Everything Looking the Same Is Not New

This started with movie posters and a search for the cure of sameness.

A decade of movie posters seemed to share about four colours, and I wrote an essay about wanting something that jolted me out of the monotony. I was looking for things that "changed my brain chemistry": a fashion show with ballet dancers as models, breaking the fourth wall on television, book chapters written in PowerPoint, and familiar things presented in unfamiliar ways.

After reading a few articles about sameness, I drafted an article proposing that platforms should build a discovery mode: an unoptimized feed that shows you things you normally wouldn't like, a non-algorithmic feed. I was rather pleased with myself.

But after some reading, I found that the complaint about sameness is much older than our algorithmic feeds. Horkheimer and Adorno were writing about culture in 1944, describing film and radio as standardized like manufactured goods. American radio consolidation later squeezed playlists to a few hundred songs, for instance. Long before social media and other recommendation-based platforms, the internet also did something similar-ish. Blog culture converged on a house style after a few years running on links and RSS. Every new distribution system eventually develops a grammar (read 'sameness') of its own.

So the problem predates the current infamous algorithms. My new question was, if people have complained about this for a century, what makes "now" different?

Speed, Reach, and AI

The loop between what "performs" and what is created used to take seasons from end to end. The duration has shortened significantly; now it can run in an afternoon or a few hours, everywhere at once. Different creative industries used to develop their own conventions and produce their own form of sameness. They could become increasingly formulaic, but in different ways. Now the same ranking systems can push creative output around the world towards the same patterns.

Generative systems move culture around and produce it as well. In a Nature feature this month, Emily Wenger says, "We have suffered as a society when we silence or ignore edge voices."

One study asked people in India and the United States to describe their own rituals and symbols, half of them using an autocomplete tool. It made the Indian participants sound more American in their word choices and removed the details, so descriptions of Diwali became less culturally specific and more generic, with descriptions that sounded alike.

What's more scary to me is that people who used ChatGPT for five days were still giving more similar answers to each other two months later, even in the absence of AI. The researchers called it a creative scar. What if it's also a cognitive scar?

Lauren Leek's Temperature Zero for Culture article explains the mechanism better than I can:

"Personalization under a standard loss function is regression to the collective mean with extra steps."

Her work on British pub closures makes this concrete. The strongest predictor of a pub surviving was how closely it resembled the median pub. The more character it had, the more likely it was to shut down, because character is hard for a bank manager to score.

I'm not against recommendations. I like Netflix saving me forty minutes of scrolling. The problem is recommendation being the only mechanism through which anything gets found. Which is how three different people have sent me the same link to something they saw on social media, despite them being strangers to each other and one of them not even having an account on the platform. No, the video sent to me did not have a million likes, so I can't attribute it to virality.

Now what?

My first answer was randomness.

My "discovery mode" is nothing new. StumbleUpon was all 'discovery mode' for sixteen years and shut down in 2018. Spotify's Discover Weekly successfully offers more diversity, novelty, and serendipity than the platform's other feeds, so it can be done. But turning up novelty alone makes the experience worse. What people respond to is serendipity, not randomness for randomness' sake. Random feeds would mostly deliver noise, because most unfamiliar things are unfamiliar for good reason.

What people value is the unexpected thing that turns out to be interesting, useful, or differently relevant. This made me realize that I may have misunderstood the movie-poster complaint. We want concept expansion even if all we have is the familiar.

Discovery means more than finding the opposite of what you already like. We need to find the edges of what we like, something adjacent enough to be meaningful but strange enough to interrupt the patterns we're used to. I admit this might be a harder problem to solve.

If we're interested in solving it, I think the principle needs to go further than precedents like StumbleUpon. The heart of the matter is that recommendation should not be the only mechanism through which we find everything on the internet.

Every optimization system needs somewhere that optimization is not all there is. Leave room for exploration, a part of the internet experience where people's work is not judged by the same metrics as the rest of the system, because those metrics will systematically favour what is already familiar.

A diverse culture needs a long tail, more outliers, and less regression to the mean. It needs room for things that are difficult to classify, difficult to predict, and maybe even initially difficult to appreciate.

We haven't run out of interesting, different things. We've stopped rewarding them. The question is whether we can build systems that help us find them regardless of whether they're rewarded.

In the spirit of finding the edges beyond what recommendation systems think we'll like, here's a list of other reflections you might enjoy.

List: Inspired by Data | Curated by Varyanne Sika | Medium Inspired by Data · Stories, insights, and reflections on the art of data storytelling. Exploring the human side of…