You May Also Like (8 page)

Liking is learning: This truism runs from entire cultures down to the individual. The exposure effects begin even before we are born. Like carrot juice as an infant?
Chances are your mother did. The odors and tastes were all around you, in the atmosphere of amniotic fluid that was your earliest dining experience.
Trained sensory panelists can even tell which women have consumed garlic pills based on the scent of their amniotic fluid alone.
Out of the womb, we strain toward the things we prefer (that is, the familiar) and make “aversive gapes” at the things we dislike.
Making faces is part of the social experience of liking and, especially, disliking: We send cues about what we are eating and look for information about what others are eating.

Simply seeing other people eating something seems to promote liking. In a classic study looking at the feeding of children in a women's prison in the 1930s, children's preferences seemed to be informed by whoever was feeding them: “
Babies who refused tomato juice were found to be fed by adults who also expressed a dislike for tomato juice.”
In a study of preschoolers, a “target” child who preferred one vegetable to another was seated with three classmates who had the opposite preference.
By the second day of the study, the target child had already switched preferences. Exposure to people, as much as food itself, influences our liking.

Mysteries still abound in our liking for food. Consider the simple question of why we should suddenly like something that we previously disliked. Very few of us “like” a substance like coffee or beer the first time we drink it, but many of us come to like it. All tastes are, in essence, “acquired tastes.” Or, as Pelchat suggests, “an acquired
liking
is really what we should say.”

And when we talk about “acquired tastes,” we should really be talking about “acquired flavors,” as Dana Small, an associate fellow in Yale University's John B. Pierce Laboratory who studies the neuropsychology of eating, suggested to me. We are not born knowing about flavors like coffee; we simply know the drink as bitter and thus bad. “The bitter is there as a sign that there is a potential toxin in whatever you're sampling,” she said. “You just want to know that; you don't want to have to learn that.”

But no one is born liking, or not liking, chicken feet. Those “gatekeeper” taste systems, after all, would not likely know feet from wing. It is all chicken. Before food even gets to us, culture has done that first big sort, sifting out the boundaries of what is acceptable to like. “
The French eat horses and frogs but the British eat neither,” notes Jared Diamond.
As with any food, the French, during a discrete historical moment, had to be taught to “learn to like” horse as food. But what we like in
taste
, as opposed to flavor, is remarkably similar around the world. As John Prescott writes in
Taste Matters
, “
The sweet taste of sucrose in water, is optimally pleasant at around 10–12 percent by weight (approximately the same as is found in many ripe fruits), regardless of whether you are from Japan, Taiwan, or Australia.”

Flavor conditioning helps us to like or dislike flavors. The benefit of this is, as Small put it, that we can “learn to like the foods that are available to us, and avoid particular foods rather than entire classes of nutrients.” When she was young, she went to a popular sailing event in her hometown of Victoria, British Columbia. With college friends, she partook of one too many drinks of Malibu and 7UP, an unholy and
intensely cloying concoction of sweet, coconut-flavored rum and citrusy soda pop. “
That was twenty years ago,” she recalled. “I can't even wear coconut suntan lotion. It makes me ill.”

Through a complex chain of activity in the brain, she said, we learn “flavor objects”—the “perceptual gestalt” of touch, taste, and smell in everything we eat. “Did this food make me sick? Did this food give me energy? You learn preferences based on the entire flavor object.” The flavor object itself is “created” by a network of neural activity, described as “
a distributed circuit including the neural representation of the odor object, unimodal taste cells, unimodal oral somatosensory cells, multimodal cells, and a ‘binding mechanism.' ” You do not just “taste” a strawberry; you virtually conjure it into being.

Coffee—the actual substance—becomes no less bitter the hundredth time we drink it than the first time we drank it. But something happens. “It
becomes
coffee,” Small said. “The brain has learned that coffee is not a potentially harmful signal.” Many of us, when first drinking coffee, add things that we like—milk and sugar. This not only weakens the bitterness but helps build positive associations with the coffee. The relationship is one-way, notes John Prescott: We do not learn to like sugar by drinking coffee; we learn to like coffee by drinking it with sugar. Add the post-ingestive signal of caffeine, and you have got a drink that we like, almost as if in spite of ourselves. You may be thinking the pleasures of caffeine or alcohol are enough to explain why we become conditioned to liking coffee or whiskey. But then why not just add those substances to what we already like? Why is it the things that are most disliked in the beginning go on to be the things we like the most?

