Decisions and the Brain
Summary
The Decision-Making System
Summary
The Brain with a Mind of its Own
Summary
The Human Condition
Summary
What Is a Decision?
In order to be able to scientifically measure decision-making, we define decisions as "taking an action." There are multiple decision-making systems within each of us. The actions we take are a consequence of the interactions of those systems. Our irrationality occurs when those multiple systems disagree with each other.
The Tale of the Thermostat
Your brain is a decision-making machine, a complex but physical thing. Like any physical process, there are multiple ways in which decisions can go wrong. Being a physical being does not diminish who you are, but it can explain some of the irrational choices you make.
The Definition of Value
Value measures how much one is willing to pay, trade, or work for a reward, or to work to avoid a punishment. Value is not intrinsic to an object, but must be calculated anew each time. This produces inconsistencies, in which different ways of measuring value can produce different orderings of one's preferences.
Value, Euphoria, and Do-It-Again Signal
Pleasure and value are different things. One can have pleasure without value and value without pleasure. They are signaled by different chemicals in the brain. In particular, the brain contains a signal for "better than expected," which can be used to learn to predict value. The lack of expected reward and the lack of expected punishment are separate processes (disappointment, relief) that require additional neural mechanisms.
Risk and Reward
Defining decision-making in terms of action-selection creates the potential for decisions that are illogical from a value-based perspective. The mammalian decision-making components, which can produce conflict between action options.
Multiple Decision-Making Systems
There are multiple neural systems that drive decision-making, each of which has different computational properties that make it better suited to drive action-selection under different conditions. We will identify four action-selection systems: Reflexes, Pavlovian, Deliberative, and Procedural. Along with motoric, perceptual, situation-categorization, and motivational support systems, these make up the decision-making system.
Reflexes
A reflex is a decision-making system because it includes taking an action. A reflex entails a simple response to a stimulus based on a rule learned over long, evolutionary timescales. Although reflexes can change within a lifespan, those changes are either predetermined genetically (as animals change from children to adults) or entail only very simple learning processes (limited primarily to habituation and adaptation effects).
Emotion and the Pavlovian Action-Selection System
The Pavlovian action-selection system consists of response sets evolved to prepare for critical species-specific behaviors such as feeding, fighting, or mating. Learning in the Pavlovian action-selection system entails recognition that one thing likely causes another, leading to taking the response set appropriate for that expected outcome. Linguistically, we categorize these response sets and identify them as emotions. In humans, many of these response sets are important to social interactions.
Deliberation
Some decisions are made by explicitly imagining the potential consequences of that decision. This imagination process is particularly good for decisions that cannot be learned slowly over time, or for major decisions that require a large upfront commitment cost. Deliberation is a computationally expensive process that requires a number of neural systems, including attention, imagination, and evaluation processes, and as such takes time to calculate which action to take.
The Habits of Our Lives
With enough experience, action-selection becomes "cached" or stored in stimulus-response pairs and response-action chains. The advantage of this system is that it can respond quickly and reliably, with the same action-sequence every time. The disadvantage of this system is that it is inflexible and is slow to change in response to changing conditions.
Integrating Information
Although our senses receive physical information from the world, we interpret those senses in terms of the objects in the world—we see a tree, we feel the wind, we hear a bird. Interpreting that information requires an integration process that takes time. We can trade off time for accuracy if need be.
The Stories We Tell
Both deliberative and procedural decision-making require recognition of the “situation” we are in. Situation recognition is a categorization process in which we recognize that this moment shares similarities with previous situations we have experienced and differs from others. Although models exist of static situation-recognition processes, in actuality, situation recognition must be about recognizing sequences of events (narratives). The neural mechanisms underlying narrative remain unknown.
Motivation
The specific goals that the decision-making systems are trying to satisfy arise from intrinsic goal functions evolutionarily prewired into our brains. However, we can also learn to identify new goals that we have learned lead to those intrinsic rewards or avoid those intrinsic punishments. The vigor with which we attempt to attain those goals depends on the level of intrinsic need and the opportunity cost of chasing those goals at the expense of others.
The Tradeoff Between Exploration and Exploitation
To take the correct actions within an environment, one has to know the structure of the environment, the consequences of those actions. Exploration is the process through which a novel environment becomes familiar. Exploitation is the process of using that knowledge to get the most out of a familiar environment. There is a tradeoff between these two processes—exploration can be dangerous, but not knowing if there are better options available can be costly.
Self-Control
We have the conscious experience that we can sometimes “override” emotional or habitual responses. This means that some process must mediate between the multiple decision-making systems. The mechanisms of that “override” process are still being studied, but some neural components, such as the prefrontal cortex, are known to be involved.
The Physical Mind
The mind is physically instantiated by the brain, but the mind is not simply software running on the hardware of the brain. The physical structure of the brain changes the computations that are available. We saw earlier that there are multiple action-selection systems, each with different advantages and disadvantages. Although it is possible to learn new tasks after brain lesions, it is not that the same software is running on another part of the brain; instead, other action-selection systems are doing their best to accomplish tasks.
Imagination
If the brain is the physical instantiation of the mind, then every mental event must be reflected in the physical realm, from imagination to hallucination to mental imagery. The current theory in neuroscience is that this physical realm is the firing of neurons— that is, just as we know that the firing of sensory neurons responds to sensory cues, and the firing of motor neurons controls muscles, this theory predicts that nonsensory mental (cognitive) events are not just reflected in but actually are the firing of neurons.
Addiction
Addiction is a disorder of decision-making. Defined this way, the question of addiction becomes one of identifying the different vulnerabilities or failure modes that have arisen within an individual’s decision-making systems. This hypothesis suggests that addiction treatments need to be tailored to the individual, based on the identification of those failure modes. Treatment processes are beginning to be tailored in just such a way.
Gambling and Behavioral Addictions
From the perspective of addiction as dysfunctional decision-making, gambling and other behaviors can also be seen as “addictions.” These nonpharmacological addictions tend to depend on vulnerabilities in the support systems (e.g., motivational systems, situation-recognition systems) more than in the action-selection systems themselves, but they do also depend, of course, on how action-selection systems use those support systems.
