Algorithms for Self-Monitoring - Building Minds That Know Themselves

Know Thyself: The Science of Self-Awareness - Stephen M Fleming 2021

Algorithms for Self-Monitoring
Building Minds That Know Themselves

Once the ability to sense external goings-on has developed, however, there ensues a curious side effect that will have vital and radical consequences. This is the fact that the living being’s ability to sense certain aspects of its environment flips around and endows the being with the ability to sense certain aspects of itself.


In the previous chapter, we saw how the perceptual systems of the brain work, taking in information from the senses and solving inverse problems to build a model of the world. We also saw how the ability to encode and track different sources of uncertainty about what we are seeing and hearing is critical to how we perceive and can also be harnessed to doubt ourselves. In the next leg of our tour of self-monitoring algorithms, we are going to start at the opposite end of the system: the parts that control our actions.

We typically think of an action as something conscious and deliberate—preparing dinner, for instance, or picking up the phone to call a friend. But the range of possible actions that we initiate over the course of a normal day is much broader. Everything that our bodies do to change things in our environment can be considered an action: breathing, digesting, speaking, adjusting our posture. And, as this set of examples illustrates, many of our actions occur unconsciously and are critical for keeping us alive.

The problem with actions, though, is that they do not always go to plan. This means that we need ways of correcting our mistakes quickly and efficiently. Any self-aware system worth its salt would not want to simply fire off actions like darts, with no way of correcting their flight after they have been thrown. Imagine reaching for your glass of red wine at a dinner party, only to misjudge and tip it off the edge of the table. You watch in horror, feeling helpless to do anything about it. And yet, sometimes, if you’re lucky, your hand will seem to shoot out of its own accord, catching the glass barely milliseconds into its descent. We will see that at the heart of being able to correct our actions in this way is an ability to predict what should have happened—bringing the glass into our hand—but didn’t.

In fact, prediction is at the heart of algorithms that can monitor themselves. Consider the predictive text on your smartphone. The only way it can correct ham-fisted errors in your typing is by having some running estimate of what word you intended to say in the first place (and if those estimates are, in fact, at odds with your intentions, then you experience maddeningly persistent sabotage). This is also intuitively true of self-awareness. We can only recognize our errors and regret our mistakes if we know what we should have done but didn’t do. The French have a wonderful phrase for this: “L’esprit d’escalier,” or staircase wit. These are the things you realize you should have said upon descending the stairs on your way out of a party.

In this chapter, we will discover a panoply of algorithms that predict and correct our actions. As we will see, brains of all shapes and sizes have exquisite fail-safe mechanisms that allow them to monitor their performance—ranging from the smallest of corrective adjustments to the movement of my arm as I reach out to pick up my coffee, all the way to updating how I feel about my job performance. Being able to monitor our actions is our second building block of self-awareness.

Predicting Our Errors

Some of the simplest and most important actions that we engage in are those aimed at maintaining our internal states. All living things must monitor their temperature, nutrient levels, and so on, and doing so is often a matter of life or death. If these states move too far from their ideal position (known as set points), staying alive becomes very difficult indeed. Consider a humble single-celled bacterium. Living cells depend on managing the acidity of their internal world, because most proteins will cease to function beyond a narrow range of pH. Even simple bacteria have intricate networks of sensors and signaling molecules on their cell surface, which lead to the activation of pumps to restore a neutral pH balance when required.

This is known as homeostasis, and it is ubiquitous in biology. Homeostasis works like the thermostat in your house: when the temperature drops below a certain point, the thermostat switches on the heating, ensuring that the ambience of your living room is kept within a comfortable range. A curious aspect of homeostasis is that it is recursive—it seeks to alter the very same thing that it is monitoring. The thermostat in my living room is trying to regulate the temperature of the same living room, not some room in my neighbor’s house. This feature of homeostasis is known as a closed-loop system. If the state it is detecting is in an acceptable range, then all is well. If it’s not—if an imbalance in pH or temperature is detected—some action is taken, and the imbalance is corrected. Homeostasis can often be left to its own devices when up and running; it is rare that a corrective action will not have a desired effect, and the control process, while intricate, is computationally simple.

