Eve, here. Rajiv Sethi describes a fascinating large-scale study of social media behavior. It considered “harmful content” which is probably very close to what is currently or often referred to as hate speech. They found that when platforms succeeded in reducing the amount of that content, the users who amplified it the most not only reduced their overall participation, but also increased their levels of promoting hateful content.
I still have doubts about the methodology of this study. We used Google to determine what constitutes abusive content, in this case hostile language directed at Muslims in India. Google’s algorithms completely failed at identifying offensive text on Naked Capitalism, even going so far as to drop all complaints even when challenged. Probably this is better, but without some evidence there is reason to doubt it.
What I find a little disappointing is that Sethi touts BlueSky as a less harmful social media platform, saying its rules for restricting content viewing and sharing are fairly consistent with research findings. Sethi argues that BlueSky represents a better compromise between the concepts of free speech and curbing hate speech than the current major platforms.
I don’t buy the idea that Blue Sky doesn’t have as much hatred based on how bad Jesse Singal was treated. Singhal has drawn the ire of Blue Sky trans activists for simply playing an impartial position. That included falsely accusing him of releasing a transgender child’s private medical records. Quillette stood up to defend her campaign of lies against journalist Jesse Singal, and why it matters. Here’s what happened to Singal at BlueSky.
This second round was prompted by the fact that I joined Bluesky. Bluesky is a Twitter alternative site based on hardcore far-left power users who get furious when people they don’t like show up. I quickly became one of the most blocked accounts on the site. And, fearing cross-contamination, these users also developed a tool that allows them to block people who follow me all at once. That way they won’t have to face the threat of seeing my content or anyone who follows me. I finally had a truly safe space.
But that’s not enough. They’ve also been actively lobbying Aaron Roderix, the site’s head of trust and safety, to have me fired (here’s one example: “You son of a bitch. You son of a bitch. You bastard, you bastard. I want you to die, block me and disappear from your existence as quickly as I typed this post. I love your buddy’s song so much. Many of these complaints come from people who appear to be so highly deregulated that it’s difficult to use Waffle House well, but who are so active online that they have no real-world impact. may give.
So they’re not content with just blocking me, blocking people who follow me, and shouting at people who don’t block me, they’re rehashing every negative rumor about me that’s been posted online since about 2017. It started circulating. Catalog, sure. They even came out with a new one called “I’m a pedophile.” (Yes, they really say that!)
Suffice it to say, this is just the first section of a long catalog of vitriolic abuse at BlueSky.
IM Doc didn’t give many details, but a group of doctors who are somewhat of a maverick on the coronavirus issue went to BlueSky and quickly returned to Twitter. They were apparently met with intense hostility. I hope he elaborates further in the comments.
Another reason I’m reading opinion regulation is that even though it’s primarily aimed at suppressing speech, Zionists have forced many governments to take up criticism of Israel’s genocide and defense of BDS. This is because it is a method that has been successful in trying to prevent this. antisemitism. Distorted notions of hate are being weaponized to censor criticism of US policy.
So perhaps the reason why BlueSky seems more approachable, Sethi says, is because some users use highly aggressive and often hateful tools to crush the expression of opinions and information that go against their ideology. One would think that this is because they are engaged in enforcing a set of norms. I don’t consider it an improvement over other countries’ standards.
Written by Rajiv Sethi, Professor of Economics, Barnard College, Columbia University, External Professor, Santa Fe Institute. Originally published on his site
The steady drumbeat of social media posts about new research in economics accelerates toward the end of the year, as interviews for faculty positions are scheduled and candidates seek to draw attention to their work. A few years ago, I came across an interesting paper that had a huge impact on the way I think about meritocracy. The paper is currently under review in a major journal, and the lead author is a faculty member at Tufts University.
This year, I have been carefully working on research on polarization, which is the topic of a seminar I plan to teach next semester. One particularly interesting new paper comes from Aarushi Kalra, a doctoral candidate at Brown University who conducted a large-scale online experiment in collaboration with social media platforms in India. The (unnamed) platform is similar to TikTok, which the country banned in 2020. There are currently several apps competing in this space, from multinational offshoots like Instagram Reels to homegrown alternatives like Moj.
