For Kuleshov, the sequential juxtaposition of content lends meaning to images that may have nothing to do with each other.
And he conducted an experiment, the so-called Kuleshov effect, to highlight this principle. Kuleshov took a clip of Ivan Ilyich Mozzhukhin (the Ryan Gosling of Tsarist Russia) staring at the camera and intercut it with some other images: a bowl of soup, a girl in a coffin, an attractive woman. When he showed this sequence to an audience, the viewers noted how the emotional state of the actor changed from cut to cut. He's hungry. He's sad. He's got the hots for that lady.
The audience praised Mozzhukhin's emotive performance. But the actor's stare that Kuleshov used was the same in each cut. Here's an example of the effect:
Editing manipulates and creates meaning by connecting (potentially unrelated) content together. It provides emphasis and perspective.
EdgeRank is a frenemy
We know this isn't true. After all, Facebook uses EdgeRank to select what is shown to each user out of all the content available. The algorithm's whole purpose is to maximize an outcome by editing content together. And until recently, we've assumed that the objective of EdgeRank was more or less to maximize the engagement and relevance of posts in each user's stream. Ain't nothing wrong with increasing relevance! Sure, EdgeRank is an editor, but it's on your side.
You're Scorsese, and EdgeRank is a Thelma Schoonmaker. Or maybe not.
Facebook's PNAS disfunction
These posts aren't necessarily related to each other. They're just content generated by a user's friends, liked pages, etc. But just like Kuleshov did for film, Facebook can do for their network -- they can stitch images and text together into a stream that meets their needs, one that conveys a concept perhaps not present in the sum total of all the social content lying on the cutting room floor.
Now, people have a problem with this. The whole experiment feels like a violation. Facebook emotionally manipulated people. And to add insult to injury, Facebook used user-generated, supposedly perspective-free content (at least free of Facebook's perspective).
The counterarguments I keep hearing are these:
- Facebook's TOS covers such experiments
- Facebook was already using and continues to use algorithms to edit your stream. This is nothing new
- Facebook didn't create content to manipulate people. They used existing content
What is allowable in data science is only partially governed by TOSs and precedence. There's an inherent creepiness to data science, so it's important that a company always ask itself, "What do our users expect from us?" Not "what is legal?" or "can we point to what we've already been doing that's similar?"
Actually, Facebook may not be naive. They may just not care. After all, their customers whom they're trying to impress and engage with using data science are their advertisers, not their users.
Counterargument (3) is where the Kuleshov Effect comes into play. Editing is powerful. If you're stewing up a pot of social slop, then you have power over the final product. A stream is nothing more than a montage of social content that constitute its ingredients. And in the creation of that stream, Facebook wields immense power even though they create none of the stream's content.
Regardless of where you fall in the debate over whether this was an appropriate experiment, its results lead to a more haunting realization. Before we get to it, let's talk about the "content problem" present in data-driven targeted marketing.
Big data has a content problem
There's only one problem: the only reason to target someone at a personal level is if you've got personalized marketing content to show that person.
Understanding a person intimately and being able to target them is nothing without something to say.
And most companies don't have anything to say. Getting a marketer to finish one monolithic piece of creative is hard enough. Imagine needing personalized content for everybody!
So shortcuts are taken ("just write 'customers like you also bought this' and then use data science to pull some product suggestions) to produce "relevantized" generic content.
No matter how sophisticated data driven targeting products get, there will always be a content gap.
But Facebook may have found a shortcut. And this is where things get depressing.
Data science: a sheep dog, corralling people toward content
Rather than tailor marketing content to a user's unique emotional make-up, Facebook has shown that they can use tangentially related (and free!) user-generated content to push a user toward marketing content generated for a more general emotional state: insecure, hungry, lonely, etc. They can edit together photos and posts in a stream to skew a user's view of reality and shift them into one of these compromised emotional states.
In other words, if they can't use data to generate enough personalized content to target people, maybe they can use data to generate vanilla people within a smaller set of emotional states. Once you have a set of vanilla people, then your American Apparel ads will work on them without customization.
As Greg McNeal put it:
"What harm might flow from manipulating user timelines to create emotions? Well, consider the controversial study published last year (not by Facebook researchers) that said companies should tailor their marketing to women based on how they felt about their appearance. That marketing study began by examining the days and times when women felt the worst about themselves, finding that women felt most vulnerable on Mondays and felt the best about themselves on Thursdays.
The marketing study suggested companies should “[c]oncentrate media during prime vulnerability moments, aligning with content involving tips and tricks, instant beauty rescues, dressing for the success, getting organized for the week and empowering stories… Concentrate media during her most beautiful moments, aligning with content involving weekend guides, weekend style, beauty tips for social activities and positive stories.” The Facebook study, combined with last year’s marketing study suggests that marketers may not need to wait until Mondays or Thursdays to have an emotional impact, instead social media companies may be able to manipulate timelines and news feeds to create emotionally fueled marketing opportunities."
This is part of the dehumanizing effect of AI and big data I wrote about a while ago. Rather than data being used to make computers more intelligent, data is being used to make humans more predictable (read: more stupid, unhappy, and willing to buy something to palliate their discontent).
Yann LeCun, who runs Facebook's AI lab, said I'm utterly wrong on this point. In his response to my last post, he contends:
"The promise of ML (and AI) is, on the contrary, to let humans be more human, to free their minds from having to reason like machines, and to let them concentrate on things that are uniquely human, like communicating with other people."
In this particular study in PNAS, we can see that the promise of data modeling at Facebook is not to "let humans be more human." It's not to "free their minds."
All of that machine reasoning isn't trying to make us more human so much as it is trying to make us more sad and predictable. And just wait until deep learning applied to image recognition can recognize and stream my selfie at Krispie Kreme next to a tagged photo of me and my love handles at the beach. Data-driven inferiority complexes for all!
The promise of data modeling at Facebook is to place us in chains made from the juxtaposition of our own content. We'll be driven into pens made of a few profitable emotional states where marketing content waits for us like a cattle gun to the skull.
That said, where else am I going to share photos of my kids with old friends? Can't do that on Twitter...I only use Twitter to express faux indignation and convenient morality concerning trending causes. Looks like I'm stuck with Zuck.