> These algorithms should probably be public. And any system which cannot be transparently explained should be shut down.

But what would even count as an explanation, and what does a "public" algorithm reveal?

There is no line of code that says "if (video.content == extreme) { show_to(EVERYONE) }".

A lot of the dangers are possible from a simple algorithm which merely performs A/B testing on whether a certain (randomly chosen, at first) video increases the amount of time a user spends on a site.

You could pass a law against A/B testing, or require companies to provide deliberately bad suggestions to make their users frustrated, but I'm not sure if that is a proportionate legislative response.

> But what would even count as an explanation

That sort of gets to the heart of it - saying "it's impossible to explain my algorithm because it is so complicated" no longer cuts it. It's the same as if I said to the FDA "This drug works in our testing for the indicated purpose but we have no idea if there are side effects" they aren't go to say "well, OK then, you can sell it". They'll say "come back when you have tested for all possible side effects and you can precisly define its behavior"

Currently Facebook, Google et al, will measure a narrow set of metrics to define the success of their algorithms (primarily relating to $$$) and then declare victory based on empirical optimisation of those metrics. They have to show that not only does the algorithm work for its stated purpose, it doesn't have undesirable side effects. They could do massive testing to demonstrate a huge range of side effects are not present. That would be similar to and probably as expensive as drug development. But an easier way to do that is to transparently explain what the algorithms do so that side effects can be predicted. If that is actually impossible for a given algorithm - well, maybe that algorithm shouldn't be used on the public at all.

If a company can't explain why a given user was shown a given video, at a given time, that system is faulty and should be prohibited. Which is to say, there's nothing particularly wrong with complicated rules and scoring systems making recommendations, but things like black box neural networks should not be able to be used for this stuff.

Let me give you an example of a good system: Spam filters not operated by Google. Algorithms decide whether or not mail gets through my corporate mail filter. It determines it based on a score, which is tallied from a set of rules. And you can drill down and see the score a spam email received, and then you can see the rules and factors that created that score. As a user, you can even generally see this, because the results are included in the message's headers in your inbox. And if you're the admin, you can then adjust those rules to fix errant behavior.

That's how a recommendation system should work. A system which can't be analyzed in that matter should not exist, and the rules and scoring applied to such systems should be disclosed in some sort of header.

The spam filter works because only you can see the scores of the emails. If the rules of Google search and Youtube were more transparent, spammers would always be one step ahead of spam detection, so everyone would lose in the end.

This is false. Spammers can and do test their own emails against spam filters. And the less transparent ones, like Gmail's often suffer longer failures to detect new strategies spammers are using.

There are enough unique filters that a spammer cannot hope to account for them all in a timely manner. Getting .001% of spam emails past Gmail's filters is much more efficient than getting 100% of emails past my personal Thunderbird spam filter.

That's a valid argument for not using Gmail, but that in the case of Youtube, almost everyone wants a single place where they can watch all the user-generated videos.

> almost everyone wants a single place where they can watch all the user-generated videos

I have my doubts this is true, so much as people are used to the relatively common controls/experience that comes with clicking a YouTube link. If I gave you a link to ocdtrekkietube.com/s8fsj2 (not a real link), and it worked as well as YouTube, I'm not sure any user would be particularly upset about it.

The biggest reason YouTube is as powerful as it is is that video hosting is expensive and few companies can eat that kind of bandwidth and storage without being an ad giant.

We're talking about Youtube's recommendation algorithms here, not their video hosting. Imagine if Youtube hosted the same videos, but provided no homepage, recommendations, or search engine. People would still gravitate towards a single source where they can watch all the content they want. There may be some more competition, such as a liberal recommendation engine, a conservative one, and one for teenagers. However, these websites/apps would likely face all the same problems that Youtube does today.

My iphone indexes photos and labels all the ones with dogs in it. I don't care how this system works and likely the answer is some black box NN. Why should this system be shut down when it causes no harm and works pretty well?

> Why should this system be shut down when it causes no harm and works pretty well?

You believe it causes no harm, but that doesn't make for a good assumption, considering image recognition NNs have been repeatedly demonstrated to be racist. So you may find it's classification of dogs might work really well, but it might label other people's coworkers or family members as dogs.

I still don't see why I should care since its only for my personal use. I agree in other areas like law enforcement and news curation this is an issue but for my personal photo search all I care about is that it works pretty well, which it does. Some mistakes are fine here.

Which is my main point, why should this apply to all tech when in some cases unexplainable mistakes are fine?

In many cases making a neural network explainable is literally impossible. Logic based AI failed in the 90s and only though neural nets have we gotten to human and superhuman level advances. If you want explainability it's going to require giving up a lot of innovation that could help people. Especially when we ourselves can't explain our own behavior psychologically.

> In many cases making a neural network explainable is literally impossible.

Then it should be illegal to use on consumer websites. It's that simple.

> If you want explainability it's going to require giving up a lot of innovation that could help people.

Four people died in the US Capitol because of this "innovation". It's time to stop pretending technology doesn't cause more harm than good in these cases.

I don't like when people try to stop or slow down technological innovation. Technologies are just tools, a nuclear bomb could just as easily have been made into a nuclear power plant. But to fully restrict them shows a lack of foresight. Yes, recommender systems can be used for bad, as yesterday, but they can also be good.

It's likely unrestricted technological innovation will inevitably cause the end of the human species, and your own example demonstrates why: In our rush to create the nuclear bomb, several times we came close to obliterating all life on earth due to simple mistakes. Several times, the hesitation of one man was the only thing that prevented our planet's complete annihilation.

The nuclear bomb is exactly why we need to block "innovation" that causes more harm than good.

And yet nuclear energy is one of the highest latent energy sources we have, orders of magnitude above solar, wind, coal and so on. If we blocked progress, then we'd have never discovered this.

The question is, how do you know something is bad before developing it? You oftentimes can't put the genie back in the bottle.