Homeostatic mechanisms, however, operate in the here and now, without caring very much about the future. A simple on-off thermostat cannot “know” that it tends to get colder at night and warmer during the day. It just switches on the heating when the temperature drops below a threshold. In the BBC comedy series Peep Show, Jez misunderstands this critical feature of thermostats, telling his housemate Mark, “Let’s whack [the boiler] up to 29.… I don’t actually want it to be 29, but you’ve got to give it something to aim for. It’ll get hotter, quicker.” Mark replies disdainfully (and accurately): “No it won’t, it’s either on or off. You set it, it achieves the correct temperature, it switches off.” You cannot trick a boiler.

The new breed of learning thermostats, such as the Nest, improves on traditional on-off devices by learning the typical rise and fall in temperature over the course of the day and the preferences of the owner for particular temperatures. A smart thermostat can then anticipate when it needs to switch on to maintain a more even temperature. The reason this is more successful than a good-old-fashioned thermostat is a consequence of a classic proposal in computer science known as the good regulator theorem, which states that the most effective way of controlling a system is to develop an accurate model of that same system. In other words, the more accurate my model of the kind of things that affect the temperature, the more likely I will be able to anticipate when I need to make changes to the heating to keep it within a comfortable range.1

The same is true when we move beyond homeostasis to actions that affect the external world. In fact, we can think of all our behavior as a form of elaborate homeostasis, in the sense that many of the things we do are aimed at keeping our internal states within desirable bounds. If I am hungry, I might decide to go and make a sandwich, which makes me feel full again. If I need money to buy ingredients to make a sandwich, I might decide to apply for a job to make money, and so on. This idea—that everything we do in life fits into some grand scheme that serves to minimize the “error” in our internal states—has both its proponents and critics in the field of computational neuroscience. But at least for many of our simpler actions, it provides an elegant framework for thinking about how behavior is monitored and controlled. Let’s take a closer look at how this works in practice.2

Who Is in Control?

In the same way that there are dedicated sensory parts of the brain—those that handle incoming information from the eyes and ears, for instance—there are also dedicated motor structures that send neural projections down to the spinal cord in order to control and coordinate our muscles. And just as the visual cortex is organized hierarchically, going from input to high-level representations of what is out there in the world, the motor cortex is organized as a descending hierarchy. Regions such as the premotor cortex are involved in creating general plans and intentions (such as “reach to the left”), while lower-level brain areas, such as the primary motor cortex, are left to implement the details. Regions in the prefrontal cortex (PFC) have been suggested to be at the top of both the perceptual and motor hierarchies. This makes sense if we think of the PFC as being involved in translating high-level perceptual representations (the red ball is over there) into high-level action representations (let’s pick up the red ball).3

One consequence of the hierarchical organization of action is that when we reach for a cup of coffee, we do not need to consciously activate the sequence of muscles to send our arm and hand out toward the cup. Instead, most action plans are made at a higher level—we want to taste the coffee, and our arm, hand, and mouth coordinate to make it so. This means that in a skilled task such as playing the piano, there is a delicate ballet between conscious plans unfolding further up the hierarchy (choosing how fast to play, or how much emphasis to put on particular passages) and the automatic and unconscious aspects of motor control that send our fingers toward the right keys at just the right time. When watching a concert pianist at work, it seems as though their hands and fingers have a life of their own, while the pianist glides above it all, issuing commands from on high. As the celebrated pianist Vladimir Horowitz declared, “I am a general, my soldiers are the keys.” In the more prosaic language of neuroscience, we offload well-learned tasks to unconscious, subordinate levels of action control, intervening only where necessary.4

Not all of us can engage in the finger acrobatics required for playing Chopin or Liszt. But many of us regularly engage in a similarly remarkable motor skill on another type of keyboard. I am writing this book on a laptop equipped with a standard QWERTY keyboard, named for the first six letters of the top row. The history of why the QWERTY keyboard, designed by politician and amateur inventor Christopher Latham Sholes in the 1860s, came into being is murky (the earliest typewriters instead had all twenty-six letters of the alphabet organized in a row from A to Z, which its inventors assumed would be the most efficient arrangement). One story is that it was to prevent the early typewriters from getting jammed. Another is that it helped telegraph operators, who received Morse code, quickly transcribe closely related letters in messages. And yet another is that Remington, the first major typewriter manufacturer, wanted to stick with QWERTY to ensure brand loyalty from typists who had trained on its proprietary system.