There must be a moment when our disliking actually shifts to liking. Small has been trying to locate it in neurologic time and space.
In one experiment, she had subjects try novel-flavored beverages that had no calories. After a few weeks, she added caloric but tasteless maltodextrin to one of the flavors. Even though they could not sense its presence, subjects liked the beverages with maltodextrin more. As with Pelchat's tea study, the “post-oral signal” coming from the gut—which is happily converting the maltodextrin into glucose—changed liking.

In Small's study, though, the beverages were all chosen to be “slightly liked.” This still does not answer how we get from disliking to liking. What if you could take a food that is intensely disliked and, in the flick of a switch, suddenly generate an intense desire?
Kent Berridge,
a neuroscientist at the University of Michigan, did just that in a Pavlovian conditioning experiment with rats. First, rats got “pulses” of a pleasant sucrose solution, along with a sound. They also got a deeply unpleasant, three-times-as-strong-as-seawater solution of “Dead Sea salt,” accompanied by a different sound.

The rats
hated
the salt—so much so that it had to be delivered into their mouths via an “implanted cannula.” And when the rats subsequently heard the respective tones, they turned either away or toward the food source and made the appropriate facial expression. Next, the rats' brains were altered with injections that triggered a kind of simulated extreme salt craving. The following day, when the rats were again presented with the tones, they immediately moved toward the Dead Sea salt, making vigorous lip-licking “pleasure” faces (the same as seen in human infants)—
before they had even tasted it in its new, “pleasant” state
. In other words, without even knowing that they liked it, they suddenly wanted it.

This might help explain not just addictive behaviors but everyday liking. In one study, Berridge and colleagues asked student subjects to identify the gender of faces seen on a computer screen. They were also shown, rather surreptitiously—at one-sixtieth of a second—angry or sad faces. Afterward, they were given a fruit beverage, which they were told was in development by a soft-drink company, and asked how much they liked it. Subjects who saw the “happy” faces reported liking the drink 50 percent more than subjects who saw the sad faces. The happy faces triggered “
mesolimbic circuits of ‘wanting' in the brains of students who viewed them, which persisted for some minutes undetected as students evaluated their own mood,” Berridge wrote. “The ‘wanting' surfaced only when an appropriate target was finally presented in the form of a hedonically laden sweet stimulus they could taste and choose to ingest or not.” It was as if they were, to paraphrase the old country song, “looking for like in all the wrong places,” finally finding something in which to express their interest.

These kinds of mechanisms might help explain how disliking turns to liking. “Tastes” enter the brain far “downstream.” Even babies possessing not much more than a brain stem are “making both a recognition and evaluation decision.” But they are not, as Berridge suggested to me, forming that “flavor object.” That happens somewhere “upstream.” In a classic study by Ivan de Araújo and colleagues, people were given
short whiffs of a mix of isovaleric acid and cheddar cheese and told it was cheese or body odor. Those in the “body odor” condition (so to speak) rated the compound lower than those who were told it was cheese. No surprise there. But the cheese people also had more activation in a wider network of brain regions, which reflects a common finding that “liking” seems to activate a larger chain of brain activity than disliking. It is as if we need to expend more energy to figure out why we like, rather than dislike, something.

Body odor and cheese read differently in the brain. But for the first steps of mental processing, “the signal is going to be the same,” Berridge said. “That signal could be modified pretty early in the pathway, however, by expectation and anticipation. How long is that signal still the same by the time you get into various parts of the brain?” So strong are those expectation and anticipation overlays, of course, that in the de Araújo study, even people who were given a “clean smell,” but were told it was cheese or body odor, had similar patterns of brain activity. They were readying themselves to like or dislike a smell that never came, a phantom pleasure or displeasure.

“In the end,” Berridge told me, “we are aware of a final product, but we're not aware of the process that gave us that product.” The bitter signal at the brain stem is the same, but somewhere in the higher cognitive processes “coffee” takes shape. Learning is interacting with taste to cause a pleasure. “Whatever pleasure we are getting is probably coming from the same basic pleasure circuits that the sweetness has a special privilege tapping into,” Berridge said. The brain has sweetened your coffee.

