To what extent do social media algorithms, and YouTube’s specifically, contribute to user’s views becoming more extreme and even radicalizing? That’s a complicated question, but researchers tried to tackle it and have come up with some interesting results.

First let’s define some terms (as the researchers did). An echochamber is a social media information ecosystem (such as a specific forum, blog, or content provider) that shares a common view, ideology, or narrative. The researcher use this precise definition: “a bounded, enclosed media space that has the potential to both magnify the messages delivered within it and insulate them from rebuttal.” A rabbit hole is a recommendation algorithm that leads a user from their initial content interest into a widening circle of related content. This is an iterative dynamic process that does not necessarily involve more viewpoint homogeneity or more extreme views. A “radicalization pathway” means that some combination of user behavior and platform algorithms leads a use to more and more extreme content and views. The trick in research is separating cause and effect – do extreme users seek extreme content, or does extreme content make users more extreme?

The researchers start, as is commonly the case, with a review of existing research. It’s probably not as bad as you are imagining. The research generally shows that users do tend to stay within mild ideological echochambers, and that this phenomenon is largely driven by user behavior, not platform algorithms. This makes sense – people will tend to sort themselves into groups of like-minded people. There is research showing this going back to 1940, long before social media. But this also refers to the average behavior, with a lot of variance. There are some sites on social media that are extreme echochambers, with careful curating of content and users. I have seen this defended as creating a “safe space” for a certain group. Those who do not follow the ideology of the group are tagged as “trolls” and quickly purged. But most of social media does not fall into this category.

Rabbit holes, on the other hand, are common on social media and are partly driven by platform algorithms that recommend content to the user. This is largely based on determining what the user likes, but it can be difficult to parse how much of this is driven by user behavior and how much by the platform algorithm itself. That is what the researchers set out to test. Here is how they did it.

They recruited 1,639 real YouTube users in late 2020. All users stayed logged into their own YouTube accounts, installed a browser extension, and followed a controlled navigation task. Half could choose videos normally (“preference” condition). Half had to follow rules like: always click the 2nd recommendation, always click the 4th recommendation, etc. In the first condition the users were making choices, but in the second condition they were not so that would tend to lead wherever the algorithm takes them.

The first finding was that echochambers are primarily driven by user behavior. In the study conservative users sought out conservative content, and liberal users sought out liberal content. But in the second condition where users followed rules rather than their own choices, content did not homogenous ideologically, and ideological differences in the offered videos largely disappears, regardless of the user’s starting point. This means that users are largely responsible for ideological echochambers.

The second finding was that rabbit holes are real and driven by the algorithm. YouTube recommends more videos related to the content being viewed, leading to an effect operationally defined as a rabbit hole – an expanding ring of content related to the original content.

Reassuringly, they did not find evidence for radicalization pathways. In either group, video recommendations did not become more and more extreme, but rather became ideologically narrower over time. This suggests the algorithm tends to reinforce initial beliefs, but not push them in a more extreme direction. This does not mean people cannot get radicalized online. But it does suggest the algorithm isn’t designed to make that happen.

The researchers also found that YouTube does tend, over time, to recommend moderately conservative content, whether or not the user is liberal or conservative. This could be an inherent bias in the algorithm, but not necessarily so. The researchers hypothesize this could be an artifact of there simply being more conservative content on the platform, or that conservative content has more “engagement signals” such as views, watch time, or likes.

This is just one study and we have to put in into context with all the other research, but it seems the consistent signal here is that user behavior on social media platforms is what mostly drives the creation of echochambers and the consuming of extreme content. These things are not inherent to the algorithms. What the algorithms do is continuously feed us related content to keep us engaged for as long a possible.

However, I do not think that this absolves the platforms of the effects of social media use. Algorithms are designed to be addictive, and they feed us content which is engaging. They are optimized for maximizing use of the platform. This may be good from a business perspective, but is likely not good from a societal perspective. Human psychology is what it is, and it is not meaningful to absolve a product or service and blame human psychology for negative outcomes. For example, fast food restaurants could say it is not their fault that people eat too many calories – that’s on them, even while they are exploiting human behavior by providing products that are ridiculously calorie dense, optimized to stimulate taste and desire, and not optimized to be healthy. Or, gambling sites can say they are blameless if their users gamble irresponsibly. To give an extreme example – drug dealers can consider themselves harmless for just filling a need that people have.

There are many industries that feed humanity’s many needs, desires, addictions, and weaknesses. Generally we recognize that there needs to be a balance between the responsibilities of the individual and the responsibilities of institutions and society. I don’t buy the libertarian argument that we should 100% rely on individual responsibility and let the market sort it out. But I also understand the limitations of the “nanny state” and the unintended consequences of infantilizing humanity by obsessively protecting them from themselves. There is a happy medium.

I do think that personal responsibility is a reasonable default, just not an absolute principle. The state has the burden of demonstrating that regulations designed to protect people from their own frailty and also from exploitation are necessary, reasonable, proportional, and efficient. This can be done first through simple transparency, but often stricter regulations are necessary (such as with highly addictive substances).

I also think it is reasonable and even necessary to consider the total cognitive burden on individuals living in our current society. You cannot just look at each case in isolation. Navigating this complex world, driven by super specialized highly technical knowledge and with the technology to precisely target our greatest vulnerabilities, can be exhausting. We can be expected to assume a reasonable burden of personal responsibility, but if this total responsibility is overwhelming to the average person, with high stakes hinging on rarified knowledge or extreme willpower, that hardly seems fair or like the kind of world we all want to live in.

What does this mean for social media? I would like to see some incremental changes, with research to see what effect they have. For example, algorithms can occasionally recommend content that is not targeted to the user, just to mix it up and expose them to new perspectives. I would like to see algorithms favor quality content, not just engaging content. I would rather this happen through good corporate citizenship and public pressure, but I am not adverse to carefully crafted regulations. At the very least, users should have transparency and control over their own algorithm preferences. For example, imagine if in the user profile of a platform there are a number of sliders – do you want recommendations based on what we think you will like, what is most popular, what is the most academic, what is likely to be the most challenging, and how much random content do you want in the mix? Do you want more scientific content, entertainment news, politics, or human interest?

Algorithms and AI increasingly control the content we see, which dramatically affects our view of reality. I think everyone has the right (and the responsibility) to control that content, at least in terms of priorities and goals. And we have to track the effect this is having on us individually and on society. Research like this helps us as least understand the phenomenon, but we have to do something with that knowledge.