Post-Traumatic Stress Disorder
In PTSD, devastating memories interact with vulnerable neural structures to produce an overwhelming emotional, episodic memory of an experience. The multiple decision-making system theory and its interacting support structures suggest a perspective on PTSD as interacting episodic and emotional memories.
Computational Psychiatry
Current psychiatric categorizations of patients are based on a taxonomy of symptoms. While this categorization seems to be reliable (in that different practitioners tend to categorize similar patients in similar ways), there seems to be a disconnect between the categorization and the underlying mechanisms. The detailed understanding of decision-making systems presented in Chapters 6–15 and the engineer’s view on the brain looking for failure modes, vulnerabilities, and dysfunction (Chapters 18–20) provide a new view on psychiatry, which is now emerging as a field called “computational psychiatry.”
What Makes Us Human?
Humans are animals with unique abilities, particularly in the realms of language, imagination, and culture. However, like our physical bodies, our mental abilities contain echoes of our related mammalian cousins. Like our decision-making systems, our tremendous abilities have precursors in other animal behaviors.
The Science of Morality
In much of current discussion, morality is said to be outside of the purview of science because science is descriptive, not prescriptive. Simply said, this means that science cannot tell you what to do. Although there is certainly a clear and important difference between description and prescription, the implications of the science of decision-making for morality are more complex than they would seem at first glance.
The Conundrum of Robotics
Free will includes three different issues wrapped up in one philosophical term: (1) the outside perspective (Why do some other beings appear as agents with free will to us?), (2) the societal perspective (Whom should we blame for criminal actions?), and (3) the internal perspective (What is the internal sensation that we identify as “consciousness”?). By separating free will into these three perspectives, I will show that the first two are not truly issues of “free will,” and that there is no reason to think that physical machines cannot experience an internal sensation of consciousness.
Because we like to think of ourselves as rational creatures, we like to define decision as the conscious deliberation over multiple choices. But this presumes a mechanism that might or might not be correct. If we are such rational creatures, why do we make such irrational decisions? Whether it be eating that last French fry that’s just one more than we wanted or saying something we shouldn’t have at the faculty meeting or the things we did that time we were drunk in college, we have all said and done things that we regret. Many a novel is based on a character having to overcome an irrational fear. Many an alcoholic has sworn not to drink, only to be found a few days later, in the bar, drink in hand.
We will see later in this book that the decisions that we make arise from an interaction of multiple decision-making systems. We love because we have emotional reactions borne of intrinsic social needs and evolutionary drives (the Pavlovian system, Chapter 8). We decide what job to take through consideration of the pros and cons of the imagined possibilities (episodic future thinking and the Deliberative system, Chapter 9). We can ride a bike because we have trained up our Procedural learning system (Chapter 10), but it can also be hard to break bad habits that we’ve learned too well (Chapter 15). We will see that these are only a few of the separable systems that we can identify. All of these different decision-making systems make up the person you are.
The idea that our actions arise from multiple, separably identifiable components has a long history in psychology, going back to Freud, or earlier, and has gained recent traction with theories of distributed computing, evolutionary psychology, and behavioral economics. 1 Obviously, in the end, there is a single being that takes an action, but sometimes it’s helpful to understand that being in terms of its subsystems. The analogy that I like, which I will use several times in this book, is that of a car. The car has a drive train, a steering system, brakes. Cars often have multiple systems to accomplish the same goal (the foot-pedal brake and the emergency/ parking brake, or the electric and gasoline engines in a hybrid like the Toyota Prius).
The psychologist Jonathan Haidt likes to talk of a rider and an elephant as a metaphor for the conscious and unconscious minds both making decisions, 2 but I find this analogy unsuitable because it separates the “self” from the “other” decision-making systems. As is recognized by Haidt at the end of his book, you are both the rider and the elephant. When a football star talks about being “in the zone,” he’s not talking about being out of his body and letting some other being take over— he feels that he is making the right decisions. (In fact, he’s noticing that his procedural/ habit system is working perfectly and is making the right decisions quickly and easily.) When you are walking through a dark forest and you start to get afraid and you jump at the slightest sound, that’s not some animal reacting— that’s you. (It’s a classic example of the Pavlovian action-selection system.) Sometimes these systems work together. Sometimes they work at cross purposes. In either case, they are all still you. “Do I contradict myself?” asks Walt Whitman in his long poem Song of Myself. “Then I contradict myself. I am large. I contain multitudes.”
So how do we determine how these multiple systems work? How do we determine when they are working together and when they are at cross-purposes? How do we identify the mechanisms of Whitman’s multitudes?
To study something scientifically, we need to define our question in a way that we can measure and quantify it. Thus, we need a measure of decision-making that we can observe, so we can compare the predictions that arise from our hypotheses with actual data. This is the key to the scientific process: there must be a comparison to reality. If the hypothesis doesn’t fit that reality, we must reject the hypothesis, no matter how much we like it.
One option is to simply to ask people what they want. But, of course, the idea that we always do what we say we want makes very strong assumptions about how we make decisions, and anyone who has regretted a decision knows that we don’t always decide to do what we want. Some readers may take issue with this statement, saying that you wanted that decision when you took the action. Just because you regret that night of binge drinking when you have a hangover in the morning doesn’t mean you didn’t want all those drinks the night before. Other readers may argue that a part of you wanted that decision, even if another part didn’t. We will come to a conclusion very similar to this, that there are multiple decision-making modules, and that the members of Whitman’s multitudes do not always agree with each other. Much of this book will be about determining who those multitudes are and what happens when they disagree with each other.
