Hacker News Reader

Comment by ethbro | Hacker News Reader

Facebook completely controls the presentation priority, no? Bolt a Watson-like truth scoring system on to evaluate content and use the resulting confidence to boost or bury content in other's feeds.

It would absolutely have to be automatically generated, but it doesn't seem impossible (just much harder NLP that Watson playing Jeopardy).

View on HN

If the scoring system does anything other than maximize engagement with Facebook, Facebook ends up making less money and performing worse on KPIs than they otherwise would. It'd be very difficult for any decision maker at Facebook to build, deploy, and maintain that feature. It'd be the sort of thing that'd only be there if otherwise the company would be under existential threat.

> under existential threat

You mean government regulation a la China because some memes really are dangerous? If you increase the velocity of transmission, then even in a free system there are some dangerously infectious falsehoods that it might be in the public good to slow. Think War of the World on Facebook.

That, or a public reaction to their product (ie, clickbait fatigue - Facebook gets really good at local optimization for the most engaging news posts but creates a news feed that convinces people that Facebook as a whole isn't worthwhile).

It basically needs to be serious enough to overwhelm the internal incentives of maximizing key performance indicators.

I think there's the global maximum vs local maxima argument in favor of doing something too. Why do people ultimately spend time on Facebook?

I would say they don't do so because of micro-optimized KPI targeting: they do so because they feel time on Facebook is valuable and makes them happy. If Facebook can't deliver increases in that, then optimizations only take them so far.

Yup. There's definitely some Goodhart's Law going on here as well. Micro-optimized KPI targeting is an inexact measure of feeling like ones time on Facebook is valuable.

Ah, but herein lies the difference. Memes seem to universally promote freedom of communication, communication about freedom, and a way to react to the oppression of the status quo. In China, this is counter to the powers that be, but in America, this is fully aligned with the anti-regulation party that is sweeping through all branches of government. Memes play into their hands, and so Facebook would never be under existential threat.

Watson scoring something for truth is only as good as the data drawn from. Who decides that data? Jeopardy questions center around long-established, non-controversial information; how does a Watson-like system evaluate unprecedented breaking news? It's not a question of the language used, it's a much, much higher-order issue that current Watson sidesteps.

Agreed, much more difficult, but ultimately the same way we do so? We believe certain ideas to be facts (hopefully based on scientific sources), then we parse incoming information in light of those.

I can't imagine an unprecedented breaking news story that has no basis in any factual information. I'm not talking about an upvoter here, but primarily a downvoter (incongruence with accepted facts being easier to prove than vice versus).

Sure, and that's the higher-order problem. If we could make machines that solve any problem the way we do, we'd be a lot closer to AGI, but for now, it's still science fiction for a computer to have beliefs, and understand information in the context of those beliefs.

I don't think ranking new information according to held beliefs is necessarily that far towards AGI (admittedly, strong NLP may be, though).

There's no fundamentally creative step in deciding "How well does this new piece of information match pieces I previously had"?

It seems to me the hardest part would be separating truthiness from virality. How do they seed legit sources in the current media climate of racing to the bottom for eyeballs.

Isn't this essentially the problem Google's been solving with continued iteration on PageRank?

From an untrustworthy collection of input relationships, how do I produce the most reliably correct output?

No because popularity does not equal truth.

It doesn't? Because in the sense that popularity is "agreement with consensus scientific opinion" then it absolutely does.

Well, to be more specific about what I mean is, Google's algorithm is successful if it finds the pages that people find valuable. It's sort of baked into the value proposition that the end user will be able to gauge the quality of results, so there is a sort of feedback loop. The problem with misinformation is there is no in-band signal for Google about its truthiness. I suppose there heuristics you can use related to fact-checking quality of certain publications, etc, but it's much trickier than general perceived quality.

Nevertheless, I agree that if anyone can do it algorithmically, it would probably be Google.

When was this ever the definition of popularity.

I'd like it if it was, but I've never witnessed popularity that embodies that statement.

Popular opinion is usually informed by the most often repeated narrative.

No, it doesn't. Sometimes, popularity blocks the process of scientific inquiry. As Kuhn pointed out. But in any case, there is no truth, as Popper pointed out.

This is basically impossible given the current and near future state of the art.

The Allen AI group has been working on solving year 4 multiple choice science tests for about 3 years. They now score a bit over 60% correct, and that is a much easier task.

What I'm suggesting we start with is not exactly the same. The analogous question in that domain is "How accurate have they been in eliminating wrong answers?"

Everything has to be right for an answer option to be correct. Only one thing has to be wrong for it to be incorrect.

And how do you think that they would do that?

Choosing a random story from politifact: http://www.politifact.com/missouri/statements/2016/nov/06/ro...

Like most false stories, there is a hint of truth to this but that doesn't make it true. Politfact spends 3 pages discussing it, and concludes:

Courts did object in three cases to decisions Kander made as secretary of state. But the specifics of those cases are much more nuanced than Blunt lets on. And the claim that Kander tried to manipulate the election is, at best, unproven.

A SOTA computer system could probably parse that last paragraph I posted, but couldn't get anywhere near getting close to doing the research needed to get to that conclusion. We are years off that (I work on research in this area)

I'm not surprised parsing something nuanced is beyond state of the art. But what about "illegal immigrants are committing an epidemic of rape against our children" or "we have a flood of illegal immigrants"?

A lot of claims that were made this cycle are rooted in measured quantities where the measurements we have disagree with the claims.

Define "epidemic of rape" or "flood of illegal immigrants".

If there is at least one rape or at least one illegal immigrant then it becomes subjective or contextual.

Is 3 rapes a flood? What if it is in a week in one town? What if one was by a former illegal immigrant? What if it is over the course of a year, but all 3 occurred in one day?

Is 20 illegal immigrants a flood? What if they all arrive in a town with a population of 100?

Who is "we"? Is a story about millions of illegal immigrants in Greece relevant? What if it about thousands of illegal immigrants working in Athens? What if Athens is in Georgia?

A real example:

"Broken Families: Raids Hit Athens' Immigrant Community Hard" vs "Athens crackdown shows no hospitality for illegal migrants"

One is a real headline from Georgia, the other from Greece. Can you tell which is which?

They can just form a group of truth checker, a ministry of sort, who's job is to inform people of the truth. They can call it the Ministry of Truth, where busy citizens who have no time to research can get their truths from.

Reply on news.ycombinator.com