The Ghost in the Scanner: A First Encounter with the Deciding Brain

Chapter 1 — The Ghost in the Scanner: A First Encounter with the Deciding Brain

On a Tuesday in October, in a basement laboratory three floors below the streets of lower Manhattan, a twenty-four-year-old graduate student named Maya slid into the bore of a Siemens 3-Tesla scanner and waited. The bore was cold. The padding pressed against the back of her skull, and a mirror mounted on the head coil angled her gaze toward a translucent screen fixed thirty centimeters from her face. On that screen a series of offers would appear, one at a time, each one a small, indifferent question: would she prefer $14 tomorrow, or $11 today? Would she rather have a 70% chance of $20, or $30 for certain? She held a four-button response box in her right hand. Her left wrist had been fitted with a pulse oximeter; her face was half-rapped in a mirror mask to keep her chin from drifting. She could not move, or speak, or sigh. All she could do — and all the experiment required her to do — was choose.

What was happening in the small room was, on the surface, banal. A young woman pressed buttons in exchange for small amounts of money, and the experimenters recorded her choices. What was happening in her head was anything but banal. Roughly 86 billion neurons were firing in patterns the experimenters could not see, representing goods she had never touched, probabilities she had never computed, and futures that did not yet exist. Hidden inside that cranial vault was a machine — or, more accurately, an uneasy federation of machines — that was computing, in real time, what Maya was worth to herself, what the future was worth relative to the present, what risk felt like in her gut, and whether the offer in front of her was "good enough" to commit a finger to. The scanner was, in a sense, a stethoscope pressed against the chest of that machine, listening to the blood-flow changes that betray where the metabolic action is.

This book is about that machine. It is about the value computation, the prediction engine, the risk register, the social simulator, and the controller that sit inside every human skull and somehow produce what we call a "decision." It is also about the new science that has emerged, in fits and starts over the last quarter century, to listen to that machine with something better than intuition. The science is called *neuroeconomics*, and its central wager is simple but profound: if choice is computation, then somewhere in the brain there must be variables that correspond to the terms in our equations — *value*, *probability*, *risk*, *effort*, *time*, *social regard* — and those variables must be measurable, manipulable, and predictable. The wager paid off. We have, in the last twenty-five years, found the value pixels, the prediction errors, the risk alarms, the social mirrors, and the control switches. We have, in the process, learned that the model of the rational economic agent that dominated twentieth-century thought was not wrong so much as it was incomplete: it described the outputs of a machine it had no instruments to inspect.

What follows in this chapter is the long preamble to that wager. We will start with the idea that classical decision theory left us — a normative scaffold so elegant that it was easy to mistake for a description of the mind. We will then watch that scaffold develop cracks, as psychologists in the 1950s and 1960s showed that real humans systematically violated the axioms. We will then ask: what would it take to put the substrate back under those equations? The answer, we will see, required a parallel revolution in the 1990s — the invention of functional magnetic resonance imaging, the discovery of reward-predicting dopamine neurons, and the founding of a small handful of laboratories willing to bet that the brain's hemodynamic blush could be read as a real-time index of mental computation. Finally, we will preview the five machines that organize the rest of the book: the valuation machine, the prediction machine, the risk machine, the social machine, and the control machine. Each is real, each is localizable, each has its own computational signature, and none of them is the homunculus that classical economics pretended was at the helm.

The Homunculus at the Helm: Classical Decision Theory and Its Hidden Assumptions

To see what neuroeconomics had to do, we have to start with the picture of the human decider that economics inherited and that neuroscience was, eventually, going to disturb. That picture is older than economics itself. It runs, in some form, all the way back to the rational soul of the Scholastics, but its modern, mathematical articulation is usually dated to a brief and famous paragraph in a Swiss pamphlet of 1738.

Daniel Bernoulli was trying to solve a puzzle. A coin would be tossed repeatedly until it came up tails; the pot would then double for every heads that preceded the first tails, and a player who accepted the gamble would be paid the contents of the pot. The expected value of the game is infinite, because the probability of a long run of heads falls off more slowly than the prize grows. Yet no rational person would pay a great deal to play. Bernoulli's resolution, in the *Specimen theoriae novae de mensura sortis*, was to propose that people do not maximize expected monetary value; they maximize the expected value of a concave function of wealth, which he called *utility*. He suggested the logarithm, in part because of a famous anecdote about a friend who would be indifferent between a thousand ducats for certain and a one-in-ten-thousand chance of ten million, even though the latter's expected value is a thousand ducats higher. With a logarithmic utility function, the marginal value of additional wealth shrinks as wealth grows, and the infinite-mean paradox dissolves. The idea was small but it carried enormous freight. A rational agent, Bernoulli implied, was someone who ordered the world by *utility* — by subjective felt value — rather than by *money*. And that ordering was stable, complete, and transitive.

The next major step came two centuries later, in 1944, when John von Neumann and Oskar Morgenstern published *Theory of Games and Economic Behavior*. They did not invent utility; they axiomatized it. If a decision-maker's preferences over a set of outcomes satisfy four conditions — completeness (any two outcomes can be compared), transitivity (if A is preferred to B and B to C, then A is preferred to C), continuity (a blend of A and C is preferred to a blend of B and C only when A and B are also preferred in a corresponding blend), and independence (mixing a lottery with a common third option does not change the preference between two others) — then there exists a real-valued utility function that the agent acts as if she is maximizing in expected value. The result is elegant and powerful. It says: a creature whose preferences satisfy these four conditions is *indistinguishable, from the outside*, from a creature that knows the value of everything and computes the expectation. The mathematics is what we now call *expected utility theory*, and the agent it describes is the textbook Homo economicus.

Leonard Savage, in 1954, took one further step. Where von Neumann and Morgenstern had assumed that probabilities were objective and known, Savage allowed them to be subjective. A decision-maker can have *personal* beliefs about the world — they may believe, for example, that there is a 60% chance of rain tomorrow even though the official forecast says 80% — and as long as her beliefs obey the axioms of probability (non-negativity, normalization, additivity) and her preferences over lotteries obey the von Neumann–Morgenstern axioms, the two can be combined into a single number — *subjective expected utility* — that she acts as if she is maximizing. The model is, again, a normative ideal. It says nothing about what real people actually do. It tells us what a perfectly coherent decider *would* do, giv