Whichever theory is correct, the English-speaking world’s QWERTY typewriter has led millions of people to acquire a highly proficient but largely unconscious motor skill. If you are a regular computer user, close your eyes and try to imagine where the letters fall on your keyboard (with the exception of the letters Q-W-E-R-T-Y!). It is not easy, and if you are like me, can only really be done by pretending to type out a word. This neat dissociation between motor skill and conscious awareness makes typing a perfect test bed for studying the different kinds of algorithms involved in unconsciously monitoring and controlling our actions. Typing can also be studied with beautiful precision in the lab: the initiation and timing of keystrokes can be logged by a computer and the movements of people’s fingers captured by high-resolution cameras.

Using these methods, the psychologists Gordon Logan and Matthew Crump have carried out detailed and creative experiments to probe how people type. In one of their experiments, people were asked to type out the answers to a classic psychological test, the Stroop task. In the Stroop, people are asked to respond to the color of the ink a word is written in—typing “blue” for blue ink and “red” for red ink, for instance. This is straightforward for most words, but when the words themselves are color words (such as the word “green” written in blue ink, “purple” written in red ink, and so on) it becomes much more difficult, and people slow down and make errors when the word and the ink color don’t match. But despite being slower to initiate typing the word, they were no slower to type the letters within the word once they had gotten started (for instance, b-l-u-e). This led to the hypothesis that there are multiple action control loops at work: a higher-level loop governing the choice of which word to type, and a lower-level loop that takes this information and works out which keys need to be pressed in which order.5

Not only are there multiple levels of action control, but the higher levels know little about the workings of the lower levels. We know this because one of the easiest ways to screw up someone’s typing is to ask them to type only the letters in a sentence that would normally be typed by the left (or right) hand. Try sitting at a keyboard and typing only the left-hand letters in the sentence “The cat on the mat” (on a QWERTY keyboard you should produce something like “Tecatteat,” depending on whether you normally hit the space bar with your right or left thumb). It is a fiendishly difficult and frustrating task to assign letters to hands. And yet the lower-level loop controlling our keystrokes does this continuously, at up to seventy words per minute! Some part of us does know the correct hand, but it’s not able to get the message out.6

Staying the Course

These experiments suggest that fine-scale unconscious adjustments are continuously being made to ensure that our actions stay on track. Occasionally, these unconscious monitoring processes become exposed, similar to how visual illusions revealed the workings of perceptual inference in the previous chapter. For instance, when I commute to work on the London Tube, I have to step onto a series of moving escalators, and I rely on my body making rapid postural adjustments to stop me from falling over when I do so. But this response is so well learned that if the escalator is broken and stationary, it’s difficult to stop my motor system from automatically correcting for the impact of the usually moving stairs—so much so that I now have a higher-level expectation that I will stumble slightly going onto a stationary escalator.7

In a classic experiment designed to quantify this kind of rapid, automatic error correction, Pierre Fourneret and Marc Jeannerod asked volunteers to move a computer cursor to a target on a screen. By ensuring that participants’ hands were hidden (so that they could see only the cursor), the researchers were able to introduce small deviations to the cursor position and observe what happened. They found that when the cursor was knocked off course, people immediately corrected it without being aware of having done so. Their paper concluded: “We found that subjects largely ignored the actual movements that their hand had performed.” In other words, a low-level system unconsciously monitors how we are performing the task and corrects—as efficiently as possible—any deviations away from the goal.8

One part of the brain that is thought to be critical for supporting these adjustments is known as the cerebellum—from the Latin for “little brain.” The cerebellum looks like a bolt-on secondary brain, sitting underneath the main brain. But it actually contains over 80 percent of your neurons, around sixty-nine billion of the eighty-five billion total. Its circuitry is a thing of regimented beauty, with millions of so-called parallel fibers crossing at right angles with another type of brain cell known as Purkinje neurons, which have huge, elaborate dendritic trees. Inputs come in from the cortex in a series of loops, with areas in the cortex projecting to the same cerebellar regions from which they receive input. One idea is that the cerebellum receives a copy of the motor command sent to the muscles, rather like receiving a carbon copy of an email. It then generates the expected sensory consequences of the action, such as the fact that my hand should be smoothly progressing toward the target. If this expectation does not match the sensory data about where my hand actually is, rapid adjustments can be made to get it back on course.9

In engineering, this kind of architecture is known as a forward model. Forward models first predict the consequences of a particular motor command and then track the discrepancy between the current state and what is expected to happen to provide small corrections as and when they are needed. Since I was a child I have loved to sail, either racing in smaller dinghies or cruising on bigger boats. When we are sailing from one place to the next, I can use a simple forward model to help counteract the effects of the tide on the position of the boat. If I plot a course to a harbor where I want to end up, the GPS tells me whether I’m left or right of the straight-line track, and I can correct accordingly without worrying about my overall heading. Often this results in the boat crabbing sideways into the tide, similar to how you would instinctively row upstream when crossing a river. To a casual observer, it looks like the adjustment for the tide was carefully planned in advance, when in fact it was the result of lots of small corrections based on local error signals.