As a first step in this identification of the multitudes, careful scientific studies have revealed that a lot of conscious “decisions” that we think we make are actually rationalizations after the fact. 3 For example, the time at which we think we decided to start an action is often after the action has already begun. In what are now considered classic studies of consciousness, Benjamin Libet asked people to perform an action (such as tapping a finger whenever they wanted to) while watching a dot move around a circle. 4 The people were asked to report the position of the dot when they decided to act. Meanwhile, Libet and his colleagues recorded electrical signals from the brain and the muscles. Libet found that these signals preceded the conscious decision to act by several hundred milliseconds. Both the brain and muscles work by manipulating electricity, which we can measure with appropriate technologies. With the appropriate mathematics, we can decode those signals and determine what is happening within the brain. Libet decoded when the action could be determined from signals in the motor areas of the brain and compared it to when consciousness thought the action had occurred. Libet found that the conscious decision to take an action was delayed, almost as if consciousness was perceiving the action, rather than instigating it. Several researchers have suggested that much of consciousness is a monitoring process, allowing it to keep track of things and step in if there are problems. 5
Much of the brain’s machinery is doing this sort of filling-in, of making good guesses about the world. Our eyes can focus on only a small area of our visual field at a time. Our best visual sensors are an area of our retina that has a concentration of cells tuned to color and specificity in a location called the fovea. This concentration of detectors at the fovea means that we’re much better at seeing something if we focus our eyes on it. Our vision focuses on a new area of the visual world every third of a second or so. The journeys between visual focusings are called saccades. Even while we think we are focusing our attention on a small location, our eyes are making very small shifts called microsaccades. If the visual world is aligned to our microsaccades so that it shifts when we do, A the cells adapt to the constant picture and the visual image “grays out” and vanishes. This means that most of our understanding of the visual world is made from combining and interpreting short-term visual memories. We are inferring the shape of the world, not observing it.
In a simple case familiar to many people, there’s a small location on our retina where the axons from the output cells have to pass through to form the optic nerve sending the visual signals to the rest of our brain. This leaves us with a “blind spot” that must be filled in by the retina and visual processing system.B Our brain normally fills in the “blind spot” from memories and expectations from the surrounding patterns. 7
In a similar way, we don’t always notice what actually drives our decision-making process. We rationalize it, filling in our reasons from logical perspectives. Some of my favorite examples of this come from the work of V. S. Ramachandran, 8 who has studied patients with damage to parts of the brain that represent the body. A patient who is physically unable to lift her arm denies that she has a problem and merely states that she does not want to. When her arm was made to rise by stimulating her muscles directly, she claimed that she had changed her mind and raised her arm because she wanted to, even though she had no direct control of the arm. In a wonderful story (told by Oliver Sacks in A Leg to Stand On), a patient claims that his right hand is not his own. When confronted with the fact that there are four hands on the table (two of his and the two of the doctor’s), the patient says that three of the hands belong to the doctor. “How can I have three hands?” asks the doctor. “Why not? You have three arms.” replies the patient.
In his book Surely You’re Joking, Mr. Feynman, the famous physicist Richard Feynman described being hypnotized. He wrote about how he was sure he could take the action (in this case opening his eyes) even though he had been hypnotized not to, but he decided not to in order to see what would happen. So he didn’t, which was what he had been asked to do under hypnosis. Notice that he has rationalized his decision. As Feynman himself recognized, even if he said “I could have opened my eyes,” he didn’t. So what made the decision? Was it some effect of the hypnosis on his brain or was it that he didn’t want to? How could we tell? Can we tell?
Many of the experiments we’re going to talk about in this book are drawn from animals making decisions. If we’re going to say that animals make decisions, we need to have a way of operationalizing that decision— it’s hard to ask animals what they think. There are methods that allow us to decode information represented within specific parts of the brain, which could be interpreted as a means of asking an animal what it thinks (see Appendix B). However, unlike asking humans linguistically, where one is asking the overall being what it thinks, decoding is asking a specific brain structure what it is representing. Of course, one could argue that assuming what people say is what they think assumes that humans are unified beings. As we will see as we delve deeper into how decisions are made, humans (like other mammals) are mixtures of many decision-making systems, not all of which always agree with each other.
Just as the Libet experiments suggest that parts of the brain can act without consciousness, there are representations in the brain that are unable to affect behavior. In a remarkable experiment, Pearl Chiu, Terry Lohrenz, and Read Montague found signals in both smokers’ and nonsmokers’ brains that represented not only the success of decisions made but also what they could have done if they had made a better choice. 9 This recognition of what they could have done is called a counterfactual (an imagination of what might have been) and enables enhanced learning. (It is now known that both rats and monkeys can represent counterfactual reward information as well. These signals appear to use the same brain structures that Chiu, Lohrenz, and Montague were studying, and to be the same structures involved when humans express regret. 10) Counterfactuals enhance learning by allowing one to learn from imagined possibilities. For example, by watching someone else make a mistake. Or (in the example used by Chiu, Lohrenz, and Montague) “if I had taken my money out of the stock market last week, I wouldn’t have lost all that money when it crashed.” While signals in both groups’ brains reflected this counterfactual information, only nonsmokers’ behavior took that information into account. If we want to understand how decisions are made and how they go wrong, we are going to need a way to determine not only the actions taken by the subject but also the information processing happening in his or her brain.
Certainly, most nonhuman animals don’t have language, although there may be some exceptions.C Nevertheless, it would be hard to ask rats, pigeons, or monkeys (all of which we will see making decisions later in the book) what they want linguistically. Given that it is also hard to ask humans what they really want, we will avoid this language problem altogether and operationalize making a decision as taking an action, because taking an action is an observable response.
This is very similar to what is known in behavioral economics as revealed preferences. 13 Economic theory (and the concomitant new field of neuroeconomics) generally assumes that those revealed preferences maintain a rational ordering such that if you prefer one thing (say chocolate ice cream) to another (say vanilla ice cream), then you will always prefer chocolate ice cream to vanilla if you are ever in the same situation. We will not make that assumption.
Similarly, we do not want to assume that a person telling you what he or she wants actually reflects the choices a person will make when faced with the actual decision. 14 Given the data that our conscious observations of the world are inferred and the data that our spoken explanations are rationalizations, 15 some researchers have suggested that our linguistic explanations of our desires are better thought of as the speech of a “press secretary” than the actions of an executive. 16 Thus, rather than asking what someone wants, we should measure decisions by giving people explicit choices and asking them to actually choose. We will encounter some of the strangenesses discovered by these experiments in subsequent chapters.
Often, experimentalists will offer people hypothetical choices. It is particularly difficult to get funding to provide people a real choice between $ 10,000 and $ 100,000, or to give them a real choice whether or not to kill one person to save five. Instead, subjects are asked to imagine a situation and pretend it was real. In practice, in the few cases where they have been directly compared, hypothetical and real decision-making choices tend to match closely. 17 But there are some examples where hypothetical and real decisions diverge. 18 These tend to be with rewards or punishments that are sensory, immediate, and what we will recognize later as Pavlovian (Chapter 8).
