The Experience Machine: How Our Minds Predict and Shape Reality

by Andy Clark

According to the new theory (called “predictive processing”), reality as we experience it is built from our own predictions.

Predictive processing speaks to one of the most challenging questions in science and philosophy—the nature of the relationship between our minds and reality. The theory, which has been steadily gaining momentum, changes our understanding of this relationship in ways that have far-reaching implications. Contrary to the standard belief that our senses are a kind of passive window onto the world, what is emerging is a picture of an ever-active brain that is always striving to predict what the world might currently have to offer.

Nothing we do or experience—if the theory is on track—is untouched by our own expectations. Instead, there is a constant give-and-take in which what we experience reflects not just what the world is currently telling us, but what we—consciously or nonconsciously—were expecting it to be telling us. One consequence of this is that we are never simply seeing what’s “really there,” stripped bare of our own anticipations or insulated from our own past experiences. Instead, all human experience is part phantom—the product

of deep-set predictions.

The idea (the main topic of this book) is that human brains are prediction machines. They are evolved organs that build and rebuild experiences from shifting mixtures of expectation and actual sensory evidence.

Predictions and prediction errors are increasingly recognized as the core currency of the human brain, and it is in their shifting balances that all human experience takes shape.

On this alternative account, the perceiving brain is never passively responding to the world. Instead, it is actively trying to hallucinate the world but checking that hallucination against the evidence coming in via the senses.

It alters how we should think about the evidence of our own senses. It impacts how we should think about the way we experience our own bodily states—of pain, hunger, and other experiences such as

feeling anxious or depressed.

This means that we can, at times, change how we feel by changing what we (consciously or unconsciously) predict.

Marr was a towering figure, whose work in neuroscience, computer vision, and AI ranks among the most important contributions ever made to cognitive science.

In Marr’s depiction, visual processing starts by detecting basic ingredients in some incoming signal—an ordered array of pixels, for instance. From there, layered

processing slowly builds toward a more comple...

Eventually, the complex detected patterns are brought into contact with knowledge and memory to deliver (though revealingly, this

part of the puzzle was never satisfactorily solved) a kind of 3D picture of the worldly scene.

The number of neuronal connections carrying signals backward in this way is estimated to exceed the number of connections carrying signals forward by a very substantial margin, in some places by as much as four to one. What is all that downward connectivity feeding information from deep in the brain to regions closer to the sensory peripheries doing?

Real neural wiring like that is costly to install and maintain. The brain, weighing in at about 2 percent of human body weight, is estimated to account

for around 20 percent of total bodily energy consumption.

Things look different, however, once we recognize the attractions of a bold new claim: that brains are nothing other

than large-scale prediction machines.

A predictive brain is a kind of constantly running simulation of the world around us—or at least, the world as it matters to us. Incoming sensory information is used to keep the

model honest—by comparing the prediction to the sensory evidence and generating an error signal when the two don’t match up.

That work goes by various names including “predictive processing,” “hierarchical predictive coding,” and “active inference.” I’ll mostly stick with “predictive processing” as a handy label for this family of theories.

which the brain is asking itself, “Given everything I know, how must the world be for me to be receiving the pattern of signals currently present?” This is the

question that perceptual systems are built to resolve.

The trick is trading intelligence and foreknowledge on the part of the receiver against the costs of encoding and transmitting all the information.

It means that the world we experience is to some degree the world we predict.

Perception itself, far from being a simple window onto the world, is permeated from the get-go by our own predictions and expectations.

Masking is a technique in which a stimulus is briefly shown then rapidly followed by another, different stimulus.

The picture looks different the second time around because improved knowledge about the world (in this case, about the original picture) is enabling your brain to make better predictions.

The first phenomenon in question is called “sine-wave speech.” A sine-wave speech recording is a recording of normal speech that has been artificially degraded in a way that replaces key parts of the sound stream with pure tone whistles. It was invented back in the early 1970s at Haskins Labs in the United States as a means of studying the

nature of speech perception—specifically, to test various theories about what parts of the sound stream are essential for hearing speech. Sine-wave speech sounds like a series of initially unintelligible ascending and descending beeps and whistles.

The striking difference in what you experience again reveals the extent to which what you hear is a construct formed in part by your own predictions.

that what we see is not simply how things are: rather, we see whatever our brain infers (guesses) as the most likely cause of the evidence coming in from the senses.

Remarkably, these self-identified “circadian profiles” correlated strongly with how the dress was perceived. Larks tended to see the dress as white and gold, owls as blue and black.

This means that just by attempting to predict the world we can acquire the knowledge that later enables us to predict the world better. This can seem a bit like magic, conjuring a good predictive model from thin air.

There’s no magic involved, but the

conjuring trick is nonetheless impressive! By trying (and failing) to predict the world, we can learn to do bette...

What this case and many others in this chapter show is that pain can sometimes be remarkably disconnected from standard bodily causes.

that “pain don’t hurt,” become much less puzzling once we realize that brains build human experience only by combining their own predictions with sensory evidence. In the same way that chronic expectations of incoming calls had caused me to feel phantom phone vibrations in my pocket, strong expectations of pain (from seeing the ripped boot and the protruding

nail) caused the construction worker to experienc...

Instead, experience always and everywhere reflects those rich webs of prior knowledge and here-and-now expectation.

“nociplastic pain.” This was defined as pain arising from abnormal processing of pain signals without any clear evidence of either tissue damage or any other recognized systemic pathology.

conscious and nonconscious expectations about our own states of pain can make a surprisingly large difference to the amounts of pain we experience.

We have seen that human experience arises at the meeting point of predictions and sensory evidence.

This result fits well with the idea that the brain’s best estimates of precision (reliability) play an integral role in shaping our experience. Only predictions that our brains estimate to be reliable get to exert a powerful influence on our sensations.

what makes placebo analgesia effective is that it activates conscious or unconscious expectations of relief. The higher the patient’s estimate of the power of the intervention, the greater its effect.

Nocebo effects are the inverse of placebo effects and occur when expectations and predictions cause us to experience unwanted symptoms rather than relief.

Self-Confirming Cycles of Pain

The results were clear. Experienced levels of pain were pulled upward or downward depending on the cue-based expectations, and these effects were reflected in the neural pain signatures.