TLDR: we should insist on auditable algorithmic feeds. Your informational environment shapes your beliefs and it does the same to the people you recruit into the system. We can instead build systems where we make the choices over what we see and the way it shapes us. As per my usual rules, I'm not going to address every nitpick in the text and will address them in the caveats and counterarguments section.
I wanted to write about why I no longer use Twitter*,and why you should avoid similar platforms that provide algorithmic feeds but refuse to reveal their algorithm. Note, I'm not a particularly ideological person; I lean more towards pragmatic decision making. Moving off platforms where you don't understand the algorithm, I've come to believe, is the pragmatic decision. If you want an algorithmic feed, you want it to be aligned with your long-term values and provide you the information that brings you closer to the person you want to be. You also want it to support the epistemic commons rather than possibly degrade them.
For almost every existing algorithmic system, you just don't know if it's achieving those objectives and their opaqueness make them challenging to study. I also suspect that I don't need to convince you that most existing systems are optimizing objectives like click rate and time spent on platform that are misaligned with anyone's long term values. We all have experienced the experience of zoning out in front of an algorithmic feed and coming back to discover time wasted and a deep sense of guilt. Systems that allow you to inspect the algorithm or build your own algorithm offer a way out. Social media algorithms are here to stay but competition and freedom in the space of algorithms offers a path to a brighter future and it's worthwhile to support platforms that are open about their optimization objectives and offer the prospect of user choice.
Now this is an important problem because...
It is fairly obvious that we have exited the era of gatekept systems. The number of viewers or subscribers for any non-algorithmic system are dwarfed by the amount of time spent and users for the algorithmic systems. Companies like CNN, Fox, MSNBC have millions of viewers for any show, a number that is immediately topped by random niche youtube videos. People spend somewhere upwards of 2 hours a day on social media; I find it hard to imagine they spend anywhere near that amount of time reading the New York times. arXiv-like models are starting to show up in every academic field with journals mostly a rubberstamp for bean-counters (note, I still think journals are important and have a valuable role to play but lets talk about that in a separate blog). It just seems incontrovertible that the place people are spending most of their time is also where they are forming most of their beliefs
This problem is also not going away because...
People are voting with their feet. Algorithmic platforms like twitter, tiktok, and youtube are more popular than ever. Furthermore, because of how unstructured and chaotic the internet is, there's basically no way out of some algorithmic process for helping you navigate it. Now note that saying people "want" these things is a bit of a weird word because it also might be fair to say that folks are addicted to these platforms. and don't necessary "want" to use them but lets respect the fact that people continue to actively seek these systems out and wil likely continue to do so.
So, if both things are true, that people want algorithms and that they are already getting most of their information from algorithms...
By an auditable algorithm, I mean one where the both the quantity it is optimizing and the resultant final ranking procedure / weights are available. Auditable, non-opaque algorithms are a wonderful opportunity. In our current world we have 1-4 actors, each pushing our collective information environment to believe in whatever they feel like. In a world of auditable algorithms, we can instead start to argue about what equilibria those algorithms should be pushing us towards. When those algorithms are auditable, we can check if we believe that they are making our epistemics better or worse or choose amongst them in ways that align with our own beliefs about who we are and what information we want to acquire.
We have also barely explored the limits of the positive benefits algorithmic feeds could have. We barely know how to study them because companies that haev algorithmic feeds restrict data accessibility and algorithm internals. Bridging-based algorithms, like BirdWatch (now called Community Notes and falsely attributed to Elon for rolling ) have been incredibly successful and suggest that we're only beginning to see the ways that these systems can held with surfacing truth. Imagine some of the following:
As an example of folks who are trying this, I loved the prosocial ranking challenge (https://rankingchallenge.substack.com/) which tried an ambitious version of this and mostly seems to have been bottlenecked by their ability to acquire enough user engagement and measure the imapct of the algorithms on users. Open API access, which basically no major social media platform now has besides bluesky, seems like it would have allowed for more rapid iteration.
In an era of reduced institutional trust, I think it's also long-term pragmatic for algorithmic feeds to make their algorithms visible. Recently, I came to believe that Twitter, before the election, decided to turn up the weights on the amount of misleading conservative information it pushed into its algorithmic feed and then switched it back after the election (note, this isn't a criticism of conservative content but specifically misleading conservative content. As a faculty, I'd prefer to avoid making explicit political statements in this way though you're welcome to read whatever you'd like into it). This strikes me, and likely many others, as true. Of course, it could be not true! It is not impossible that I'm imagining this, that it didn't happen, but in the absence of algorithmic visibility there is nothing Twitter can do to convince me it didn't happen. I certainly feel like I experienced it and quick experiments I made with new accounts also backed up for me that they might still be doing it. Sure, at this point they could release their ranking procedure but I'm not exactly going to believe them that whatever they release now is representative of their procedure prior to Nov. 4th. Consequently, I refuse to use twitter as I do not want to help this system sustain itself through network effects.
