[throwawaygh]

> No primary moderation action should be made based on human input

1. That means no HN.

2. I normally don't have to remind people of this at places like HN, but... algorithms are written by... humans! Supervised algos use data labeled by... humans!

> Automated moderation should look for identifiable harms (i.e. illicit content, directed threats, terrorism)

Why do you list terrorism separately from directed threats?

What is the line/difference between "terrorism" and an "undirected threat"?

Are militia groups that don't make directed threats terrorists? Are radical religious groups that don't make directed threats terrorists? What if they are run by actual terrorists but none of the speech amounts to a directed threat?

Speaking of which, what is a terrorist organization? Is the KKK? What about small white nationalist or black power militia groups? What about QAnon? What about antifa? What about BLM? What about Westboro Baptist? What about the Black Panthers?

There are people -- elected officials -- who think each of those is a terror organization.

So, defining terrorist organization is absolutely a political fight. Maybe we avoid that and just talk about directed threats/ Ok. Does that mean that Al Qaeda allowed to operate on FB as long as they don't make directed threats? In fact, that FB is prohibited from not allowing Al Qaeda on as long as they don't make directed threats? That seems like not a solution anyone is going to get behind.

We haven't even gotten past the "obviously terrorism=bad" and we already have to declare whether BLM, QAnon, Westboro, or militia groups are "terrorists". Which some senators believe is the case and is a 100% political question.

> illicit content

Is Ginsberg's Howl illicit? Is a picture of two women kissing illicit? What about non-sexualized nude breasts? What about nude male bodies? What about an erect penis but in a non-erotic context? Will the dominant answers to these questions be the same in 50 years?

Lots of people would say a site that allows pictures of heterosexual kissing but not not pictures of homosexual kissing is obviously taking a political position, but that was outside the realm of "political opinion" when I entered adulthood! Any public homosexual display of affection was obviously illicit.

> absolutely nothing should be removed or blocked based on vague and nebulously defined concerns over "misinformation".

What does vague mean? What does nebulously defined mean? What is the difference between misinformation and libel? What is the difference between misinformation and dangerous information? Is it impressible to remove a video that's targeted at kids and encourages huffing glue as a fun and harm-free activity?

Anyone who has moderated a forum knows that such an algorithm is going to have all sorts of holes and perceived biases. I've never written an automod that some user doesn't get pissed off about.

More generally: that's just straight-up moderation, it has nothing to do with tweaks to recommendation algos.

What if Twitter realizes that people leave the site if they see stuff about abortion but stay if they see stuff about LGBT rights? Again, viewpoint-neutral, Americans just one day start yawning about abortion and really polarize on LGBT stuff. Can they prioritize posts about LGBT rights over posts about abortion as long as the content served up on the preferred topic is viewpoint-neutral and the only algorithmic goal is more lingering eyeballs?

If no to that, how about sports news vs. SCOTUS decision news?

If yes to that, what about COVID case counts vs. Jobs Report numbers?

Even more generally: anyone who's stayed up to date on robust machine learning knows that defining good notions of robustness -- and political neutrality is a type of robustness -- is very much an open problem. So even if we had a precise definition of political neutrality, which I don't think we do, "simply create an algorithm that has that property" is very much an open algorithmic problem.

In fact, there are even some impossibility theorems in this space. So even if we can define neutrality in a perfectly neutral way -- which we can't -- this might be like passing a constitutional amendment that demands a voting system has all of: Non-dictatorship, unrestricted domain, monotonicity, IIA, and non-imposition. You can legislatively demand "the perfect voting system", but the universe is not obliged to ensure the existence of such a thing. Same for some types of robust ML, and no one knows which side of an impossibility theorem some precise-enough-to-code notion of political neutrality might fall on.

Which also brings up the REAL question: are tweaks to recommendation algorithms allowed? Obviously we can't ask FB/Twitter to freeze their recommendation algos -- it's their core product. So. If they notice an "obvious bias" and tweak the algorithm to correct for it, who decides whether that was a biased human intervention or a totally appropriate bug fix? Oh, right, a politically appointed FTC.

I think that "politically neutral" is impossible to formalize in code because it is a fundamental contradiction in terms. But even if it does, I suspect that any reasonable lists of formal specifications might be either mathematically impossible to train a classifier to respect or else at least AGI-complete to actually implement. But if you disagree, I'm happy to clone the Github repo and mess around with your proposal.

[adamiscool8]

>1. That means no HN.

No, it means less 230 protection for HN. Stop conflating this with destruction of the platform, it's becoming like "net neutrality". Remember when tweaking that killed the internet?

>What is the line/difference between "terrorism" and an "undirected threat"? Speaking of which, what is a terrorist organization?

