One in every of my favourite rhetorical and analytical strikes is becoming a member of issues collectively (exhibiting that two totally different, seemingly unrelated concepts are facets of the identical phenomenon) and taking them aside (resolving a paradox by demonstrating that what seems to be one, contradictory factor is definitely two various things which were lumped collectively).
“Taking issues aside” is a really helpful framework for understanding AI. How can we resolve the (seeming) paradox that some expert staff report fantastic outcomes from their work with AI, whereas others are filled with dire warnings in regards to the lurking defects of their AI-assisted outputs? Easy: the primary group are “centaurs” (people who’re assisted by machines) and the second are “reverse centaurs” (people who’ve been pressed into service as peripherals for machines):
https://pluralistic.web/2025/12/05/pop-that-bubble/#u-washington
What are we to make of the individuals who’ve been fired by bosses who changed them with AI, in gentle of the truth that AI is demonstrably not capable of do their (former) jobs? Once more, it is easy in case you separate out two distinct phenomena: “AI can do your job” is the primary. The second is: “Your boss is a credulous dolt who’s infinitely attractive for changing lippy staff with pliable machines, which made him a simple mark for an AI salesman who satisfied him to fireside you and change you with an AI that may’t do your job”:
https://pluralistic.web/2025/03/18/asbestos-in-the-walls/#government-by-spicy-autocomplete
That is additionally a helpful transfer for understanding the AI funding bubble. It is not simply billionaires who do not assume different persons are as actual as they’re and consequently their jobs may be accomplished by chatbots. It is also billionaires who imagine that bosses may be bought AI and do not care if the AI is flawed, as a result of that is your boss’s downside after he buys the AI and fires you. They do not must imagine in AI in an effort to assume it is a good funding: like an investor betting that Joe Rogan can promote thousands and thousands of {dollars}’ price of peptides to determined younger males, they’re assessing the gross sales potential, not the deserves of the factor on the market:
https://pluralistic.web/2026/08/03/andor/#both
As helpful as “taking issues aside” is, “placing issues collectively” can also be an important method for assessing, critiquing and enhancing AI. In a stellar essay entitled “Temperature Zero for Tradition: Why Every little thing Is Beginning to Look the Identical” by the info scientist Lauren Leek, we get a top-notch instance of “placing issues collectively”:
https://laurenleek.substack.com/p/temperature-zero-for-culture-why
Leek’s essay is a kind of fabulous, wide-ranging, cross-disciplinary items, relating city design, music developments, artificial LLM crowds, Netflix suggestion algorithms, and a number of other different topics, all in search of to resolve a(nother) (seeming) paradox: how is it that we’ve got a lot potential selection, however every part is so manifestly the identical?
The reply is sophisticated and nuanced, however Leek’s foundational level is that in a data-driven society, “predictions” are self-fulfilling prophecies. As Leek places it: “As soon as prediction shapes the alternatives in entrance of us, we lose the power to inform the distinction between what folks needed and what the system made straightforward to need.”
It is a pervasive concern throughout many domains. Leek says that economists name it “performativity,” whereas machine studying researchers name it “mannequin collapse” and urbanists name it “placelessness.”
“Performativity” describes how, as soon as a market has been modeled by economists, that mannequin turns into the muse for financial coverage, which pushes the market to adapt to the mannequin:
https://press.princeton.edu/books/paperback/9780691138497/do-economists-make-markets
“Mannequin collapse” describes how machine studying fashions which can be educated on their very own predictions change into extremely bland, with all selection disappearing from the system’s predictions:
https://pluralistic.web/2024/03/14/inhuman-centipede/#enshittibottification
That is massively consequential: it is why bias proliferates by way of predictive policing algorithms: prepare a mannequin with information from racist stop-and-frisks and it’ll predict that every one the weapons and medicines in a metropolis are to be present in Black and brown peoples’ pockets. Flip these predictions into suggestions telling cops the place to go search for weapons and medicines and they’ll double down on racist stops, producing much more biased coaching information, which turns into nonetheless extra bias within the predictions:
“Placelessness” is the urbanist’s identify for “when all over the place optimises towards the identical template.” I consider it as Flinstones Syndrome, the place the identical background is looped behind Fred and Barney as they drive by way of Bedrock. In New York Metropolis, it is Citibank-bodega-Chipotle-Walgreens; within the Chicago suburbs, it is the strip malls with a Chili’s, a gasoline station, and an enormous field retailer.
Leek proposes that these are all expressions of the identical underlying phenomenon, a failure mode of knowledge science that takes a world of “granular private information” and arrives at a world the place “personalisation produc[es] extra sameness.”