In this kind of algorithm, what matters is keeping track of deviations from what I planned to happen. This means that information coming in from my senses can simply be ignored if it is in line with what I expect—another feature of predictive control that is borne out by experiments on human volunteers. For instance, if I were to passively move your arm, your brain would receive information that your arm is moving from the change in position of the joints and muscles. But if I move my arm myself, this sensory feedback is dampened, because it is exactly in line with what I expected to happen (this is also why you can’t tickle yourself). These neural algorithms set up to detect deviations in our movements can lead to some counterintuitive phenomena. If you throw a punch in a boxing ring or a bar brawl, your unfortunate punch-recipient would feel the punch more acutely on his face than you, the punch-thrower, would on your hand. This is because the punch-thrower expects to feel the punch, whereas the punch-recipient does not. If the punch-recipient decides to retaliate, he will then throw a stronger punch than the punch-thrower believes he has thrown, in order to match the punch he feels he has just received. And so on, in a vicious cycle of escalating tit for tat. If you have ever sat in the front seat of a car while two kids are fighting in the back, you will know how this scenario can play out.10

What all these experiments tell us is that there are a range of self-monitoring processes churning away in the absence of conscious awareness. We are able to act smoothly and swiftly, usually without thinking about it, thanks to the power and flexibility of predictive control. When I step onto a moving escalator during rush hour, small corrections to my posture are made based on local algorithms, just as subtle adjustments to the course of a sailing boat are made to take into account effects of the tide. But if the deviation in what we expected to happen is large—if we are missing the target by a mile—then minor corrections or adjustments are unlikely to be able to bring it back into line. This is the point at which unconscious adjustments to our actions morph into conscious recognition of making an error.

From Detecting Errors to Learning About Ourselves

One of the first studies of how we become aware of our errors was carried out by psychologist Patrick Rabbitt in the 1960s. He designed a difficult, repetitive task involving pushing buttons in response to sequences of numbers. The actual task didn’t matter too much—the clever part was that he also asked people to push another button if they detected themselves making an error. Rabbitt precisely measured the time it took for these additional button presses to occur, finding that people were able to correct their own errors very quickly. In fact, they realized they had made an error on average forty milliseconds faster than their fastest responses to external stimuli. This elegant and simple analysis proved that the brain was able to monitor and detect its own errors via an efficient, internal computation, one that did not depend on signals arriving from the outside world.

This rapid process of error detection can lead to an equally rapid process of error correction. In a simple decision about whether a stimulus belongs to category A or B, within only tens of milliseconds after the wrong button is pressed, the muscles controlling the correct response begin to contract in order to rectify the error. And if these corrective processes happen fast enough, they may prevent the error from occurring in the first place. For instance, by the time our muscles are contracting and we are pressing the send button on a rash email, we might have accumulated additional evidence to suggest that this is not a good idea and withhold the critical mouse click at the last moment.11

A couple of decades after Rabbitt’s experiment, the brain processes that support internal error detection were beginning to be discovered. In his PhD thesis published in 1992, the psychologist William Gehring made electroencephalograph (EEG) recordings from participants while they performed difficult tasks. EEG uses a net of small electrodes to measure the changes in the electrical field outside the head caused by the combined activity of thousands of neurons inside the brain. Gehring found that a unique brain wave was triggered less than one hundred milliseconds after an error was committed. This rapid response helps explain why Rabbitt found that people were often able to very quickly recognize that they had made an error, even before they were told. This activity was labeled the error-related negativity (ERN), which psychologists now affectionately refer to as the “Oh shit!” response.12

We now know that the ERN occurs following errors on a multitude of tasks, from pressing buttons to reading aloud, and is generated by a brain region buried in the middle of the frontal lobe: the dorsal anterior cingulate cortex (dACC). This tell-tale neural signature of self-monitoring is already in place early in human development. In one experiment, twelve-month-old babies were flashed a series of images on a computer screen, and their eye movements recorded. Occasionally one of the images would be a face, and if the babies looked toward it, they would get a reward in the form of music and flashing colored lights. The occasions on which the baby failed to look at the face are errors in the context of the experiment—they did not perform the action that would get them the reward. On these occasions, EEG recordings showed a clear ERN, although somewhat delayed in time compared to what is typically seen in adults.13