This problem, of a user deciding the algorithm is working to push the folks around them in a direction they disagree with is only going to show up more as ranking-based systems become more enshrined. This is going to happen over and over and over again and in every direction. By providing continual visibility to their ranking algorithm, they can build institutional trust and allow themselves to actually step into their new role as informational gatekeepers.
I'm not going to come out here and say we have the answer. I'm pretty excited about bluesky, since the firehose, the APIs, and the toolchains emerging around it mean that it's straightforward for users to build the type of marketplace I'm advocating for. It's also possible for researchers to study the results of particular algorithms and their effects on users by examining the semantic content of user behavior over time. In particular, I've played a little bit with Graze which lets you build your own algorithm atop the firehose. Bluesky certainly has problems with an alternative type of hateful conduct though I'd argue it's less than other sites and that it's possible to tune the algorithm to prevent you from being recommended it. The final answer might be a different website built atop the firehose. In the meantime, I'm going to keep supporting bluesky as it's the closest to being able to realize the vision outlined here.
In here I list a set of counterarguments I believe I am likely to see. Note, I'm having trouble steelmanning a lot of them. Read at your own peril / boredom.
Boring and a distraction from my point. My point is that many people want algorithmic feeds and so pointing out that you can just not use an algorithmic feed is missing the argument entirely. Furthermore, it fails to acknowledge the reality of the internet which is that in its current state it is an unstructured firehose in which there is a lot of valuable information but it's not obvious how to get to it. There doesn't seem to be a real way out of applying some kind of filtering and discovery process.
Of course, it could be true that algorithmic feeds are bad and cannot be rescued! I'm certainly open to the possibility that they interact with the human psyche in such a way that the majority of people will always opt in to algorithms that do ourselves long-term damage. It's something I do mull over though I don't really see a future where the government simply comes down and bans any sort of algorithmic content recommendation. Nor do I see a future where the culture collectively agrees that these things are bad for us and we learn to avoid them. Maybe hand-curated content resumes its popularity and we each rely on a well-aligned human to make suggestions for us, kind of like receiving book recommendations in your local bookstore (though note, this is also a kind of algorithmic recommendation). Doesn't seem super plausible though I do think it's also a kind of beautiful vision of a less lonely version of the internet. So, I sort of think they're here to stay and we should figure out how to make the best of them.
Boring and a distraction from my point. A point I've made is that many people want algorithmic feeds and that algorithmic feeds can be pretty valuable if built correctly and so pointing out that you can just not use an algorithmic feed is not engaging with the argument.
Plausible! Certainly if you think of a company like Google, the carefully tuned weights of its recommendation system / search engine are a lot of its marketplace advantage in a world where anyone can now index the internet. However, my claim is that sites can capture a user base by abandoning this advantage and offering an excellent algorithmic marketplace / toolchain and then sealing its dominance in through network effects. Sure, another company could take the most popular algorithms and repackage them but users are lazy and want to stay where their ongoing conversations already are.
Sometimes when you criticize twitter or something people do, someone will say "but freedom of speech" or some variant of this. I don't even know how to steelman this one which is why this section is dismissive and rude. Since I can't steelman it or ideological turing test it, I can't really argue against it. I just don't even really know what I'm arguing with.
I'm unclear how indicating that I believe social media is increasingly important and so we should construct versions of it that are positive, likely to increase the amount of truth in the commons, and under user control constitutes virtue signaling. This again is a position I struggle to steelman but that I also occasionally hear. Since I can't steelman it or ideological turing test it, I can't really argue against it. I just don't even really know what I'm arguing with.
Correct, I do have a twitter use policy and continue to use it under some restrictions. I post two things there (1) exhortations that people should leave twitter and (2) papers that my students write. I do not feel comfortable pausing the latter as promoting the work of my students is among my #1 most important jobs and so I am willing to accept slight compromises in service of it. Otherwise, I do not use twitter nor do I engage with the comments. I believe this minimal use is unlikely to help it stick around and is just running around in the remains of a decaying empire.
I already agreed with you that bluesky has problems. However, it has the possibility of resolving its problems in a way that competitors currently are unwilling to explore (probably because they like having control over their users).