To those wonderful examples, I would add one other one, from the world of financial coverage: Goodhart’s Regulation, which holds that “When a measure turns into a goal, it ceases to be measure”:
https://en.wikipedia.org/wiki/Goodhartpercent27s_law
Goodhart’s Regulation captures all kinds of phenomena. When Google first deployed Pagerank, they confirmed that by counting the inbound hyperlinks to all of the pages on the internet, you might extract a sign about which pages had been most necessary (as a result of there was no cause to hyperlink to a web page until you discovered it noteworthy).
However as soon as Pagerank turned the dominant means by which internet customers discovered pages, counting hyperlinks stopped being helpful: first, as a result of folks used Pagerank to seek out one of the best pages and hyperlink to them, making it unattainable for brand spanking new pages to get the inbound hyperlinks wanted to supersede incumbent pages; and second, as a result of it is easy for fraudsters to create inbound hyperlinks for low-quality pages in bulk, as soon as there is a cause to take action.
Counting inbound hyperlinks was a world-beating retrospective means of predicting which web page would finest match a searcher’s question, however as soon as it formed the world it sought to research, it ceased to be potential strategy to predict which web page would finest match your queries.
Leek is an excellent information scientist and a good higher science communicator, with a knack for crisp, readily understood explanations. How can a world of granular, extremely assorted information flip right into a world of homogeneous decisions? Easy: begin with a set of things (“cuisines, genres, store sorts”) and a normal algorithm for sorting them. Let customers select from these suggestions. The mode (common) of these decisions “will get proven extra, so it will get picked extra, so the mannequin grows extra assured the mode is what folks need, and the tails starve.” Run this for a number of rounds and the evenly distributed catalog of decisions “collapses onto one dominant possibility.”
That is intrinsic within the decisions we make in designing suggestion algorithms, tilting them in direction of the probability of a profitable suggestion. A recommender that desires to succeed each time will make the most secure potential suggestions, “so an algorithm that’s unsure about you, and it’s at all times at the very least a little bit unsure, hedges towards the typical.”
Then she busts out an attractive, excellent little statistics aphorism: “Personalisation beneath a normal loss perform is regression to the collective imply with further steps.” That’s to say, “regression to the imply” (the tendency of assorted issues to change into extra standardized) can’t be averted with the usual personalization algorithm. That algorithm goes to play it protected, exhibiting you issues which can be broadly palatable, and since your decisions are constrained to the typical, you’ll select common issues.
That is how suggestion programs – and different analytical instruments that produce predictions which can be then became motion – drive so many numerous phenomena (streets, markets, media suggestions) into sameness. The truth that these recommenders are self-fulfilling prophecies signifies that “they do not must be proper,” solely “listened to.”
This explains the sameness of so lots of London’s excessive streets. Leek examines 640 procuring streets, characterizing 18,000 meals locations unfold out throughout them, flagging all of the chain eating places. Her evaluation exhibits that any two London streets will, on common, share about half of their “meals profile.”
Clearly, that is most pronounced on streets with chain shops, and it does not take that many chain shops earlier than a avenue’s sameness shoots up: “A comparatively small variety of repeated names is sufficient to make in any other case totally different streets resemble each other extra.” So why do streets with chains resemble each other a lot? As a result of the chains use an algorithm (weighting footfall, proximity to coach stations, demographics, and rivals) to resolve the place to place their eating places. If a avenue with a Gail’s Bakery on it appears like each different avenue with a Gail’s Bakery, that is as a result of Gail’s solely places its eating places in locations which have extremely comparable traits, measured to a excessive diploma of accuracy and managed by a slim set of tolerances.
In different phrases, each avenue that feels prefer it ought to have a Gail’s will ultimately get a Gail’s, whereupon that avenue will really feel much more like all the opposite streets which have a Gail’s, as a result of it’ll share yet another widespread issue with these different streets (a Gail’s).
Leek factors right here to her earlier work on pub closures within the UK. The UK has skilled an epidemic of pub closures, with hundreds of pubs disappearing since 2016:
https://laurenleek.substack.com/p/britain-lost-14000-third-places-they
Her analysis discovered that the largest predictor of a pub surviving was its similarity to the median pub; which is to say that the extra distinctive a pub was, the extra “character” it had, the extra seemingly it was to shut. Pubs which can be totally different from the typical pub are tougher to categorize, which implies they’re tougher for a financial institution supervisor to evaluate for creditworthiness or for a landlord to justify extending a long-term lease to. The algorithms used to allocate capital and actual property are additionally recommenders, they usually additionally drive selection out of the system.