The idea of finding some precise moment and place where disliking turns to liking is complicated by the fact that the same areas of the brain that are excited by liking are also activated by disliking. The amygdala, for example, seems to respond in equal measure to things we like and things we dislike. Perhaps someday scientists will discover a
meh
circuit—a discovery that suggests that at heart we are actually fairly ambivalent about most things and that it is some particular firing of synapses, or the person we had lunch with that day, or what song is on the radio, that eventually pushes us one way or the other.

It is striking to realize how strongly we stand by our likes and dislikes,
given how open they are to distortion and manipulation, as much by our own brains as by outside influences. Perhaps we instinctively sense the fragility and arbitrariness of these preferences, and so cling to them even harder. What is clear is that food is where we find the most intensely personal relationship with our own taste, literally and metaphorically. As Beauchamp had told me, at Monell, “The most important decision every human being makes every day is whether to put something in their mouth or not.” This was once a life-or-death decision; now it is just personal taste.

And yet that only seems to make the decisions that much more elaborate, bringing that much more insecurity to our choices. Back at the Chinese restaurant, Rozin described our “affective” relationship with food as “very fundamental, very basic, and it's frequent. Not as frequent as breathing, but breathing isn't a matter of taste.” He paused to collect the last bit of sweet-and-sour shrimp, plucked it into his mouth, then added, “Same hole.”

LIKING IN A NETWORKED AGE

 

One night, as I trolled for something to watch on Netflix, a film called
The Rocking Horse Winner
popped up (“Because you enjoyed:
Psycho, Annie Hall, Fargo
”). I clicked through to find a 1949 British adaption of a D. H. Lawrence story about a boy who could predict the winners of actual horse races by riding his toy horse faster. The story, like the film, was new to me.

Here, I thought, was the genius of algorithmic recommendation systems: picking some obscure film out of the historical dustbin, based on some unseen alchemy that was beyond me. What linked
The Rocking Horse Winner
to Woody Allen's iconic comedy, Alfred Hitchcock's shocker, and the Coen brothers' dark midwestern gothic? And what in my own rating activity had summoned this cinematic
ménage à quatre
? What if I had
loved
the Hitchcock but not liked
Annie Hall
—would that have triggered some other recommendation?

Greg Linden, who helped create Amazon's pioneering algorithms, reminds us that we should not ascribe them too much power in finding some odd, uncannily prescient suggestion. “The computer,” he said,
“merely performs an analysis of what humans are doing.”
And yet their own creators have admitted that the ever more complex mathematical regimes can become, in effect, HAL-like “black boxes” whose precise behavior can no longer actually be determined or predicted (something we humans can at least identify with).

I have occasionally bristled at the recommendations of Netflix—
an Adam Sandler film? Are you kidding me?
But the flip side of having access to so many films is expending more time in deciding what to watch. And so I have come to accept that in an age of often bewildering choice in which I no longer have time to read back issues of
Cahiers du Cinéma
or flip through cutout bins at record shops, there might be some benefit to off-loading some of my decision making and discovery process to a computer, the way we have largely outsourced our memory lapses to Google.

For a time, anyway, I rigorously trained my Netflix algorithm. I rated each film I had seen and studied the predictions of what I might like with Talmudic intensity. I wanted the thing fine-tuned, able to handle the twisting contours of my taste profile. I wanted it to know that just because I loved
The Evil Dead
did not necessarily mean I liked most other horror films. I wanted it to know not just that I liked something but
why
I liked it. I wanted more than it could give.

And so when I found myself at Netflix's headquarters, in a red-tile-roofed building—half old Hollywood, half La Quinta Inn—in Los Gatos, California, I had rating stars in mind. They were a borderline obsession. I would spend strange amounts of time pondering whether the 2.9 predicted rating for me warranted seeing the film (the distance from 2.9 to 3.0 had a gnomic power). Watching a film in the 1 ratings seemed almost illicit. And when I came across a film I had not seen but that had a 4.7 predicted rating for me, I could feel the room move.