A classic example of this I call the parable of the jellybeans. Sarah Boysen and Gary Berntson tried to train chimpanzees to choose the humble portion. 19 They offered the subject two trays of jellybeans. If he reached for the larger tray, he got the smaller one and the larger one was given to another chimpanzee; however, if the deciding chimpanzee reached for the smaller tray, he got the larger tray and the smaller one was given to another chimpanzee. When presented with symbols (Arabic numerals that they had previously been trained to associate with numbers of jellybeans), subjects were able to choose the smaller group, but when the choices were physical jellybeans, they were unable to prevent themselves from reaching for the larger group of jellybeans. Other experiments have found that more linguistically capable animals are more able to perform these self-control behaviors. 20 This may be akin to our ability to talk ourselves out of doing things that we feel we really want: “I know I’m craving that cigarette. But I don’t want it. I really don’t.”
A similar experiment is known colloquially as the marshmallow experiment. 21 Put a single marshmallow in front of a child sitting down at the kitchen table. Tell the child that if the marshmallow is still sitting there in five minutes, you’ll add a second marshmallow to it. Then leave the room. It is very, very difficult for children not to reach for the marshmallow. It is much easier for children to wait for two pennies or two tokens than for two marshmallows. We will discuss the marshmallow experiment in detail in the chapter on self-control (Chapter 15). Studies of decision-making in psychology (such as the marshmallow experiment) as well as studies in behavioral economics and the new field of neuroeconomics tend to measure choices within the limitation of discrete options. In the real world, we are rarely faced with a discrete set of options. Whether it be deciding when to swing the bat to hit a baseball or deciding where to run to on a playground, we are always continuously interacting with the world. 22 As we will see later, some of the mechanisms that select the actions we take are not always deliberative, and do not always entail discrete choices.
So where does that leave us in our search for a definition of decision-making? We will not assume that all decision-making is rational. We will not assume that all decision-making is deliberative. We will not assume that decision-making requires language. Instead, we define decision-making as the selection of an action. Obviously, many of the decisions we take (such as pushing a lever or button— say on a soda machine) are actual actions. But note that even complex decisions always end in taking an action. For example, buying a house entails signing a form. Getting married entails making a physical statement (saying “I do”). We are going to be less concerned about the physical muscle movements of the action than about the selection process that decided on which action to take. Nevertheless, defining “decision-making” as “action-selection” will force us to directly observe the decisions made. It will allow us to ask why we take the actual actions we do. Why don’t those actions always match our stated intentions? How do we choose those actions over other potential actions? What are the systems that select actions, and how do those systems interact with the world? How do they break down? What are their vulnerabilities and failure-modes? That is what the rest of this book is about.
The decision-making that you do arises from the physical processes that occur in your brain. Because your brain is a physical thing, it has vulnerabilities, what one might call “failure-modes” in the engineering world. We see these vulnerabilities in our susceptibility to bad choices (Do you really need a new sports car?), in our susceptibility to addictions (Why can’t you just quit smoking those cigarettes?), in our inability to control our emotions or our habits (Why do you get so angry about things you can’t change?). We would like to believe that we are rational creatures, capable of logic, always choosing what’s best for us. But anyone who has observed their own decisions (or those of their friends) will recognize that this is simply not correct. We are very complex decision-making machines, and sometimes those decision-making processes perplex us. Understanding how the brain makes decisions will help us understand ourselves. To understand those vulnerabilities, we need to understand the mechanism of decision-making in our brains.
Today, we are all familiar with complex machines, even complex machines that make decisions. The simplest machine that makes a decision is the thermostat—when the house is too hot, the thermostat turns on the air conditioning to cool it down, and when the house is too cold, the thermostat turns on the furnace to heat it up. This process is called negative feedback—the thermostat’s actions are inversely related to the difference between the temperature of the room and the temperature you’d like it to be (the set-point). But is the process really so simple? Taking the decision-making process of the thermostat apart suggests that even the simple decision-making process of the thermostat is not so simple.
The thermostat has three key components of decision-making that we will come back to again and again in this book. First, it perceives the world—the thermostat has a sensor that detects the temperature. Second, it determines what needs to be done—it compares that sensor to the set-point and needs to increase the temperature because it is too cold, needs to decrease the temperature because it is too hot, or doesn’t need to do anything because the temperature is just right. Finally, it takes an action—it turns on either the furnace or the air conditioning.
In the artificial intelligence literature, there was an argument through much of the 1980s about whether a thermostat could have a “belief.” 1 Fundamentally, a belief is a (potentially incorrect) representation about the world. Clearly, a working thermostat requires a representation of the target temperature in order to take actions reflecting the temperature of the outside world. But can we really say that the thermostat “recognizes” the temperature of the outside world? The key to answering this question is that the thermostat does not take actions based on the temperature of the world, but rather on its internal representation of the temperature of the world. Notice that the internal representation might differ from the real temperature of the room. If the sensor is wrong, the thermostat could believe that the room is warmer or cooler than it really is and take the wrong action.
One of the key points in this book is that knowing how the brain works allows us a better perception of what happens when something goes wrong. I live in Minnesota. In the middle of winter, it can get very cold outside. Imagine you wake up one morning to find your bedroom is cold. Something is wrong. But simply saying that the thermostat is broken won’t help you fix the problem. We need to identify the problem, to diagnose it, if you will.
Maybe something is wrong with the thermostat’s perception. Perhaps the sensor is broken and is perceiving the wrong temperature. In this case, the thermostat could think that the house is fine even though it is too cold. (Notice the importance of belief here— the difference between the thermostat’s internal representation and the actual temperature can have a big impact on how well the thermostat makes its decision!) Maybe the set-point is set to the wrong temperature. This means that the thermostat is working properly— it has correctly moved the temperature of your house to the set-point, but that’s not what you wanted. Or, maybe there’s something wrong with the actions available to the thermostat. If the furnace is broken, the thermostat may be sending the signal saying “heat the house” but the house would not be heating correctly. Each of these problems requires a different solution. Just as there are many potential reasons why your bedroom is too cold and knowing how a thermostat works is critical to understanding how to fix it, when smokers say that they really want to quit smoking, but can’t, we need to know where each individual’s decision-making process has gone wrong or we won’t be able to help. Before we can identify where the decision-making process has broken down, we’re going to need to understand how the different parts of the brain work together to make decisions. Many readers will object at this point that people are much more complicated than thermostats. (And we are.)
