03 April 2019

Recommendation algorithms are the biggest problem in science communication today


Having been interested in science communication (and being a low-level practitioner) for a while, I recognized that there are a lot of old problems that recycle themselves. Like, “Should we have scientists or journalists communicating science to the public?” (Why not both?) “Is there value in debunking bad science, or does it make people dig in and turn off?” “Why can’t people name a living scientist?” “You have to tell a story...”

Having followed social controversies about science (particularly the efforts of creationists to discredit evolution) for a long time, I was almost getting bored seeing the same issues go ‘round and ‘round. But now I think we are truly facing something new.

Recommendation algorithms on social media.

You know these. These are the lines of computer code that tells you what books you might like on Amazon based on what you’ve bought in the past. It’s the way Facebook and Instagram deliver ads that you kind of like seeing in your feed. And it’s how Netflix and YouTube shows you want you might want to watch next.

I think recommendation algorithms may be the number one problem facing science communication today.

And of these, YouTube seems to be particularly bad. How bad is YouTube? Pretty bad, in two ways. First, it is the biggest social media platform. Only Facebook rivals it.


YouTube is the most widely used online platform among U.S. adults. 73% of US adults say they use YouTube online or on their cellphone.

Second, YouTube is very effective at convincing people of false stuff.

Interviews with 30 attendees revealed a pattern in the stories people told about how they came to be convinced that the Earth was not a large round rock spinning through space but a large flat disc doing much the same thing.

Of the 30, all but one said they had not considered the Earth to be flat two years ago but changed their minds after watching videos promoting conspiracy theories on YouTube.

You watch one thing, and YouTube recommends something even a little crazier and more extreme. Because YouTube wants you to spend more time on YouTube. People are calling this “rabbit hole effect.”

It’s not just happening over scientific facts, either. Many have noted that YouTube is having a similar effect in politics, with many branding it a “radicalization machine.”

Reporter Brandy Zadrozny summarizes YouTube’s defense thus:

YouTube’s CPO says the rabbit hole effect argument isn’t really fair because while sure, they do recommend more extreme content, users could choose some of the less extreme offerings in the recommendation bar.

So we need humans need to fix the problem, right? Human content moderation is the answer! Well, maybe not, because repeated exposure to misinformation makes you question your world view.

Conspiracy theories were often well received on the production floor, six moderators told me. After the Parkland shooting last year, moderators were initially horrified by the attacks. But as more conspiracy content was posted to Facebook and Instagram, some of Chloe’s colleagues began expressing doubts.

“People really started to believe these posts they were supposed to be moderating,” she says. “They were saying, ‘Oh gosh, they weren’t really there. Look at this CNN video of David Hogg — he’s too old to be in school.’ People started Googling things instead of doing their jobs and looking into conspiracy theories about them. We were like, ‘Guys, no, this is the crazy stuff we’re supposed to be moderating. What are you doing?’”

Mike Caulfield noted:

I will say this until I am blue in the face – repeated exposure to disinformation doesn’t just confirm your priors. It warps your world and gets you to adopt beliefs that initially seemed ridiculous to you.

Propaganda works. Massive disinformation campaigns work. Of course, people with a point of view and resources have known this for a long time. Dana Nucitelli noted:

(T)he American Petroleum Institute alone spent $663 million on PR and advertising over the past decade - almost 7 times more than all renewable energy trade groups combined.

Meanwhile, scientists are still hoping that just presenting facts will win the say. I mean, the comments coming from the first day of a National Academy of Sciences colloquia feel much like the same old stuff, even though the title seems to hint at the scope of the problem (“Misinformation About Science in the Public Sphere”). It feels like the old arguments about the best way that individual scientists can personally present facts, ignoring the massive machinery that most people are connected to very deeply: the social, video, and commercial websites that are using recommendation algorithms to maximize time people spent on site.

The good news is that as I was writing this, the second day seems to be much more on target. And one speaker is saying that the biggest source of news is still television. Sure, but what about all the other information people get that is not news?

