As a devoted listener of the Decoder podcast, I’m no stranger to the topics of content moderation and social media algorithms, but ever since that one certain rich guy bought that one certain bird website the internet has been abuzz with hot takes on these subjects.

It seems clear that algorithmic content recommendation is the source of real world harms, but at least some people think it’s actually pretty OK.

It’s a nuanced topic, and in my opinion, the right answer will be equally nuanced. Maybe there is a reasonable middle ground?

Algorithmic content recommendations only exist because they are profitable for the social media companies, but why, again, are they bad?

In my observation, this setup has a few failure modes:

1. Perpetually fresh content feeds are purposefully addictive and are actively eroding everyone’s attention span¹.

2. Objectively harmful content (self-harm or pro-anorexia are examples we’ve seen on TikTok lately²) may be surfaced automatically to people who are most likely to be harmed by it.

3. Modern algorithms are black boxes and create emergent echo chambers that content consumers can scarcely understand let alone manage³.

We’re past the point of asking whether recommendation algorithms at scale are good or bad. They sure are great for the companies that created them, but objectively they are bad for us.

So why not simply abandon recommendations?

As with most things, there are two sides to the coin. On the one hand, content recommendations at scale cause direct harm. But on the other hand, there is something to be said for discovery.

The amount of content generated on any given social media platform now is beyond comprehension. Every minute, Twitter receives over a half million tweets, Instagram receives some 60,000 photos, and YouTube receives over 500 hours of video.

Increasingly, people are turning to these platforms to see more than content generated by others they know personally. TikTok is bucking the trend of relying on a “social graph” to drive recommendations and the other platforms are following suit.

For good reason: I personally follow authors, journalists, makers, and musicians alongside the people I know from my personal and professional lives. My feeds would be boring (to me) if they only contained photos of meals and hot takes from people I know.

How do we discover the people we want to follow without allowing some AI to find them for us?

Without some manner of recommendation, everyone would have to search for things they like, which has two fatal flaws:

1. People are bad at describing things (which might be solved by a sufficiently intuitive search system).

2. You can only search for things if you know that they exist.

It’s the second one that is damning. I definitely follow people who are doing things I wouldn’t have dreamed of or ever searched for. Thanks, YouTube.

So recommendations are harmful, but recommendations are necessary. What do?

I’m trying to resist the urge to give in to technosolutionism and propose some new, simpler, better algorithm that will decisively fix all of our problems. Although some different technological approach to recommendations may work, my proposal is more fundamental.

To my mind, the only way to create a system that truly acts to serve its users alone is for the users to own it. You can see the beginnings of what that looks like on Mastodon, the most popular platform within the so-called “fediverse.” Mastodon has no algorithm at all, and most discovery is driven by hashtags.

No algorithm means you’ll never see something and not know why; you only see posts from people you follow, and those “boosted” by people you follow, and the difference is explicitly shown.

Mastodon’s use of hashtags feels like the OG Instagram, where hashtags was the only way anyone found you. You can even follow a hashtag now, and those posts will appear in your home timeline.

By viewing the “federated timeline,” you can view all of the posts made publicly by anyone followed by anyone on your home server. Each Mastodon server also suggests trending hashtags.

Mastodon is only one platform, and may or may not prove itself to be as useful, effective, or fun as the others. I use it only as an example of what social media discovery features look like when designed by the users themselves.

I think the position I’ve come around to is that complex, AI-driven recommendation algorithms may not be necessary. A naive algorithm could surface things that are unexpected, but wouldn’t exhibit the intrinsic bias of its training data or the humans who created it.

“See what you ask for, or see everything.” Mastodon’s tag-based discovery solves the problem of finding what you want to see, but if you want to diversify your feed, you can view that federated timeline, which is simply chronological. It’s a firehose, and on some servers it can be overwhelming, but it’s all-or-nothing.

This way, you never see “only the concerning posts our computer brain thought you’d like.” You’ll see those, too, but along with everything else in chronological order for you to make the final call.

This is a hard problem. Maybe the hardest problem in communication at scale today. What do you think? Are recommendation algorithms good? Bad? Neither?