Many readers will then conclude that people are not machines. Back when negative feedback like the thermostat was the most complicated machine mechanism that made decisions, it was easy to dismiss negative feedback as too simple a model for understanding people. However, as we will see later in the book, we now know of much more complicated mechanisms that can make decisions. Are these more complicated mechanisms capable of explaining human decision-making? (I will argue that they are.) This leaves open some difficult questions: Can we be machines and still be conscious? Can we be machines and still be human?
The concept of conscious machines making decisions pervades modern science fiction, including the droids C3P0 and R2D2 in Star Wars, the android Data of Star Trek: The Next Generation, the desperate Replicants of Ridley Scott’s Blade Runner, and the emotionally troubled Cylons of Battlestar Galactica. Star Trek: The Next Generation spent an entire episode (The Measure of a Man) on the question of how the fact that Data was a machine affected his ability to decide for himself whether or not to allow himself to be disassembled. In the episode, the judge concludes the trial with a speech that directly addresses this question—“ Is Data a machine? Yes. Is he the property of Starfleet? No. We have all been dancing around the basic issue: does Data have a soul? I don’t know that he has. I don’t know that I have. But I have got to give him the freedom to explore that question himself. It is the ruling of this court that Lieutenant Commander Data has the freedom to choose.” 2 We will examine the complex questions of self and consciousness in detail at the end of the book (The Conundrum of Robotics, Chapter 24), after we have discussed the mechanisms of decision-making. In the interim, I aim to convince you that we can understand the mechanisms of our decision-making process without losing the freedom to choose.
I don’t want you to take the actual process of the thermostat as the key to the story here, anymore than we would take jet planes as good models of how swans fly. And yet, both planes and swans fly through physical forces generated by the flow of air over their specially shaped wings. Even though bird wings are bone, muscle, and feathers, while airplane wings are metal, for both birds and planes, lift is generated by airflow over the wings, and airflow is generated by speed through the air. If we can understand what enables a 30-ton airplane to fly, we will have a better understanding of how a 30-pound swan can fly. Even though the forward push through the air is generated differently, both birds and planes fly through physical interactions with the air. We will use analogous methods of understanding decision-making processes to identify and fix problems in thermostats and in ourselves, because both use identifiable computational decision-making processes.
A good way to identify where a system (like a thermostat) has broken down is a process called “differential diagnosis,” as in What are the questions that will differentiate the possible diagnoses? My favorite example of this is the show CarTalk on National Public Radio, in which a pair of MIT-trained auto mechanics (Tom and Ray Magliozzi) diagnose car troubles. When a caller calls in with a problem, the first things they discuss are the basics of the problem. (Anyone who has actually listened to CarTalk will know that the first things Tom and Ray discuss are the person’s name, where the caller is from, and some completely unrelated jokes. But once they get down to discussing cars, they follow a very clear process of differential diagnosis.) A typical call might start with the caller providing a description of the problem—“ I hear a nasty sound when I’m driving.” And then Tom and Ray will get down to business— they’ll ask questions about the sound: “What is the sound? Where is it coming from? Does it get faster as you go faster? Does it still happen when the engine is on but the car is not moving?” Each question limits the set of possible problems. By asking the right series of questions, one can progressively work one’s way to identifying what’s wrong with the car. If we could organize these questions into a series of rules, then we could write a computer program to solve our car problems. (Of course, then we wouldn’t get to hear all the CarTalk jokes. Whether this is good or bad is a question of taste.)
In the 1980s, the field of artificial intelligence developed “expert systems” that codified how to arrange these sorts of question-and-answer rules to perform differential diagnosis. 3 At the time, expert systems were hailed as the answer to intelligence— they could make decisions as well as (or better than) experts. But it turns out that most humans don’t make decisions using differential diagnoses. In a lot of fields (including, for example, medicine), a large part of the training entails trying to teach people to make decisions by these highly rational rules. 4 However, just because it is hard for people to make decisions by rule-based differential diagnoses does not mean that humans don’t have a mechanism for making decisions. In fact, critics complained that the expert systems developed by artificial intelligence were not getting at the real question of what it means to be “intelligent” long before it was known that humans didn’t work this way. 5 People felt that we understood how expert systems work and thus they could not be intelligent. A classmate in college once said to me that “we would never develop an artificial intelligence. Instead, we will recognize that humans are not intelligent.” One goal of this book is to argue that we can recognize the mechanisms of human decision-making without losing our sense of wonder at the marvel that is human intelligence.
Some readers will complain that people are not machines; they have goals, they have plans, they have personalities. Because we are social creatures and much of our intelligence is dedicated to understanding each other, we have a tendency to attribute agency to any object that behaves in a complex manner. 6 Many of my friends name their cars and talk about the personality of their cars. My son named our new GPS navigation system “Dot.” When asked why he named the GPS (the voice is definitely a woman’s), he said, “So we can complain to her when she gets lost—‘ Darn you, Dot!’”
A GPS navigator has goals. (These are goals we’ve programmed in, but they are goals nonetheless.) Dot’s internal computer uses her knowledge of maps and her calculation of the current location to make plans to achieve those goals. You can even tell Dot whether you prefer the plans to include more highways or more back-country scenic roads. If Dot were prewired to prefer highways or back-country scenic roads, we would say she has a clear personality. In fact, I wish I had some of Dot’s personality— when we miss an exit, Dot doesn’t complain or curse, she just says “recalculating” and plans a new route to her goal.
The claim that computers can have goals and that differences in how they reach those goals reflects their personality suggests that goals and plans are simple to construct and that personality is simply a difference in underlying behavioral variables and preferences. We explain complex machines by thinking they’re like people. Is it fair to turn that on its head and explain people as complex machines?
In this book, I’m going to argue that the answer to this question is “yes”—your brain is a decision-making machine, albeit a very complex one. You are that decision-making machine. This doesn’t mean you’re not conscious. This doesn’t mean you’re not you. But it can explain some of the irrational things that you do.