This similar phenomenon acts on tradition. In an age of music suggestion algorithms, hit songs are altering; at the moment’s songs use a smaller vocabulary of distinctive phrases and repeat these phrases extra usually:
Vocabulary richness, distinct phrases relative to size, has fallen by greater than 1 / 4 because the early Sixties, whereas the share of repeated traces has climbed by practically a 3rd. The trendy hit says much less and says it extra usually, as a result of the hook that works will get repeated.
However that is not the entire story! Whereas every music resembles itself extra (“saying much less extra usually”), inside that constraint, there’s way more selection at the moment than earlier than: a given music’s (constrained) vocabulary has grown extra distinct when in comparison with all the opposite songs’ vocabularies. Songs repeat the phrases they use, however the phrases repeated in songs are getting extra totally different.
For Leek, that is the important thing to understanding the entire phenomenon and (extra importantly) doing one thing about it. Music suggestion programs optimized for a singable hook, however didn’t optimize on any of the opposite variables in songs, so these dimensions acquired a broader vary, even because the optmized variable bought flatter and narrower.
Which means the tendency of recommenders to “flatten the world” is not a single blunt final result: it will depend on which dimension we select to flatten by way of suggestion, and who chooses to flatten that dimension.
A media recommender optimizes for consumption, exhibiting you a tractable set of issues it believes you may watch, learn or hearken to. While you select from amongst this restricted set, the recommender takes notice of that reality and exhibits you extra of the identical, pushing every part to a greige median. All the films, books and songs you may need preferred that had been omitted from that preliminary set are excluded from being really useful sooner or later. The options of that media that you simply may need appreciated “decay out of consideration.” They’re by no means examined for desirability. The mannequin collapses.
How badly does it collapse? Leek cites Movietweetings’ information on which films folks watch: out of one million public film rankings, half relate to the highest 2% of films within the set. There’s 38,000 movies within the set, however simply 380 titles account for 40% of the rankings. Leek argues (persuasively) that this is not as a result of recommenders are good at “understanding your style” – fairly, they’re good at “narrowing the menu.”
Leek relates this to her work on creating LLM “personas” – artificial populations meant to imitate the tastes and proclivities of actual teams of individuals, you could interrogate “earlier than you spend cash asking precise people.” Whereas this may be helpful for a lot of purposes, “it fails in precisely the way in which this complete essay is about.”
Leek went to monumental lengths to breed the traits that make folks attention-grabbing to review in combination, painstakingly replicating the ways in which social connections, psychological outlook and demographic elements predict folks’s beliefs. The outcome was a set of LLM personas with “elaborate tales” about how they differed from each other, however whose survey responses about deliberate actions had been homogeneous in a means that actual populations are usually not.
This, Leek writes, is identical drive that homogenizes different data-driven predictors. As a result of she’d ordered her LLM to breed the statistically validated relationships between various factors that predict an individual’s beliefs, every artificial persona was a homogenized common. It is just like the paradox of “The Common Man,” the place army uniforms sized to the typical of all service personnel match nobody, as a result of nobody is common:
https://archive.org/particulars/DTIC_AD0010203
The factor is (as Leek factors out) the concept that artificial personas are a great way to grasp the preferences of an actual inhabitants shouldn’t be a innocent delusion: it is a product that is being actively bought to governments, campaigning politicians and entrepreneurs. It is a self-fulfilling prophecy that drives governance, political campaigns and product design to the identical homogeneous median that’s making each procuring avenue in London really feel the identical.
This issues. As Leek writes, ecologists have lengthy understood the significance of selection for systemic resilience: they name it “the insurance coverage worth of biodiversity.” A various system has reservoirs of species and variation that will not be optimized for a way issues stand now, however that may transfer into niches created when issues change in ways in which lay waste to the beforehand dominant organisms. As anybody whose favourite banana went extinct can inform you, homogeneity works properly, however variety fails properly:
https://en.wikipedia.org/wiki/Gros_Michel
The brittleness of algorithm-induced homogeneity is compounded by the truth that recommenders obscure the true preferences of individuals. In case you watch two Scandinavian crime dramas after Netflix recommends them to you, it’ll hold exhibiting you extra Scandy crime for the subsequent decade – even when there’s one other sort of programming that you simply’d vastly choose (if solely you knew about it). Which means decision-makers who select which exhibits will get made sooner or later will carry on funding their protected Danish detectives, to the exclusion of no matter may emerge from the identical bizarre attractor that produced the Okay-Pop Demon Hunter fortune.
Transpose this failure mode onto states, financial institution managers and landlords, and we see complete ranges of insurance policies, companies and actions that by no means come into existence, regardless of the recognition, prosperity and pleasure they could convey us.
