We awake every day to find things basically the same today as they were yesterday, and breaking free from that routine means redirecting our attention. Phil Connors (played by Bill Murray) was a weatherman for a local television station in the movie Groundhog Day. Sent to cover a Groundhog Day ceremony, Connors becomes trapped in a single repeating day: February 2. Every day he wakes up and the people around him do the exact same things they did the day before, with the exact same results. All he can control is his own focus, effort, and attention. He can perfectly predict each day, but initially can’t find his way out of the loop.
Many of our daily choices are decided by recommendation algorithms. Knowingly or unknowingly, we delegate decisions to these bits of software. We let them decide where we buy things, what we buy, films we watch, songs we listen to, restaurants we go to, where we travel, and so much more. These systems are at work in the applications you use everyday, from TikTok to Google Maps. It's surprising how much of our lives now rely on them.
When you open Amazon to shop, the items you see are carefully chosen by their algorithm based on everything it knows about you and your shopping habits, and Netflix has spent billions perfecting its movie recommendation system. Spotify pushes music to you that it knows you’ll like, disappointing teen boys everywhere whose mixtapes couldn’t compete. Uber even tries to match you with the right drivers and cars. We assume that all of these recommendations are an attempt to give us what we want. But that is not always true.
Dating apps use recommendation algorithms to match potential partners. Smart thermostats learn from when and how you adjust the temperature in your house. Health insurance companies use algorithms to steer doctors toward less expensive tests and referrals. Airlines have one that pushes you toward selecting certain seats on the plane. Drive-through menus have gone digital, adjusting price and availability constantly. With the ever-increasing use of artificial intelligence, these recommendation algorithms keep getting better. But it’s worth asking this question: better at what, exactly?
Political economist and social data scientist Dr. Lauren Leek argues that these recommendation algorithms are designed primarily to avoid being wrong, and often guess at what you will like by relying on what is generally popular. The popular choice then gets picked more often across everyone, and that is treated as proof that the algorithm’s choice was correct. Meanwhile the lesser-known, more eccentric, or more creative options fade out of view. Leek’s research demonstrates this using London high streets.
Comparing a map of 640 main streets of London, her research mapped 18,000 restaurants. She flagged any restaurant name appearing three or more times across the city. The research found that if you pick any two main streets in London, on average, they will share almost half of the same food profile. Leek ranked the streets from the one with the smallest share of chain restaurants to the one with the largest. Then she split the ranked list into four equal groups; group one had the fewest chains, and group four had the most. The streets in group one had the most distinctiveness overall, while the streets in groups two, three, and four were remarkably similar to each other and to the rest of London.
Leek calls this a convergence layer. A street needed only a couple of chain restaurants to lose its individuality. In other words, just a couple of the common chain restaurant names on a street gives it a lot in common with all the other streets. Like the recommendation algorithm that defaults to feeding you what is most popular, chain restaurants seem to pick locations that resemble other locations where they’ve had success. The restaurants use data to decide where a new branch is most likely to succeed. The answer, of course, is locations that have high similarity to where the chain has done well. And each new location makes the street more similar to other streets in the city, and that uniformity makes the next predicted location an even safer bet for the chain.
What I notice here is that the chain restaurants are not optimizing for customer preference. They are using certainty of profitability as a proxy for customer preference. It doesn’t matter whether you prefer Starbucks coffee; it’s the choice that is right there when you want coffee. It has to be just good enough that you’ll stop when you pass it on your way to work, stumble out of your hotel on the first day of vacation and see it right across the street, or notice one inside the Target store where you buy your bath towels and toilet bowl cleaner. The location matters more than the quality of the coffee, and most of us will accept the most popular option. And it is most popular because of its carefully chosen locations. This self-reinforcing loop makes everything in our world move toward the average.
Recommendation algorithms do not discover your preferences, they create them. And once they’ve created the preferences, they reinforce them because they are popular, causing them to grow ever more popular. In sociology, we know this as performativity. It is a predictive model that alters the environment until its prediction comes true. This makes each day’s activities look less like conscious choices, and more like Phil Connors reliving February 2 over and over again. At one point in the movie he abandons all hope of tomorrow being any different from today.
The health insurance algorithm optimizes to reduce its costs, not to increase the quality of your healthcare. The dating app wants you to match, but not for a long-term relationship because they need to keep you using the platform. Your thermostat does learn from your home temperature adjustments, but through agreements with electric utility companies, it also nudges small temperature changes based on directives from the power company. Digital menus aren’t just showing what they think you’ll like most, but rather the most profitable item that you will be willing to eat. The airline gets you to pay more for an assigned seat by separating you from a traveling partner or sticking you in a middle seat. The most popular restaurant on DoorDash is probably the ghost kitchen with a 30% platform commission rate or the restaurant that is not getting enough DoorDash orders and is at risk of switching to a different delivery platform. When we accept these popular choices, they feed back into our profiles as preferences and are used to justify giving us the same choices next time.
Claude, ChatGPT, and all the artificial intelligence tools don’t make this better. By their very nature, these tools all pull toward the average. They are trained on the world’s data and rewarded for giving the answer that is most popular, and therefore most safe. Currently this shows up most in A.I. writing. As Claude has gotten increasingly capable at coding, its English prose has become horrific to the point that it has been called “Claudish.” The writing that comes from the most advanced models in the world is decidedly average. And the companies that wrote basic recommendation algorithms are using these models to make their algorithms more powerful and accurate. This power gives companies even more control over what popular means, and over what you do, watch, hear, eat, and so much more. The power makes them even more capable of creating the consumer preferences that will optimize their bottom-line.
In the movie, Phil Connors eventually breaks out of his February 2 loop, but not before some difficult times. Going through a phase of depression and nihilism from all the sameness, the weatherman says, “You want a prediction about the weather. . . I’ll give you a winter prediction: It’s gonna be cold, it’s gonna be grey and it’s gonna last you for the rest of your life.” But eventually Connors realizes that although he wakes up every day and the people around him all do and say the same things they did the day before, he still has control over his daily choices. He uses his limitless rehearsal time to learn to play the piano, stops trying to manipulate his way to a relationship with the person he loves, finds unusual ways to show generosity to the people around him, and spends an authentic and transparent day with the woman he’s fallen for. Even within the forced framework each day brought, Connors discovered he still had the freedom to make different choices.
I like to call this embracing serendipity. I cannot change what the algorithms feed me, and many times I know I’ll accept the popular without knowing what the other choices really were. But in big and small ways, I know I can also reintroduce authenticity through noticing beauty in the people, things, and places I wasn’t prompted to see. I can discover my preferences by an unplanned walk that leads me into a local bookstore I’ve never heard of. I can try a restaurant that doesn’t have any reviews yet. I can find a place to live by walking through a neighborhood. I can decide to see a movie I don’t think I’ll like. Or go to a library and not search for anything in particular. I might discover that what is popular is not always my preference. Who knows? Maybe things tomorrow won’t be the same as they were today.