We can think of the ERN as a special case of a “prediction error” signal. Prediction errors do exactly what they say on the tin—they keep track of errors in our predictions about the future, and they are a central feature of algorithms that can efficiently learn about the world. To see how prediction errors help us learn, imagine that a new coffee shop opens up near your office. You don’t yet know how good it is, but they have taken care to buy a top-of-the-line espresso machine and get the ambience just right. Your expectations are high—you predict that the coffee will be good before you’ve even tasted it. When you sip your first cup, you find that it’s not only good—it’s one of the best cups of coffee you have had in a long time. The fact that the coffee was better than expected leads you to update your estimate, and it becomes your new favorite stop on the way in to work.

Now let’s imagine a few weeks have gone by. The baristas have become complacent and the coffee is no longer as good as it used to be. It might still be good, but compared to what you expected, this is experienced as a negative error in your prediction, and you might feel a little more disappointed than usual.

The ability to make and update predictions depends on a famous brain chemical, dopamine. Dopamine is not only famous, but it is also commonly misunderstood and often referred to as the “pleasure” chemical in the popular media. It is true that dopamine is boosted by things that we enjoy, from money to food to sex. But the idea that dopamine simply signals the rewarding character of an experience is incorrect. In the 1990s, a now classic experiment was carried out by the neuroscientist Wolfram Schultz. He recorded signals from cells in the monkey midbrain that produce dopamine and deliver it to other brain areas. Schultz trained the monkeys to expect a drop of juice after a light was switched on in the room. Initially, the dopamine cells responded to the juice, consistent with the pleasure theory. But over time, the animals began to learn that the juice was always preceded by the light—they learned to expect the juice—and the dopamine response disappeared.14

An elegant explanation for the pattern of dopamine responses in these experiments is that they were tracking the error in the monkeys’ prediction about the juice. Early on, the juice was unexpected—just like the unexpectedly good coffee from the new shop. But over time, the monkeys came to expect the juice every time they saw the light, just as we would come to expect good coffee every time we walked into the cafe. Around the same time that Schultz was performing his experiments, the computational neuroscientists Peter Dayan and Read Montague were building on classic work on trial-and-error learning in psychology. A prominent theory, the Rescorla-Wagner rule, proposed that learning should only occur when events are unexpected. This makes intuitive sense: If the coffee is just the same as yesterday, I don’t need to alter my estimate of the goodness of the coffee shop. There is no learning to do. Dayan and Montague showed that versions of this algorithm provided an excellent match to the response of dopamine neurons. Shortly after Schultz, Dayan, and Montague’s work was published, a series of studies by my former PhD adviser Ray Dolan discovered that the neural response in regions of the human brain that receive dopamine input closely tracks what one would expect of a prediction error signal. Together, these pioneering studies revealed that computing prediction errors and using them to update how we experience the world is a fundamental principle of how brains work.15

Now that we’re armed with an understanding of prediction errors, we can begin to see how similar computations are important for self-monitoring. Occasionally we directly experience positive or negative feedback about our performance—on an assignment at school, for instance, or when we learn we have beaten our personal best over a half-marathon distance. But in many other areas of everyday life, the feedback may be more subtle, or even absent. One useful way of thinking about the ERN, then, is that it reflects an internal signal of reward or, more specifically, the absence of reward. It conveys the difference between what we expected (to perform well) and what actually happened (an error).

Consider sitting down to play a simple melody on the piano. Each note has a particular sound to it, but it would be strange to say that any note is “better” or “worse” than another. Played alone, an A is no more rewarding than a G-sharp. But in the context of a melody such as the opening of Grieg’s Piano Concerto in A Minor, mistakenly playing a G-sharp instead of an A is immediately jarring, and we would wince at the clash. Even if no external feedback is involved, playing a wrong note is an error in how we expected to perform. And by keeping track of these performance errors, the brain can estimate whether it is doing well or badly—even in the absence of explicit feedback.16

By definition, errors are not usually made when we expect to make them—otherwise we would have been able to take steps to prevent them from happening. This feature of human error is used for comic effect in one of my favorite sketches from The Fast Show. A genial old man called Unlucky Alf turns to the camera and says in a broad northern English accent: “See that down there? They’re digging a bloody great hole at the end of the road. Knowing my luck I’ll probably fall down that.” We watch and wait as he slowly ambles off down the road, at which point an enormous gust of wind picks up and he is blown into the hole. The sketch is funny because of the preparation and foresight that went into unsuccessfully avoiding disaster. It is more typical that we are surprised by errors precisely because we didn’t see them coming, and, like Homer in The Simpsons, exclaim “D’oh!” when we recognize them after the fact.