Understanding how the human decision-making system works has enormous implications for understanding who we are, what we do, and why we do what we do. Scientists study brains, they study decision-making, and they study machines. By bringing these three things together, we will begin to get a sense of ourselves. In this book, I will discuss what we know about how brains work, what we know about how we make decisions, and what we know about how that decision-making machine can break down under certain conditions to explain irrationality, impulsivity, and even addiction.
The definition of value
Summary
Illogical value calculations
Summary
Measuring value
Summary
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
Content
The most interesting thing about these results (of which there are many) is that even when they are illogical (like the disease example above), they often still sound right. There are other examples where we can see that decisions are irrational, and don’t sound right when they are pointed out to us, but which humans definitely show when tested. The best examples of most of these come from the advertising and marketing worlds.
A person goes into an electronics store to buy a television and finds three televisions on sale, one for $100, one for $200, and one for $300. Imagine that the three televisions have different features, such that these are reasonable prices for these three TVs. People are much more likely to buy the $200 TV than either of the other two. If, in contrast, the person goes into the store and finds three televisions on sale, the same $200 TV, the same $300 TV, and now a fancier $400 TV, the person is more likely to buy the $300 TV. In the first case, the person is saying that the $200 TV is a better deal than the $300 one, but in the second case, the person is saying that the $300 TV is the better deal. Even though they’ve been offered the same televisions for the same prices, the decision changed depending on whether there is a $100 TV or a $400 TV in the mix. This is called extremeness aversion and is a component of a more general process called framing. In short, the set of available options changes your valuation of the options. This is completely irrational and (unlike the other examples above) seems unreasonable (at least to me). Yet it is one of the most reliable results in marketing and has probably been used since the first markets in ancient times through to the modern digital age. 10
Products in advertisements used to compare themselves to each other. Tylenol would say it was better than aspirin and Coke would say it was better than Pepsi. I remember an RC Cola commercial with two opposing teams fighting about which drink they preferred, Coke or Pepsi, while a third person sits on the sidelines drinking an RC Cola, out of the fray, smiling. Advertisements today rarely mention the competition. This is because one of the heuristics we use is simply whether we recognize the name or not. 11 So although RC Cola was trying to communicate that Coke and Pepsi were the same, while RC Cola was different, what people took from the advertisement was a reminder that everyone drank Coke and Pepsi. Just mentioning the name reminds people that it exists and reinforces the decision to choose it. Familiarity with a brand name increases the likelihood that one will select it.
It is election season as I write this. All around my neighborhood, people have put out signs for their candidates. The signs don’t say anything about the candidates. Often they don’t even have the party the candidates belong to. Usually, it’s just the name of the candidate, and sometimes, an appeal to “vote for” the candidate. What information do I get from seeing a sign that says nothing other than “Vote for X”? I suppose that one may need to be reminded to vote at all, but then why does the sign include “for X” on it? (One of my neighbors does have a large handmade sign she puts out every election season that just says “VOTE!” on it.) It is true that one can get a sense of the grouping of candidates from people who put out multiple signs: given that I like person X for state representative, seeing that all the people with person X also have person Y for county prosecutor, while all the people who have person X’s opponent have person Z for county prosecutor, might suggest to me that I would like person Y over person Z for county prosecutor. But lots of people have just one sign out. All that tells me is that lots of people like person X. Why would knowing that lots of people like something suggest that I would too?
There are actually three things that these single-sign houses are doing. First, they are increasing my familiarity with that name. As with the products, just knowing the name increases one’s likelihood of voting for someone. Second, if lots of people like a movie, it’s more likely to be a good movie than one everybody hated. Using the same heuristic, if everyone likes a candidate, isn’t that candidate more likely to be a good choice? And third, we like to be on the winning team. If everyone else is going to vote for someone, he or she is likely to win. Of course, we’re supposed to be voting based on who we think is a better choice to govern, not who is most popular. But it’s pretty clear that a lot of people don’t vote that way.
These effects occur because we don’t actually calculate the true value of things. Instead, we use rules of thumb, called heuristics, that allow us to make pretty good guesses at how we value things. 12 If you like all the choices, picking the middle one is a pretty good guess at good value for your money. If you’re making a decision, familiarity is a pretty good guess. (Something you’re familiar with is likely to be something you’ve seen before. If you remember it, but don’t remember it being bad, how bad could it have been?)
Some economists have argued that evolutionarily, heuristics are better than actually trying to calculate the true value of things because calculating value takes time and heuristics are good enough to get us by. 13 But a lot of these nonoptimal decisions that we’re making are taking time. Knowing that Programs A and C are the same and that programs B and D are the same in the Kahneman and Tversky flu example above doesn’t change our minds. This makes me suspect that something else is going on. It’s not that heuristics are faster and we are making do with “good enough.” I suspect that these effects have to do with how we calculate value. We cannot determine how humans calculate value unless we can measure it. So, again, we come back to the question of how we measure value.
What is “value”? What does it mean to value something? In psychological theories, value is thought of as that which reduces needs or which alleviates negative prospects1—when you are hungry, food has value; when you are thirsty, water has value. Pain signals a negative situation that must be alleviated for survival. In economics, the concept of revealed preferences says that we value the things we choose, and make decisions to maximize value. 2 Although this is a circular definition, we will use it later to help us understand how the decision-making process operates. In the robotics and computer simulation worlds where many of the mechanisms that we will look at later have been worked out, value is simply a number r for reward or -r for punishment. 3 The robotics and computer science models simply try to maximize r (or minimize -r).
In many experiments that study decision-making, value is measured in terms of money. But of course, money has true value only in terms of what it can buy. 4 Economists will often tell us that money is an agreed-upon fiction— I am willing to sell you this apple for a green piece of paper that we both agree is worth $1 because I am confident that I can then use the paper later to buy something that I want. The statement that money is valuable only in terms of what it can buy is not entirely true either: money can also have value as a symbol of what it implies for our place in the social network. 5 This is one reason why extremely well-paid professional athletes fight for an extra dollar on their multimillion dollar salaries— they want to be known as the “best-paid wide receiver” because that identification carries social value. Money and value are complex concepts that interact in difficult ways.