However Leek does not finish with this worrisome notice. As a substitute, she identifies this complete factor – mannequin collapse, placelessness, performativity, even Goodhart’s Regulation – as an expression of one of many best-understood tradeoffs in pc science: “exploration vs exploitation”:
Any system studying from suggestions has to divide its effort between exploiting what already scores properly and exploring choices it hasn’t tried, in case they’re higher.
Pc scientists have lengthy understood that specializing in exploitation to the exclusion of exploration is a entice that locks you into “the primary first rate possibility” so you may by no means uncover one of the best one.
Which signifies that this algorithmic homogeneity has a well-understood corrective: “forcing exploration again in.” The issue is that markets hate this type of exploration. An organization that lives and dies by what number of clicks it will get is rarely going to sacrifice 20% of its visitors by exhibiting its customers bizarre, untested choices that rating worse than the median as a result of these bizarre issues have by no means had an opportunity to show that they’re fascinating.
It is a basic market failure, and, as Leek factors out, there are regulatory responses within the UK (the Digital Markets, Competitors and Shoppers Act) and the EU (the Digital Providers Act), each of which require the biggest platforms to open up their suggestion programs, however thus far, regulators have centered on “on-line harms” fairly than selection (although the DSA does require platforms to supply algorithmic suggestions that aren’t based mostly in your private traits).
Leek identifies this willingness of states to set situations for algorithm design as a way by which “exploration” may be pressured again into the system. She’s additionally bullish on interoperability, in order that customers can go away platforms with unhealthy recommenders, with out dropping entry to their media or social circles. As she writes, “the deepest self-discipline on a feed that has trapped you is the credible capability to go away it and take your information with you.” I could not agree extra:
https://pluralistic.web/2023/01/08/watch-the-surpluses/
She’s much less hopeful about particular person responses. Demanding that you simply be an “adventurous shopper” is a means of letting programs off the hook. When each avenue has the identical eating places and each bookshop has the identical books and the folks in your life are all locked into considered one of two social media platforms, “selecting correctly” solely will get you thus far. Procuring is not politics!
https://pluralistic.web/2026/05/21/purity-culture/#stop-fucking-that-chicken
Leek is an outstanding author. After studying this piece yesterday, I despatched it to half a dozen folks after which learn every part else in Leek’s publication archives. Not solely is all of it sensible, however I additionally realized that she’d written one of the crucial memorable articles about cities and platforms I’ve learn within the final yr, “How Google Maps quietly allocates survival throughout London’s eating places – and the way I constructed a dashboard to see by way of it”:
https://laurenleek.substack.com/p/how-google-maps-quietly-allocates
I ought to have added Leek’s publication to my RSS reader once I learn that final December. I’ve rectified that oversight! What a implausible thinker, scientist and communicator! If she is not being relentlessly pestered by editors and literary brokers providing her a ebook deal, then it actually does show that the recommender programs are elevating the tasteless median over the totally, delightfully spiky outliers.
#25yrsago Terrible, silly Wired report on Dutch hacker camp https://internet.archive.org/internet/20011007084604/https://www.wired.com/information/tradition/0,1284,46033,00.html
#25yrsaog Wonderful NYT story in regards to the inside contradictions of the DMCA https://memex.craphound.com/2001/08/13/excellent-nyt-story-about-the/
#20yrsago Our defective instinct about open programs https://www.ft.com/content material/64167124-263d-11db-afa1-0000779e2340
#20yrsago Defending in opposition to the final plot received’t save us from the subsequent one https://www.schneier.com/weblog/archives/2006/08/terrorism_secur.html
#20yrsago NBC: Hair-gel terrorists posed no danger final week https://internet.archive.org/internet/20060813194630/http://www.msnbc.msn.com/id/14320452/
#15yrsago AT&T merger leak: it’s all about elevating costs and lowering competitors https://internet.archive.org/internet/20110920222524/http://www.broadbandreports.com/shownews/Leaked-ATT-Letter-Demolishes-Case-For-TMobile-Merger-115652
#10yrsago What’s inside a Tiki Fowl? https://miehana.blogspot.com/2016/08/fancy-feathers-restoring-tiki-room-birds.html
#5yrsago Finish of the road for Reaganomics https://pluralistic.web/2021/08/13/post-bork-era/#manne-down
#5yrsago Good cities are neither, 2021 version https://pluralistic.web/2021/08/13/post-bork-era/#our-streets
#1yrago Maga’s boss class assume they’re proof against American carnage https://pluralistic.web/2025/08/13/then-they-came-for-me/#boss-politics
“The Reverse-Centaur’s Information to AI,” a brief ebook about being a greater AI critic, Farrar, Straus and Giroux, June 2026https://us.macmillan.com/books/9780374621568/thereversecentaursguidetolifeafterai/
“Canny Valley”: A restricted version assortment of the collages I create for Pluralistic, self-published, September 2025 https://pluralistic.web/2025/09/04/illustrious/#chairman-bruce
“Enshittification: Why Every little thing Abruptly Obtained Worse and What to Do About It,” Farrar, Straus, Giroux, October 7 2025https://us.macmillan.com/books/9780374619329/enshittification/
“Picks and Shovels”: a sequel to “Purple Group Blues,” in regards to the heroic period of the PC, Tor Books (US), Head of Zeus (UK), February 2025 (https://us.macmillan.com/books/9781250865908/picksandshovels).