The experiment’s platform has around 200 million monthly users, of which Kalra managed to serve around 1 million and track an additional 4 million as a control group. 1 This process involves replacing algorithmic curation with randomized feeds, the purpose of which is to determine the impact on users. Exposure and engagement with harmful content. In particular, the authors were interested in viewing and sharing content that was classified as abusive based on Google’s Perspective API and specifically targeted at India’s Muslim minority.
The results are great. Those in the treatment group who had the most prior exposure to harmful substances (based on the algorithm’s response to prior involvement) responded to decreased exposure as follows: Overall engagement decreased and people spent less time on the platform (subsequent research showed they spent more time on competing sites). But they also increased their share of harmful content, conditional on encountering it. In other words, the reduction in sharing of harmful content was less than the reduction in exposure to harmful content. They also increased their active searches for such material on the platform, resulting in slightly more exposure than the treatment users who had the least exposure at baseline.
Some might argue here that switching to a randomized feed is a very blunt measure, and not something that platforms would implement or that regulators would prefer. Even those who were most exposed to harmful content under algorithmic curation were primarily consuming non-toxic fodder. For example, the proportion of content classified as harmful was approximately 5% of the feeds in the most exposed quintile at baseline, with the remaining posts catering to other types of interest. It is therefore not surprising that this intervention led to a sharp decline in engagement.
This is very clear when looking at the quintile of treated users who had the least exposure to harmful content at baseline. For this set of users, switching to a randomized feed resulted in a statistically significant increase in exposure to harmful posts.
Source: Figure 1 from Kalra (2024)
These users refused to engage with harmful content at baseline, and the algorithm avoided serving such content accordingly. However, this was not possible with randomized feeds. As a result, even these users experienced a significant drop in engagement.
Source: Figure 3 (right panel) from Kalra (2024)
In principle, it is possible to imagine interventions that degrade the user experience to some extent. The authors use model-based counterfactual simulations to investigate the impact of randomizing only a portion of the feeds of selected users (those most exposed to harmful content at baseline). This is interesting, as existing moderation policies typically target content rather than users, limiting exposure to only content classified as harmful while preserving algorithmic curation more generally. Or it may be worth investigating the impact of reducing it. I think the model and data make it possible.
But there is an elephant in the room: the specter of censorship. From a legal, political and ethical perspective, this is more relevant to policy-making than platform profitability. Although not necessarily codified in law, the idea that people have a right to access content that others might consider to be antisocial or abusive exists in many cultures. deeply embedded. In such an environment, it is natural for platforms to suppress political speech with suspicion.
At the same time, there is no doubt that conspiracy theories spread online can have devastating real-life effects. One way to get around this dilemma is to take advantage of composable content moderation. This gives users great flexibility in labeling content and deciding which labels to activate.
As explained in a previous post, this appears to be the approach taken at Bluesky. This platform allows you to hide abusive replies from all users, thereby slowing down the strategy of spreading abusive content by replying to prominent posts. The platform also allows users to detach their posts when they are quoted, forcing those who want to mock or ridicule them to use (less effective) screenshots instead.
Bluesky is currently experiencing some serious growing pains. 2 However, I am optimistic about this platform in the long run. This is because users can fine-tune content moderation, which should allow for a variety of experiences and isolation from attacks without requiring much central management. Censorship or banishment.
It was interesting to see entire communities (such as academic economists) move to another platform while keeping their content and connectivity largely intact. Such mass migrations are relatively rare, as network effects entrench platform usage. However, once it occurs, it is difficult to reverse it. This gives Bluesky some leeway as it tries to find a way to address complaints in a consistent manner. I think the platform will thrive if it avoids bans and blocks and prioritizes labeling and decentralized moderation. This should allow those who prioritize safety to protect themselves from harm without silencing the most controversial voices among us. Such voices can sometimes turn out to be the most prophetic in retrospect.