Part of the problem is that you cannot just ask people How much do you value this thing? We don’t have units to measure it in. Instead, value is usually measured in what people will trade for something and in the decisions they make. 6 This is the concept called revealed preferences—by examining your decisions, we can decide what you value. Theoretically, this is very simple: one asks How much will you pay for this coffee mug? The price you’ll pay is the value of the coffee mug. If we measure things at the same time, in the same experiment, we shouldn’t have to worry about inflation or changes in the value of the money itself.
The problem is that people are inconsistent in their decisions. In a famous experiment, Daniel Kahneman and his colleagues Jack Knetsch and Richard Thaler divided their subjects into two groups: one group was asked how much money they would pay for a $6 (in 1990) Cornell University coffee mug and the other group was first given the coffee mug and then asked how much money they would accept for the mug. The first group was willing to pay much less on average ($3) than the second group was willing to accept for it ($7). This is called the endowment effect and is a preference for things you already own. 7
Simply phrasing decisions in terms of wins or losses changes what people choose. 8 The interesting thing about this is that even once one is shown the irrationality, it still feels right. This effect was first found by Daniel Kahneman and Amos Tversky, who made this discovery sitting in a coffee shop in Israel, asking themselves what they would do in certain situations. When they found themselves making irrational decisions, they took their questions to students in their college classes and measured, quantitatively, what proportion made each decision in each condition.
In the classic case from their 1981 paper published in the journal Science: 9
Imagine that the U.S. is preparing for the outbreak of an unusual Asian disease, which is expected to kill 600 people. Two alternative programs to combat the disease have been proposed. Assume that the exact scientific estimate of the consequences of the programs are as follows:
Problem 1: If Program A is adopted, 200 people will be saved. If Program B is adopted, there is 1/3 probability that 600 people will be saved, and 2/3 probability that no people will be saved. Which of the two programs would you favor?
Problem 2: If Program C is adopted, 400 people will die. If Program D is adopted, there is 1/3 probability that nobody will die, and 2/3 probability that 600 people will die. Which of the two programs would you favor?
A careful reading of the two conditions shows that Program C is identical to Program A, while Program D is identical to Program B. Yet, 72% of the first group chose program A and 28% chose program B, while only 22% of the second group chose program C and 78% chose program D! Kahneman and Tversky interpret this as implying that we are more sensitive to losses than to gains— we would rather risk not gaining something than we would risk losing something. This is clearly a part of the story, but I will argue that there is a deeper issue here. I will argue that “value” is something that we calculate each time, and that the calculation process doesn’t always come up with the same “rational” answer.
With animals, we can’t ask them how much they value something; we can only offer them something and determine how much they’ll pay for it. Usually, this is measured in terms of the amount of effort an animal will expend to get the reward. 14 How many lever presses is the animal willing to make for each reward?
Alternatively, we can give the animal a direct choice between two options: 15 Would the animal rather have two apple-flavored food pellets or one orange-flavored food pellet? Would it rather have juice or water? We can also titrate how much of one choice needs to be offered to make the animal switch— if the animal likes apple flavor better than orange flavor, would it still prefer half as much apple to the same amount of orange?
Finally, we can also measure the negative consequences an animal will put up with to get to a reward. 16 Usually, this is measured by putting a negative effect (a small shock or a hot plate) in between the animal and the reward, which it has to cross to get to the reward. (It is important to note that crossing the shock or hot plate is entirely up to the animal. It can choose not to cross the punishment if it feels the reward is not valuable enough.) A common experiment is to balance a less-appealing reward with a more-appealing reward that is given only after a delay. 17 Because animals are usually hungry (or thirsty) when running these experiments, they don’t want to wait for reward. Thus delay becomes a very simple and measurable punishment to use—how long will the animal wait for the reward?
Of course, humans are animals as well and all of these experiments work well in humans: How much effort would a smoker spend to get a puff of a cigarette? 18 How much money will you trade for that coffee mug? 19 What negative consequences will you put up with to achieve your reward? 20 How long will you wait for that larger reward? 21 Will you wait for a lesser punishment, knowing that it’s coming? 22
These four ways of measuring value all come down to quantifiable observations, which makes them experimentally (and scientifically) viable. Economists, who study humans, typically use hypothetical choices (“ What would you do if …?”) rather than real circumstances, which, as we will see later, may not access the same decision-making systems. It has been hard to test animals with hypothetical choices (because hypothetical choices are difficult to construct without language), but there is some evidence that chimpanzees with linguistic training will wait longer for rewards described by symbols than for immediately presented real rewards. 23 (Remember the chimpanzees and the jellybeans?) This, again, suggests that language changes the decision-making machinery. More general tests of hypothetical rewards in animals have been limited by our ability to train animals to work for tokens. Recent experiments by Daeyeol Lee and his colleagues getting monkeys to work for tokens may open up possibilities, but the critical experiments have not yet been done. 24
One of the most common ways to measure willingness to pay for something is a procedure called the progressive ratio—the subject (whether it be human or not) has to press a lever for a reward, and each time it receives the reward, the number of times it has to press the lever for reward increases, often exponentially. The first reward costs one press, the second two, the third four, the fourth eight, and so on. Pretty soon, the subject has to press the lever a thousand times for one more reward. Eventually, the animal decides that the reward isn’t worth that much effort, and the animal stops pressing the lever. This is called the break point. Things that seem like they would be more valuable to an animal have higher break points than things that seem like they would be less valuable. Hungry animals will work harder for food than satiated animals. 25 Cocaine-addicted animals will work harder for cocaine than for food. 26 A colleague of mine (Warren Bickel, now at Virginia Tech) told me of an experimental (human) subject in one of his experiments who pulled a little lever back and forth tens of thousands of times over the course of two hours for one puff of nicotine!
It’s been known for a long time that drugs are not infinitely valuable. When given the choice between drugs and other options, both human addicts and animals self-administering drugsA will decrease the amount of drug taken as it gets more expensive relative to the other options. 27 Drugs do show what economists call elasticity: the more expensive they get, the less people take. This is why one way to reduce smoking in a population is to increase the tax on cigarettes.