“The Bezzle”: a sequel to “Purple Group Blues,” about prison-tech and different grifts, Tor Books (US), Head of Zeus (UK), February 2024 (thebezzle.org).
“The Misplaced Trigger:” a solarpunk novel of hope within the local weather emergency, Tor Books (US), Head of Zeus (UK), November 2023 (http://lost-cause.org).
“The Web Con”: A nonfiction ebook about interoperability and Huge Tech (Verso) September 2023 (http://seizethemeansofcomputation.org). Signed copies at E book Soup (https://www.booksoup.com/ebook/9781804291245).
“Purple Group Blues”: “A grabby, compulsive thriller that may go away you understanding extra about how the world works than you probably did earlier than.” Tor Books http://redteamblues.com.
“Chokepoint Capitalism: Methods to Beat Huge Tech, Tame Huge Content material, and Get Artists Paid, with Rebecca Giblin”, on the way to unrig the markets for inventive labor, Beacon Press/Scribe 2022 https://chokepointcapitalism.com
“The Publish-American Web,” a geopolitical sequel of types to Enshittification, Farrar, Straus and Giroux, 2027
“Unauthorized Bread”: a middle-grades graphic novel tailored from my novella about refugees, toasters and DRM, FirstSecond, April 20, 2027
“Enshittification, Why Every little thing Abruptly Obtained Worse and What to Do About It” (the graphic novel), Firstsecond, 2027
“The Memex Methodology,” Farrar, Straus, Giroux, 2027
In the present day’s high sources:
At the moment writing:
“As soon as Is Enemy Motion,” a science fiction novel in regards to the origins of recent technofascism. In the present day’s phrases: 546 (4161 complete).
“The Publish-American Web,” a sequel to “Enshittification,” in regards to the higher world the remainder of us get to have now that Trump has torched America. Fourth draft accomplished. Submitted to editor.
A Little Brother brief story about DIY insulin PLANNING
This work – excluding any serialized fiction – is licensed beneath a Artistic Commons Attribution 4.0 license. Meaning you should use it any means you want, together with commercially, supplied that you simply attribute it to me, Cory Doctorow, and embody a hyperlink to pluralistic.web.
https://creativecommons.org/licenses/by/4.0/
Quotations and pictures are usually not included on this license; they’re included both beneath a limitation or exception to copyright, or on the premise of a separate license. Please train warning.
Weblog (no adverts, monitoring, or data-collection):
Pluralistic.web
E-newsletter (no adverts, monitoring, or data-collection):
https://pluralistic.web/plura-list
Mastodon (no adverts, monitoring, or data-collection):
https://mamot.fr/@pluralistic
Bluesky (no adverts, potential monitoring and data-collection):
https://bsky.app/profile/doctorow.pluralistic.web
Medium (no adverts, paywalled):
https://doctorow.medium.com/
Tumblr (mass-scale, unrestricted, third-party surveillance and promoting):
https://mostlysignssomeportents.tumblr.com/tagged/pluralistic
“When life offers you SARS, you make sarsaparilla” -Joey “Accordion Man” DeVilla
READ CAREFULLY: By studying this, you agree, on behalf of your employer, to launch me from all obligations and waivers arising from any and all NON-NEGOTIATED agreements, licenses, terms-of-service, shrinkwrap, clickwrap, browsewrap, confidentiality, non-disclosure, non-compete and acceptable use insurance policies (“BOGUS AGREEMENTS”) that I’ve entered into together with your employer, its companions, licensors, brokers and assigns, in perpetuity, with out prejudice to my ongoing rights and privileges. You additional signify that you’ve the authority to launch me from any BOGUS AGREEMENTS on behalf of your employer.
ISSN: 3066-764X
