At the Bottom of the Box was Hope

Leek proposes that these are all expressions of the same underlying phenomenon, a failure mode of data science that takes a world of "granular personal data" and arrives at a world where "personalisation produc[es] more sameness."

To these excellent examples, I'd add another one, from the world of monetary policy: Goodhart's Law, which holds that "When a measure becomes a target, it ceases to be a good measure":

https://en.wikipedia.org/wiki/Goodhart%27s_law

Goodhart's Law captures a wide variety of phenomena. When Google first deployed Pagerank, they showed that by counting the inbound links to all the pages on the web, you could extract a signal about which pages were most important (because there was no reason to link to a page unless you found it noteworthy).

But once Pagerank became the dominant means by which web users found pages, counting links stopped being useful: first, because people used Pagerank to find the best pages and link to them, making it impossible for new pages to get the inbound links needed to supersede incumbent pages; and second, because it's easy for fraudsters to create inbound links for low-quality pages in bulk, once there's a reason to do so.

Counting inbound links was a world-beating retrospective way of predicting which page would best match a searcher's query, but once it shaped the world it sought to analyze, it ceased to be a good prospective way to predict which page would best match your queries.

Leek is a brilliant data scientist and an even better science communicator, with a knack for crisp, readily understood explanations. How can a world of granular, highly varied data turn into a world of homogeneous choices? Simple: start with a set of items ("cuisines, genres, shop types") and a standard algorithm for sorting them. Let users choose from those recommendations. The mode (average) of those choices "gets shown more, so it gets picked more, so the model grows more confident the mode is what people want, and the tails starve." Run this for a few rounds and the evenly distributed catalog of choices "collapses onto one dominant option."

This is intrinsic in the choices we make in designing recommendation algorithms, tilting them towards the likelihood of a successful recommendation. A recommender that wants to succeed every time will make the safest possible recommendations, "so an algorithm that is uncertain about you, and it is always at least a little uncertain, hedges toward the average."

Then she busts out a beautiful, perfect little statistics aphorism: "Personalisation under a standard loss function is regression to the collective mean with extra steps." That is to say, "regression to the mean" (the tendency of varied things to become more standardized) cannot be avoided with the standard personalization algorithm. That algorithm is going to play it safe, showing you things that are broadly palatable, and because your choices are constrained to the average, you will choose average things.

This is how recommendation systems – and other analytical tools that produce predictions that are then turned into action – force so many diverse phenomena (streets, markets, media recommendations) into sameness. The fact that these recommenders are self-fulfilling prophecies means that "they don't have to be right," only "listened to."

This explains the sameness of so many of London's high streets. Leek examines 640 shopping streets, characterizing 18,000 food places spread out across them, flagging all the chain restaurants. Her analysis shows that any two London streets will, on average, share about half of their "food profile."

Obviously, this is most pronounced on streets with chain outlets, and it doesn't take that many chain outlets before a street's sameness shoots up: "A relatively small number of repeated names is enough to make otherwise different streets resemble one another more." So why do streets with chains resemble one another so much? Because the chains use an algorithm (weighting footfall, proximity to train stations, demographics, and competitors) to decide where to put their restaurants. If a street with a Gail's Bakery on it feels like every other street with a Gail's Bakery, that's because Gail's only puts its restaurants in places that have highly similar characteristics, measured to a high degree of accuracy and controlled by a narrow set of tolerances.

In other words, every street that feels like it should have a Gail's will eventually get a Gail's, whereupon that street will feel even more like all the other streets that have a Gail's, because it will share one more common factor with those other streets (a Gail's).

Leek points here to her earlier work on pub closures in the UK. The UK has experienced an epidemic of pub closures, with thousands of pubs disappearing since 2016:

https://laurenleek.substack.com/p/britain-lost-14000-third-places-they

Her research found that the biggest predictor of a pub surviving was its similarity to the median pub; which is to say that the more distinctive a pub was, the more "character" it had, the more likely it was to close. Pubs that are different from the average pub are harder to categorize, which means they're harder for a bank manager to assess for creditworthiness or for a landlord to justify extending a long-term lease to. The algorithms used to allocate capital and real estate are also recommenders, and they also drive variety out of the system.

This same phenomenon acts on culture. In an age of music recommendation algorithms, hit songs are changing; today's songs use a smaller vocabulary of unique words and repeat those words more often:

Vocabulary richness, distinct words relative to length, has fallen by more than a quarter since the early 1960s, while the share of repeated lines has climbed by nearly a third. The modern hit says less and says it more often, because the hook that works gets repeated.

But that's not the whole story! While each song resembles itself more ("saying less more often"), within that constraint, there's far more variety today than before: a given song's (constrained) vocabulary has grown more distinct when compared to all the other songs' vocabularies. Songs repeat the words they use, but the words repeated in songs are getting more different.

For Leek, this is the key to understanding the whole phenomenon and (more importantly) doing something about it. Music recommendation systems optimized for a singable hook, but did not optimize on any of the other variables in songs, so those dimensions acquired a broader range, even as the optmized variable got flatter and narrower.

This means that the tendency of recommenders to "flatten the world" isn't a single blunt outcome: it depends on which dimension we choose to flatten through recommendation, and who chooses to flatten that dimension.

