shy away from some human intervention in these processes.

AI and machine learning are most effective these days when they help assist people (flagging potentially malicious code, bubbling up anomalies, etc.), and it isn't that expensive to get a pair of eyeballs to double check conclusions!

I'd love to see an ad system with manual review for all ads... and have a review fee when ads are changed/cycled.

It actually can be automated to a quite some degree. This is basically a something called expert system. You could create automated system to do these reviews from human reviewers and their expert domain knowledge. It might never be 100% accurate and might require human intervention from time to time but a high level of automation of this process can be achieved.

Current Machine Learning techniques rely on being trained by manually generated data. Google's virtual Assistant learned to understand voice because Google setup Google 411 and had millions of participants train it. Google developed OCR models by setting up Google reCAPTCHA and having millions of people train it.

Manual prompts to users are a great way to develop training data and being able to distribute them at scale to millions of users means you can develop training data in very short order.

On the contrary, the lack of human review is why both the Play Store and the Chrome Web Store (especially this one) are dumpster fires with rampant malware.

Automated tasks are not good at outsmarting humans. When you want to review a human's work for security, you need humans somewhere along the process.

Automation can help those humans do their jobs, but it's simply not a solution here.

You don't need to be that skilled, do you? Isn't this something you can train thousands of people to do over the course of a couple months?