Elasticity measures how much the decision to do something decreases in response to increases in cost. Luxuries are highly elastic; nonluxuries are highly inelastic. As the costs of going to a movie or a ballgame increase, the likelihood that people will go decreases quickly. On the other hand, even if the cost of food goes up, people aren’t about to stop buying food. Some economists have argued that a good definition of addiction is that things we are addicted to are inelastic. This allows them to say that the United States is “addicted to oil” because we are so highly dependent on driving that our automobile use is generally inelastic to the price of oil. However, again, we face an interesting irrationality. The elasticity of automobile use is not linear28—raising the price of gasoline from $1 per gallon to $2 had little to no effect on the number of miles driven, but when a gallon of gasoline crossed the $3 mark, there was a sudden and dramatic drop in the number of miles driven in 2005. People suddenly said, “That’s not worth it anymore.” Of course, people then got used to seeing $3 per gallon gasoline and the number of miles driven has begun to increase again.
This irrationality is where the attempt to measure value gets interesting. In a recent set of experiments, Serge Ahmed and his colleagues found that even though the break point for cocaine was much higher than for sweetened water, B when given the choice between cocaine and sweetened water, almost all of the animals chose the sweetened water. 29 (Lest readers think this is something special about cocaine, the same effects occur when examining animals self-administering heroin. 30) Measuring value by how much the animal was willing to pay said that cocaine was more valuable than sweetened water. (Cocaine had a higher break point than the sweetened water did.) But measuring value by giving the animals a choice said that the sweetened water was more valuable. (They consistently chose the sweetened water over the cocaine.) The two measures gave completely different results as to which was more valuable.
One of my favorite examples of this irrationality is a treatment for addiction called Contingency Management, where addicts are offered vouchers to stay off drugs. 31 If addicts come into the clinic and provide a clean urine or blood sample, showing that they have remained clean for the last few days, they get a voucher for something small—a movie rental or a gift certificate. Even very small vouchers can have dramatic effects. What’s interesting about this is that raising the cost of a drug by $3 would have very little effect on the amount of drug an addict takes (because drugs are inelastic to addicts), but providing a $3 voucher option can be enough to keep an addict clean, straight, and sober. Some people have argued that this is one of the reasons that Alcoholics Anonymous and its 12-step cousins work.C It provides an option (going to meetings and getting praised for staying sober) that works like a voucher; it’s a social reward that can provide an alternative to drug-taking. 32
So why is it so hard to measure value? Why are we so irrational about value? I’m going to argue that value is hard to measure because value is not something intrinsic to an object.
First, I’m going to argue in this book that there are multiple decision-making systems, each of which has its own method of action-selection, 33 which means there are multiple values that can be assigned to any given choice. Second, in the deliberative experiments we’ve been looking at, we have to calculate the value of the available options each time. 34 (Even recognizing that options are available can be a complex process involving memory and calculation.) The value of a thing depends on your needs, desires, and expectations, as well as the situation you are in. By framing the question in different ways, we can guide people’s attention to different aspects of a question and change their valuation of it. In part, this dependence on attention is due to the fact that we don’t know what’s important. A minivan can carry the whole family, but the hybrid Prius gets better gas mileage, and that convertible BMW looks cooler. It’s hard to compare them.
This means that determining the true value of something depends on a lot of factors, only some of which relate to the thing itself. We’ll see later that value can depend on how tired you are, on your emotional state, and on how willing you are to deliberate over the available options. 35 It can even be manipulated by changing unrelated cues, such as in the anchor effect, in which unrelated numbers (like your address!) can make you more likely to converge to a higher or lower value closer to that number. 36
Economists argue that we should measure value by the reward we expect to get, taking into account the probability that we will actually get the reward, and the expected investment opportunities and risks. 37 If we tried to explore all possibilities and integrate all of this, we could sit forever mulling over possibilities. This leads us to the concept of bounded rationality, introduced by Herb Simon in the 1950s, which suggests that the calculation takes time, that sometimes it’s better to get to the answer quickly by being less complete about the full calculation, and that sometimes a pretty good job is good enough. 38 Instead, we use heuristics, little algorithms that work most of the time.
The problem with this theory is that even when we are given enough time, we continue to use these heuristics. Economists and psychologists who argue for bounded rationality (such as Gerd Gigerenzer39) argue that evolution never gives us time. But if the issue were one of time, one would expect that the longer one was given, the more rational one would be. This isn’t what is seen. People don’t change their minds about taking the middle-value television if they are given more time; in fact, they become more likely to pick the middle television, not less, with more time.40
We pick the middle television in the three-television example because one of the algorithms we use when we want to compare the choices immediately available to us is to find the middle one. Another example is that we tend to round numbers off by recognizing the digits. 41 $3.99 looks like a lot less than $4.00. Think about your “willingness to pay” $3.75 per gallon for gasoline relative to $3.50 per gallon. If both options are available, obviously, you’ll go to the station selling it for $3.50. But if it goes up from $3.50 to $3.75 from one day to the next, do you really stop buying gasoline? Now, imagine that the price goes up from $3.75 to $4.00. Suddenly, it feels like that gasoline just got too expensive. We saw a similar thing at the soda machine by my office. When the cost went up from 75¢ to $1, people kept buying sodas. But then one day it went up from $1 to $1.25, and people said, “It’s not worth it” and stopped. (It’s now $1.50 and no one ever buys sodas there anymore.) Sometimes the cost crosses a line that we simply won’t put up with anymore.
Part of this is due to the mechanisms by which we categorize things. Whole dollar amounts draw our attention and we are more likely to pick them. In a recent experiment, Kacey Ballard, Sam McClure, and their colleagues asked people which they would prefer, $7 today or $20 in a week. 42 (This is a question called delay-discounting, which we will examine in detail later in our chapter on impulsivity [Chapter 5].) But then they asked the subjects to decide between $7.03 today and $20 in a week. People were more likely to pick the $20 over $7.03 than $20 over $7 even. In a wonderful control, they then asked the subjects to decide between $7 today and $20.03 in a week. This time, people were more likely to pick the $7. There is no way that these decisions are rational, but they do make sense when we realize (1) that making the decision requires determining the value of the two options anew each time and comparing them, and (2) that we use heuristics that prefer even dollar amounts to compare them.D
Value is an important concept to understanding decision-making, but our brains have to calculate how much we value a choice. That calculation is based on heuristics and simple algorithms that work pretty well most of the time but can be irrational under certain conditions. How do we actually, physically determine value? What are the neurophysiological mechanisms? That brings us to the differences between pleasure, pain, and the do-it-again signal.