I knew I was not alone: The company had awarded a million dollars, in its famous “optimization prize,” to the computer scientists who devised a 10 percent improvement to Netflix's predicted ratings. Many smart people had invested hours thinking about things like the so-called
Napoleon Dynamite
Problem—or what to do with movies that seem to polarize taste along less than predictable lines. Here, in Los Gatos, was, I imagined, a kind of benevolent Stasi of taste that knew everything about how people watched films, a massive repository of human predilection. I wanted to know things that I knew were proprietary
and they would not tell me: How responsive was the algorithm to rating? If I gave a movie 2.7 stars that Netflix had predicted was for me 3.2, how quickly would this divergence ripple through my rating ecosystem? What film had the widest distribution of extremely negative and extremely positive ratings?

All this is why I could practically hear the needle scratching across the record when I sat down with Todd Yellin, the company's vice president of product innovation, in the
Top Gun
room (all of Netflix's rooms are named after films or television shows), and he proceeded to tell me, “My first job here was director for product personalization. I led the effort of how to get ratings, how to get better predictions out of those ratings, where to put them in the user interface.” So far, so good. Then he said, “As we've broadened the scope of personalization, we deemphasized ratings predictions over the years.”

I let it sink in.
Deemphasized
. I probably looked a bit crestfallen. I could tell Yellin sensed my disappointment. I had come here to understand the world's most sophisticated engine of predictive taste in movies, and I was being told that taste—at least as expressed through ratings—was being deemphasized. “We've gotten more people to hit stars than anyone in the universe on movies and TV shows,” said Yellin. “And we've come up with many algorithmic recipes to improve the accuracy of that prediction.” But that, he said, was state of the art circa “2005 or 2006.” My geeky “stars” questions suddenly seemed like horribly out-of-style fashions. So after all that extensive time and effort toward building the perfect ratings-based recommendation system, Netflix walked away?

Not quite. “We still have people rate, we still find that very useful information,” said Yellin. “It's just secondary.” Two things happened to dim the usefulness of stars. The first, suggested Xavier Amatriain, the company's director of recommendation systems, is that the company was getting close to some kind of terminal velocity in taste prediction. “It's pretty much like many things in the algorithmic world,” he told me. “It takes you 20 percent of your time to get to that 90 percent accuracy, then 80 percent of your time you're trying to get that final 10 percent accuracy.” It was less than clear that the investment in getting that final 10 percent, and the complexity it would add—to a recommendation system already groaning with “Restricted Boltzmann Machines,”
“Random Forests,” and “Latent Dirichlet Allocations”—would actually pay off.

Something else had changed. Since the Netflix Prize, Netflix had gone from a strictly DVD-by-mail company to, largely, a video-streaming service. “What people wanted to do when they were giving us a rating,” Amatriain said, “was expressing a thought process. You added something to your queue; you watched it two days later. And then you were expressing an opinion you knew was going to have feedback in the long term.” But with instant streaming, “it's a very different concept. You don't like it, it's okay. You just switch to something else. The cost of switching is much lower.”

With streaming, Netflix might have had less explicit feedback, but it had more implicit
behavior
. “We're able to get real-time data play,” Yellin said, “which is richer than what they say about what they want.” Netflix knows infinitely more about what and
how
you watch: when you watch something, where you watch, the moment you stop in a movie, what you watch next, whether you watch something twice. What you
search
for—another taste signal. Yellin tells me all this with passionate intensity. With his jittery patter and angular, slightly drawn look, accentuated by a lack of hair, he comes across like a hyper-knowledgeable video store clerk, back in the days when they existed. But he is a video store clerk who has been given an omnipotent glimpse into what the people in this country have been sticking into their VCRs—and what parts they hit “rewind” on. If there is a violation of privacy here, its most salient feature is that you cannot hide from your own taste.

The arrival of companies like Netflix, with their petabytes of data on people's likes and dislikes, all those thumbs and favorites, offered unprecedented insight into what had long been an elusive field: the formation of judgments, the expression of preferences, the mechanics of taste. This vast range of online activity—“electronic word of mouth,” as it is called—is where abstract, “unaccountable” notions of taste run into the empirical order of the Internet, with its collaborative filtering algorithms, its sprawling data sets, its seemingly perpetual record of activity. Any one review—or thumb or like—is essentially useless. It suffers from what Ray Fisman has called
the “cheap talk” problem. The aggregate level is where, through sheer numbers, the noise can be filtered, the outliers marginalized, and statistical consensus achieved.