A media recommender optimizes for consumption, showing you a tractable set of things it believes you'll watch, read or listen to. When you choose from among this limited set, the recommender takes note of that fact and shows you more of the same, pushing everything to a greige median. All the movies, books and songs you might have liked that were omitted from that initial set are excluded from being recommended in the future. The features of that media that you might have appreciated "decay out of consideration." They are never tested for desirability. The model collapses.

How badly does it collapse? Leek cites Movietweetings' data on which movies people watch: out of a million public movie ratings, half relate to the top 2% of movies in the set. There's 38,000 films in the set, but just 380 titles account for 40% of the ratings. Leek argues (persuasively) that this isn't because recommenders are good at "knowing your taste" – rather, they are good at "narrowing the menu."

Leek relates this to her work on creating LLM "personas" – synthetic populations meant to mimic the tastes and proclivities of real groups of people, that you can interrogate "before you spend money asking actual humans." While this would be useful for many applications, "it fails in exactly the way this whole essay is about."

Leek went to enormous lengths to reproduce the traits that make people interesting to study in aggregate, painstakingly replicating the ways that social connections, psychological outlook and demographic factors predict people's beliefs. The result was a set of LLM personas with "elaborate stories" about how they differed from one another, but whose survey responses about planned actions were homogeneous in a way that real populations are not.

This, Leek writes, is the same force that homogenizes other data-driven predictors. Because she'd ordered her LLM to reproduce the statistically validated relationships between different factors that predict a person's beliefs, each synthetic persona was a homogenized average. It's like the paradox of "The Average Man," where military uniforms sized to the average of all service personnel fit no one, because no one is average:

https://archive.org/details/DTIC_AD0010203

The thing is (as Leek points out) the idea that synthetic personas are a good way to understand the preferences of a real population is not a harmless delusion: it's a product that's being actively sold to governments, campaigning politicians and marketers. It's a self-fulfilling prophecy that drives governance, political campaigns and product design to the same homogeneous median that is making every shopping street in London feel the same.

This matters. As Leek writes, ecologists have long understood the importance of variety for systemic resilience: they call it "the insurance value of biodiversity." A diverse system has reservoirs of species and variation that may not be optimized for how things stand now, but that can move into niches created when things change in ways that lay waste to the previously dominant organisms. As anyone whose favorite banana went extinct can tell you, homogeneity works well, but diversity fails well:

https://en.wikipedia.org/wiki/Gros_Michel

The brittleness of algorithm-induced homogeneity is compounded by the fact that recommenders obscure the true preferences of people. If you watch two Scandinavian crime dramas after Netflix recommends them to you, it will keep showing you more Scandy crime for the next decade – even if there's another kind of programming that you'd vastly prefer (if only you knew about it). This means that decision-makers who choose which shows will get made in the future will keep on funding their safe Danish detectives, to the exclusion of whatever might emerge from the same weird attractor that produced the K-Pop Demon Hunter fortune.

Transpose this failure mode onto states, bank managers and landlords, and we see whole ranges of policies, businesses and activities that never come into existence, despite the popularity, prosperity and joy they might bring us.

But Leek doesn't end with this worrisome note. Instead, she identifies this whole thing – model collapse, placelessness, performativity, even Goodhart's Law – as an expression of one of the best-understood tradeoffs in computer science: "exploration vs exploitation":

Any system learning from feedback has to divide its effort between exploiting what already scores well and exploring options it hasn’t tried, in case they’re better.

Computer scientists have long understood that focusing on exploitation to the exclusion of exploration is a trap that locks you into "the first decent option" so you can never discover the best one.

Which means that this algorithmic homogeneity has a well-understood corrective: "forcing exploration back in." The problem is that markets hate this kind of exploration. A company that lives and dies by how many clicks it gets is never going to sacrifice 20% of its traffic by showing its users weird, untested options that score worse than the median because these weird things have never had a chance to prove that they are desirable.

This is a classic market failure, and, as Leek points out, there are regulatory responses in the UK (the Digital Markets, Competition and Consumers Act) and the EU (the Digital Services Act), both of which require the largest platforms to open up their recommendation systems, but so far, regulators have focused on "online harms" rather than variety (though the DSA does require platforms to offer algorithmic recommendations that are not based on your personal traits).

Leek identifies this willingness of states to set conditions for algorithm design as a means by which "exploration" can be forced back into the system. She's also bullish on interoperability, so that users can leave platforms with bad recommenders, without losing access to their media or social circles. As she writes, "the deepest discipline on a feed that has trapped you is the credible ability to leave it and take your data with you." I couldn't agree more:

https://pluralistic.net/2023/01/08/watch-the-surpluses/

She's less hopeful about individual responses. Demanding that you be an "adventurous consumer" is a way of letting systems off the hook. When every street has the same restaurants and every bookshop has the same books and the people in your life are all locked into one of two social media platforms, "choosing wisely" only gets you so far. Shopping isn't politics!

https://pluralistic.net/2026/05/21/purity-culture/#stop-fucking-that-chicken

Leek is a superb writer. After reading this piece yesterday, I sent it to half a dozen people and then read everything else in Leek's newsletter archives. Not only is it all brilliant, but I also realized that she'd written one of the most memorable articles about cities and platforms I've read in the last year, "How Google Maps quietly allocates survival across London’s restaurants – and how I built a dashboard to see through it":

https://laurenleek.substack.com/p/how-google-maps-quietly-allocates