As an anthropologist of AI, I study how the technology raises questions about what it means to be human. Amid increasingly intelligent machines, which of our capabilities are seemingly no longer unique to us, and what does that say about who we really are? What even is this “we”? These kinds of questions saturate the field of AI research, and that’s one of my motives for studying it. I moved to the Bay Area to figure out how responses to these questions might come from within generative AI research and deployment. In other words, I’m observing how the AI thinks and works anthropologically.
Last weekend, I explored this kind of anthropology with a diverse mix of AI experts at an AI Salon hosted by Ian Eisenberg and Liz O’Sullivan. The meeting’s topic, proposed this time by Liz, was “Common Sense.” As the Salon officially asked:
How do we bring “common sense” into our AI systems? How universal is “common sense” anyways?
These are the kinds of questions I’ve been tracking in AI, and after two hours of rich discussion, I felt compelled to write further. The ideas I develop here owe much to the Salon participants and our conversation that afternoon.
What It Takes to Define Common Sense
Clifford Geertz, a cultural anthropologist, offers an evocative definition of common sense:
When we say someone shows common sense we mean to suggest more than that he is just using his eyes and ears, but is, as we say, keeping them open, using them judiciously, intelligently, perceptively, reflectively, or trying to, and that he is capable of coping with everyday problems in an everyday way with some effectiveness.
Common sense is the taken-for-granted knowledge that allows us to effectively live in a world shared by other people and other forces beyond our control. It is a matter of being in the world and being aware in the world. And it is, ideally, knowledge that goes without saying.
As helpful as this is (and helps helpful enough that I’ll return to it), an important piece remains missing: What specific kinds of knowledge does common sense actually include? If one were to create a database of common-sense assertions for ML training or evaluation, what should one add? Table manners? The rules of the road? The fact that rain is wet—or that one should find shelter to avoid it?[1] To demarcate some domain of knowledge as common sense in a way that holds in all cases and satisfies all interested parties is tough enough for one particular time and place. Fully defining common sense across all times and places, then, as any globally deployed AI system would require, might be impossible. Every group may have a common sense, but what that sense entails may differ across them.[2]
Common Sense is a Shifter
Linguistic anthropology offers a potential way out of this definitional challenge. Instead of asking what common sense is, ask how, why, and to what end people invoke the concept. That’s to say that maybe we should treat the slipperiness of common sense as a feature, not a bug. This means taking a pragmatic approach, one common across ethnographic research, linguistic anthropology, and the philosophical tradition of Pragmatism that inspires this work.
To follow this approach, I start from the rough observation that common sense is rarely invoked when it works correctly. As knowledge that normally goes without saying, common sense ideally remains in the background. For the most part, we only mention common sense when there’s a problem with it.[3] For example, when someone seems to lack it. Of course, we only learn this after the fact. We experience common sense in its absence—during a faux pas, for example, or worse. And we’re equipped with emotions for detecting common sense: when it’s others who lack it, there’s frustration, and when it’s us, there’s embarrassment.
But what knowledge is actually missing is not necessarily determined in advance. Instead, what common sense actually refers to depends on the actual situation. What common sense means is a function of its context—namely the contexts in which it goes wrong. It is therefore what linguists call a shifter. Shifters are words whose meaning depends on the context in which they are used. Other shifters include words like here, there, I, and her. Shifters lack a fixed referent; instead, they describe relations between subjects and objects. For shifters like pronouns and prepositions, those relations are typically descriptive. For shifters like common sense, they’re normative.
Thinking about common sense as a shifter changes the meaning of a question like “can AI systems have common sense?” The answer is most certainly yes, insofar as AI systems can behave in ways that human observers might not expect, to act in ways beyond what should have gone without saying, to elicit frustration or embarrassment. But the common sense an AI system would lack in such cases is not necessarily located in its memory; instead, it’s a function of the system’s behavior in the world. As a shifter, common sense describes a relation, and relations need multiple parties.
How ML can “Shift” Common Sense
But our relations with machines are different than relations with other humans, and this is where things get interesting. Human and machine behavior are different. Compared to humans, machines can work on different scales in space and time. Their actions can be far larger or far smaller, far quicker or last far longer. They may have sensors—and pretty good ones, indeed—but they don’t have eyes and ears that work like ours. What it is to be like a robot or other machine is so different that, as Rodney Brooks argues, they effectively live in a different world than us.
So if common sense is, returning to Geertz, a matter of using one’s eyes and ears and keeping them open, then machines must have a different relationship with common sense. What is common for us (whatever that really is) may not be common for them, and vice versa. Machines will inevitably act beyond what for humans may reasonably go without saying.
This is not necessarily new. Software is loaded with breaches of common sense, regardless of whether the software uses machine learning or deterministic logic. Developers are constantly called upon to reign in their programs—to make them “just work.” But as AI systems take on new capabilities and configurations, the ways in which common sense can be breached may increase in variety. For example, autonomous vehicles can detect the presence of rain, but their relationship with rain is different than that of their human passengers. As a result, they might pull over just short of a protective overhang, exposing their passengers to the elements when they could have stayed dry. A human driver might also lack this common sense (and good manners), but in a world run by autonomous systems, these kinds of events may become more common and pronounced.
As we interact with new kinds of technical systems, knowledge we had always taken for granted may no longer go without saying. In other words, new kinds of knowledge might come to be called common sense because we’ll discover new kinds of knowledge are not as common as we thought. As a result, trying to teach common sense to machines before they enter the world will inevitably fall short. The pressing question is not how to prepare for new kinds of common sense, but rather how to prepare for the breaches that will inevitably bring them to our attention.
[1] This last example comes from Geertz.
[2] This is why common sense is, for Geertz, a cultural system.
[3] In this sense, common sense is like infrastructure in that it’s often only visible when it breaks down, as Susan Leigh Star observes. Perhaps we could think of common sense as a kind of infrastructural knowledge.
Spencer Kaplan 