Commonsense and Statistical World-making

Marx famously claimed that philosophers only interpret the world, whereas the point is to change it.

However, many philosophers do look at the relationship between self and world as a kind of metabolism. A phenomenologist, for example, would argue that there is no perceiving (or, by extension, judging) the world without changing it. Perception and cognition are never passive, they are inherently intentional, tool-like processes. We work in a kind of loop with the objects of our interest. We see a thing in one state, we poke it so we can see it in another state. Or we change our own state to see if the object changes relative to it. Or we wait until time changes both.

Philosophies do indeed change the world. On the purely speculative side, we have words like “idea” and “real,” which are drawn from a dialect of Platonic, which long ago filtered into everyday use, to the point that we cannot even think of thought without the concept of the idea or of existence without the concept of the real. Philosophy has changed the world by changing our common sense.

Philosophy’s descendants–psychology, sociology, political economy, etc.–all rely heavily on tons of untestable theory. Even theories from the mathematical sciences invariably conjure a world rife with assumptions about order, causality, time and space, elegance, simplicity, etc. And the so-called revolutions in science are just such reversals in our most basic concepts of the world–in our common sense. Now we are not the sump of the universe, now a triangle can have three right angles, now a thing might be in two places at once.

As they say, everything’s a nail to a hammer. Whatever conceptual tool we use to engage the world remakes the world in its image. If, today, the so-called soft sciences can only gain a veneer of respectability by applying statistical methods to their theories, then these theories will inevitably tend to fashion something sharp around the edges. But that does not mean that a sociopolitical theory is only that which can be quantified. To apply a Marxist analogy, the quantifiable parts of a theory only represent its exchange-value, which is different from its value per se.

For example, you may have two opposing theories about gun-ownership, which qualitatively have little in common. One side might argue that freedom from tyranny is the only value that really matters, while the other side might argue that guns are inherently hateful things[1]. Thus we have an opposition between too qualitative values. But for rhetorical efficacy and putatively for the sake of scientifically-informed policy decisions, the complex of ideas and values surrounding gun ownership is reduced to a magnitude, which is in turn reduced to a boolean: how many people do guns kill, then, do guns per se kill people? But this was never quite the question for either side; thus the problem is muddled. And this all despite the fact that there is no observational study or experiment that has yet proven outright that guns do or don’t kill people. It is not a problem particularly amenable to observational studies. The sample sizes are never large enough and we can never control for enough lurking variables.

Alongside the power that experimental and observational statistical studies have to shape common sense as a kind of rhetorical tool, information technologies now allow for new applications of these statistical tools, which shape us in a perhaps more subterranean manner. While a statistical study will affect our idea of a thing, an algorithm will set the terms for our modes of interaction with others. That is, the algorithms that mega-aggregators like Facebook and Google use to sort content for their end users are essentially complex statistical tools. Consider Reddit—one of the few platforms where we know the formula it uses to rank “best” comments. The single input variable is a boolean: upvotes for success; downvotes for failure. Then something called a binomial confidence interval is used to converge on the likelihood of the next vote being an upvote. The higher this likelihood, the higher the rank of the comment.

By contrast, Facebook and Google gather far more variables, with the apparently good intention of further optimizing how people interact. In the Reddit case, many micro-conversions (as we call them in the industry) go unaccounted for. What about when someone looks at a comment and neither upvotes nor downvotes it? What about when someone responds to a comment? What about when someone starts writing a response to a comment but stops early? What about when the language used in that aborted response was extremely negative?  Especially in a so-called closed garden like Facebook, these variables are a dime a dozen, and many of them are gathered.

But the problem with chasing these chimerical perfect algorithms is that the sample size necessary to judge variables’ effect grows (at least) exponentially relative to the number of variables, and the ability to infer causation becomes nigh impossible. Imagine that each variable is like a dimension in space. Going from two variables to three is like going from 2D space to 3D. And so on. You can intuit how much more difficult it is to actually gauge trajectories in these multidimensional spaces. Thus, in place of conscious interpretation, these platforms blindly optimize for the variables that seem most profitable (engagement, click-through rates, likes, etc.), remaining essentially agnostic about values.

But values do matter. A statistic is, by definition, a sample of a theoretically complete population. It stands to reason that when aggregators like Facebook use statistical methods to decide what content the user consumes, there is an “ideal” population lurking somewhere in the background. What does this population look like? We can easily guess, since these aggregators have already gone a long way to making it a reality. It likely looks like me in the morning—supine—groggily thumbing through tweets, occasionally giggling or angrily responding to someone I disagree with. When statistics are not used merely to observe the world (already an implicit intervention) but to shape the world, this shadow population must be explicit! We must be able to imagine the kind of world that we are seeking to create with these tools and, perhaps more importantly, to judge between worlds.

There are no perfect statistical methods, as there are no perfect systems of collective decision making. Whatever is expressed in the system is expressed at the expense of something else. But if we are to work towards “better” tools, we would do well to not only work from the bottom up, but from the top down. Instead of aimless optimization, we need at least a provisional teleology, what Gayatri Spivak might call “premature essentialization.” We must be mindful that whenever we use a statistical tool, we are making a judgment about what kind of world we want to live in.