The algorithm problem requires a deep reorienting of thinking. I don’t quite know what that is yet. It is true that this is a technological problem, and technological fixes may be relevant. But I think Danah Boyd is right that we can’t just change the algorithms, although I think changing algorithms is necessary. We have to change people, too, beyond a “If they only knew the facts” kind of way. Because many do know the facts, and it don’t settle matters. But changing culture is hard.

I think combating the algorithm problem might require strong political action to regulate YouTube in the way television networks in the US were (and, in other countries, still are) regulated. But the big social technology companies are spending millions in lobbying efforts in the US.

Update, 17 May 2019: A hopeful tweet on this matter.

Promising data point for YouTube’s anti-conspiracy push (on 1 topic, at least): the people in my flat-earth FB group are very mad that the site has stopped shoveling lunatic videos into their feeds.

Update, 27 June 2019:

A good illustration of this phenomenon recently appeared in a piece for MEL magazine about an increasingly disturbing trend — women whose once-promising romantic relationships implode after their boyfriends become “redpilled.” For the benefit of the blissfully uninitiated: to be “redpilled” means to internalize a set of misogynistic far-right beliefs popular with certain corners of the internet; the product of a noxious blend of junk science, conspiracy theory, and a pathological fear of social progress.

The men interviewed in the piece, once sweet and caring, started changing after going down a rabbit hole of extremist political content on YouTube and involving themselves in radical right-wing online communities. Convinced of their absolute correctness, these men became at first frustrated, then verbally abusive once they realized their female partners did not always agree with their new views.

From here.

Update, 25 August 2019: Okay, this is political information rather than scientific, but a new preprint describes how YouTube facilitates getting users into increasingly extreme political content. The dataset here is impressive: “360 channels, more than 79 million comments, more than 2 million recommendations.”

One of the authors concludes, “YouTube has to be held accountable and take action.”

Update, 31 August 2019: Pinterest gets it.

When Pinterest realized in 2018 that the search results for many health-related terms – such as “vaccines” or “cancer cure” – were polluted with non-scientific misinformation, the visual social media site took a radical step: it broke the search function for those terms. “If you’re looking for medical advice, please contact a healthcare provider,” a message on the otherwise blank page read.

On Wednesday, Pinterest announced a new step in its efforts to combat health misinformation on its platform: users will be able to search for 200 terms related to vaccines, but the results displayed will come from major public health organizations, including the World Health Organization (WHO), Centers for Disease Control, American Academy of Pediatrics (AAP) and Vaccine Safety Net.

Update, 29 December 2019: Fascinating thread by Arvind Narayanan about this issue by on Twitter and why it.s so hard to study. Excerpt, with my emphasis added:

I spent about a year studying YouTube radicalization with several students. We dismissed simplistic research designs... by about week 2, and realized that the phenomenon results from users/the algorithm/video creators adapting to each other. ...

After tussling with these complexities, my students and I ended up with nothing publishable because we realized that there’s no good way for external researchers to quantitatively study radicalization. (Emphasis added. -ZF) I think YouTube can study it internally, but only in a very limited way. If you’re wondering how such a widely discussed problem has attracted so little scientific study before this paper, that’s exactly why. Many have tried, but chose to say nothing rather than publish meaningless results, leaving the field open for authors with lower standards. (Emphasis added. -ZF)

In our data-driven world, the claim that we don’t have a good way to study something quantitatively may sound shocking. The reality even worse — in many cases we don’t even have the vocabulary to ask meaningful quantitative questions about complex socio-technical systems. ... And I want to thank the journalists who’ve been doing the next best thing — telling the stories of people led down a rabbit hole by YouTube’s algorithm.


External links

The trauma floor
Study blames YouTube for rise in number of Flat Earthers
Google and Facebook can’t just make fake news disappear

YouTube’s algorithms can drag you down a rabbit hole of conspiracies, researcher finds
Google, Facebook set 2018 lobbying records as tech scrutiny intensifies
Americans are smart about science
The magical thinking of guys who love logic

1 comment:

Christina Pikas said...

I think the problem is specifically YouTube and FaceBook recommendations more than recommender systems in general. Like if you're reading a science story on The Atlantic and it recommends another of their stories - that's a good rabbit hole.