The government has a clear processes to designate foreign and domestic terrorist organizations. [0] Let the actual politicians engage in that political fight. Social media companies can use the result.

>What is the difference between misinformation and libel?

Actual malice? If the standard works for newspapers, why can't it work for social media companies?

>More generally: that's just straight-up moderation, it has nothing to do with tweaks to recommendation algos. [...] If they notice an "obvious bias" and tweak the algorithm to correct for it, who decides whether that was a biased human intervention or a totally appropriate bug fix?

None of this relates. Content should not be removed or suppressed based on any political preference or designation, and that includes a fig leaf of facial neutrality. Whether it's recommended to some and not others != suppression, and it's trivial to show that your systems are based on user action not partisan interest.

These aren't sticky questions at all, they're just ways to navel gaze and avoid the obvious solutions that are inconvenient to certain actors.

[0] https://www.state.gov/terrorist-designations-and-state-spons...

[majormajor]

> No, it means less 230 protection for HN. Stop conflating this with destruction of the platform, it's becoming like "net neutrality". Remember when tweaking that killed the internet?

Really? If HN starts only moderating based on "identifiable harms (i.e. illicit content, directed threats, terrorism)" then it'll quickly become a cesspool and lose the community.

On the other hand, if they continue to apply posting guidelines, how many banned users suing HN over "politically motivated censorship" and shit like that do you think it takes for them to decide it's not worth it? Content removed because someone was an abusive jerk suddenly becomes, in plaintiff's claims, content removed because the moderators didn't like their politics. Now spend your $$$$ to defend against that claim!

You're sticking your head into the sand over what the unintended consequences of your proposals would be because you really really really want to believe it would only have the intended consequences that you like.

(Look at what you do when you bring up newspapers: newspapers have extremely limited user-generated content, because of the standards you're proposing extending. Again: there goes HN.)

The only stuff that would survive would be the stuff with big userbases, big pockets, and the ability to throw a lot of moderating power at stuff. Which all sounds to me more like traditional broadcast media - which is historically claimed to be also unfair to the same conservatives who are making the most noise about this stuff. So... good luck with that.

[throwawaygh]

>>> No primary moderation action should be made based on human input

>> 1. That means no HN.

> No, it means less 230 protection for HN.

I'd be fascinated to hear what dang thinks about HN's future existence if this hypothetical law where "No primary moderation action should be made based on human input" applied to HN.

It seems impossible to (a) run a healthy forum or (b) avoid lawsuits or even jail. E.g., can you link me to a github repo that automatically catches 100% of libel? Or even 100% of child porn (or I guess actual porn as a proxy for that problem)? Removing libel and other illegal content without "primary moderation action"s that are based on "human input" is not currently possible.

(BTW: that's NOT what Hawley's bill does! It allows human moderation, you just have to keep the political appointees happy.)

>> What is the difference between misinformation and libel?

> Actual malice? If the standard works for newspapers, why can't it work for social media companies?

Because newpapers have a few journalists. Not hundreds of millions of users.

This has to be done arithmetically or it's financially reckless to allow free-form comments at all. If it's so easy to algorithmically identify libel with 100.00% accuracy, go do it!

Given that there are regularly court cases that hinge on whether some statement raised to the level of libel -- cases that even get appealed and where highly trained judges disagree -- I'm willing to bet the problem is AGI-complete. And then some.

> The government has a clear processes to designate foreign and domestic terrorist organizations. [0] Let the actual politicians engage in that political fight. Social media companies can use the result.

> Content should not be removed or suppressed based on any political preference or designation

So politicians get to define what terrorism means and companies should suck it up and implement whatever the politicians in power decide.

So, if some powerful GOP senator designates BLM a terrorist organization, and social media companies then remove all BLM content, is that not "removing or suppressing based on political preference"? What about pro-2A militias? What about QAnon?

By the way, what about "illicit content"? If some hard core right-winger takes over Twitter tomorrow, can they ban pictures of homosexuals kissing as "illicit content"?

Hawley -- whose bill doesn't even do what you suggest -- is just shifting power over content moderation decisions from companies to political appointees. That's all. It's not neutral, it is based on human input, and it's primarily just a shift in decision making power.

Dressing this up as "neutral" is obvious bullshit. Hawley wants Twitter to understand that his political party is their ultimate master when they choose which speech to amplify on their platform. This is his explicit and openly stated goal. It is about power, not neutrality.

But anyways, this argument is easy to resolve in your favor. You propose not Hawley's bill, but a hypothetical different one where human input can't be a primary consideration. So, you're claiming that a formal specification of the political neturality of an NLP classifier exists. I've build a lot of classifiers, and I don't believe you. Show me the code.