A powerful way of implementing self-monitoring, then, is to create predictions for how we expect to perform and keep track of whether or not we performed as intended. If we make a mistake, this is registered as a negative error in our prediction of success. Remarkably, there is a beautiful symmetry between the brain circuitry involved in detecting external rewards—such as whether the coffee was better or worse than expected, or whether we received a recent bonus at work—and those involved in tracking internal errors in our performance. Both appear to rely on dopamine. For instance, in zebra finches, introducing unexpected sounds into what the birds hear of their own birdsong leads to reduced firing in dopamine neurons. These particular dopamine neurons project to another brain region involved in song learning, as if the dopamine is signaling whether a recent effort was good or bad—sort of like a resident talent-show judge in the bird’s brain. The same circuits tracking internal errors in song production also track external rewards, just as we would expect if a common prediction error algorithm is involved in learning about both the world and ourselves.17

Becoming Self-Aware

Let’s recap. We have encountered cases in which error correction is applied at multiple levels of the system—from detecting and correcting changes in internal states (homeostasis) to ensuring our actions remain in line with what we intended to do. Many of these forms of self-monitoring are widespread in the animal kingdom and appear early in human development.18 We have also seen that algorithms for estimating uncertainty and monitoring our internal states and actions are ubiquitous features of complex, self-regulating systems such as the human brain. These two building blocks form the core of what psychologists refer to as implicit metacognition—forms of self-monitoring that often proceed automatically and unconsciously. In contrast, explicit metacognition refers to those aspects of metacognition of which we are also consciously aware. When I become convinced that I have made a hash of a task at work, then I am engaging explicit metacognition.

A useful (if coarse) analogy for the relationship between implicit and explicit metacognition is the interaction between the pilots and autopilot of a modern airliner. The aircraft has an electronic “brain” in the form of its autopilot, which provides fine-grain self-monitoring of the plane’s altitude, speed, and so on. The pilots, in turn, perceive and monitor the workings of the plane’s autopilot, and such monitoring is governed by the workings of their (biological) brains. The interaction between pilot and autopilot is a rudimentary form of “aircraft awareness”—the pilots are tasked with being aware of what the plane’s autopilot is doing and intervening where necessary. The same is true of the explicit and implicit aspects of metacognition, except now the interaction is all taking place within a single brain.

This does not mean there is the equivalent of an internal pilot sitting there monitoring what is happening in our heads. The concepts and models we use to describe how the mind works tend to be different from the concepts and models we use to describe its implementation in neural hardware. As an analogy, it makes sense to talk about the words of this book “existing” in my word-processing software, but it makes less sense to search for the words in the 1s and 0s zipping around my laptop’s circuit board. In the same way, we might talk of self-awareness as involving something that “monitors” and “observes” other cognitive processes (a psychological or computational level of analysis), but this does not mean there is an actual monitor or observer to be found when we peer inside the head. The field of cognitive neuroscience is making us increasingly familiar with the notion that there is no single place in the brain where feelings happen or decisions are made, and the same is true of metacognition—there is no single location where self-awareness “happens.”19

At the psychological level, though, the picture from cognitive science is that many of the relevant processes needed to “fly” the mind and body can be handled on autopilot, without involving explicit metacognition. A vast array of predictions and prediction errors are triggering continual adjustments to keep our mental planes flying straight and level—but for the most part they remain hidden from view, just as airline pilots are often oblivious to the continual adjustments an autopilot is making to keep their plane locked to a height of thirty thousand feet.

Many animals have this rich capacity for implicit metacognition—they can sense when they are uncertain and track errors in their actions. This helps them pass metacognitive tests such as Smith’s uncertain-response test that we encountered in the previous chapter. Human infants as young as twelve months old also display sophisticated capacities for implicit metacognition. But by the time we reach adulthood, most of us have also acquired an explicit form of metacognition that allows us to consciously think about our own minds and those of others.20

The question that remains, then, is why? Why did we gain a remarkable ability to become aware of ourselves? Implicit metacognition—our vast array of unconscious autopilots—seems to be doing just fine without it. Why did evolution bother making any of this conscious?