Sociologists like Pierre Bourdieu, who probably thought more about taste than anyone else (and whom we shall revisit later), had always faced the problem of self-reporting: Asking people what they like is not the same thing as observing what they do. The beauty of the Internet is that regardless of what people say, you can see, with increasing fidelity, their actual behavior. Almost every aspect of human taste that Bourdieu was interested in is, every day, being cataloged online, in numbers beyond any sociologist's dream. What music do you like? (Spotify, Pandora.) What is your ideal human face? (OkCupid, Match.​com.) What is the ideal subject of a photograph? (Flickr, Instagram.)

So where Netflix once relied much more on what people
said
they liked—the itself rather novel bedrock upon which recommendation systems were founded—it was now focusing more on what people actually
watched
. “There are many advantages to this,” says Amatriain. “One is the way people rate: They rate in an aspirational way—what they would like to be watching or how they would like to be watching.” As Carlos Gómez Uribe, Netflix's director for product innovation, told me, “A relatively high fraction of people tell us that they often watch foreign movies or documentaries. But in practice that doesn't happen very much.”

Netflix had always sensed this gap between people's aspirations and their behavior. It could, to cite one example, track how long a DVD sat at a user's house, presumably unplayed. “Al Gore's
Inconvenient Truth
,” said Yellin, to nods around the table. “We noticed that movie would stay at people's houses forever. That was a great cup holder.” But now the level of scrutiny was more real-time, more visceral: You quit that Bergman film for
Dodgeball
? You just created a data point.

People, Yellin suggested, “want to feel good about themselves. Or they could even be self-delusional about the image they have of themselves—what kinds of things they say they like, how many stars they'll give a particular title, and what they actually watch.” You might give five stars to
Hotel Rwanda
and two stars to
Captain America
, he said, “but you're much more likely to watch
Captain America
.”

There is nothing particularly new in this. Economists since Thorstein Veblen have talked about our conspicuous “signals” of taste, whether honest or not. They usually only flow upward: People do not
rate
Captain America
five stars,
Hotel Rwanda
two stars, and then secretly watch the latter. The sociologist Erving Goffman famously described the way we present ourselves as a “dramaturgical” act: “
We find that upward mobility involves the presentation of proper performances and that efforts to move upward and efforts to keep from moving downward are expressed in terms of sacrifices made for the maintenance of front.”

We all have wanted, at one time or another, to appear as an idealized self. “
I'm actually a quite different person,” as the playwright Ödön von Horváth wrote, “I just never get around to being him.” Think of the moment in
Play It Again, Sam
where Woody Allen's character is scrambling, ahead of a date, to array his coffee table with respectable books (“You can't leave books lying around if you're not reading them,” his friend complains, to which he replies, “It creates an image”). What is curious about the Netflix data is that they were private; there is no one to see your tasteful choices or interesting queue. As Yellin suggested, the dramaturgy involved here is directed at one's
self
.

Which leads to the interesting question posed by the anthropologist Robert Trivers and his psychologist colleague William von Hippel: “Who is the audience for self-deception?” Goffman wrote that people are often compelled to maintain standards “because of a lively belief that an unseen audience is present who will punish deviations from these standards.” Hence the guilt in “guilty pleasure,” a subject I will treat later. If deception itself is an evolutionarily useful strategy, “fundamental in animal communication,” self-deception too becomes “
an offensive strategy evolved for deceiving others.” Woody Allen's character, by displaying those books, is making himself feel better, as well as helping to convince
himself
that he is the sort of person who reads those books, which will thus help convince his date.

This does not mean it cannot be jarring when the mirror is held up to self-deception. A rather common complaint Netflix hears is “Why are you recommending all these two- or three-star movies to me?” In other words, why are you giving me stuff I will not like? Netflix, however, is not in business to turn you into a cineast. It wants to keep you signed up with Netflix. It is like a casino using clever math to keep you on the machines.
It wants to recommend things you will watch. “Engagement,” Netflix calls it. “When someone rates a movie like
Schindler's List
,” says Gómez Uribe, “it tends to be pretty high—as opposed to one of the silly comedies I watch, like
Hot Tub Time Machine
.” But if you
give people nothing but four- or five-star films, “that doesn't mean that user will actually want to watch that video on a Wednesday night after a long day at work.”