I speak two languages and am currently learning the third. Often, the search for the term in the right language miserably fails. For instance, I speak English but will insert a word or two in German. It is not because I want to make it sound fancy, but because I am unable to quickly find the right word. I may swap “almost” with “fast” and back (and ”fast” is “almost” in German). When I speak like that, people around me are confused. But not because they don’t understand German, but because they don’t anticipate me answering in German when we were speaking English just before that. I assume that at this point the reader expects me to say what the topic and the title of the book that I have read are. Without further ado, I fulfill your expectations. The book is “The experience machine: how our minds predict and shape reality” by Andy Clark.

The main idea of the book and much of Andy Clark’s work as a philosopher is that the brain is a predictive machine. There are four building blocks of the predictive machine that are useful to understand.

1. A generative model is the internal probabilistic model that the brain (or computational system) uses to generate predictions about sensory input. The brain assumes that some hidden causes in the world (we may also say “latent variables”) give rise to sensory signals, and the generative model is the model of the causes. Example: when you read the sentence in a familiar language, you don’t just read it word by word. Instead, you slide through the sentence because you know the general rules of grammar, such as which type of word usually follows an adjective (a noun), where the verb is in the sentence, or what is to be expected after a comma or colon. 

2. Using a generative model, the brain can generate predictions of incoming sensory signals based on its current beliefs. Continuing the example. When you read the sentence, you at the same time generate the predicted sentence in your brain (it happens automatically without much awareness, unless the language is not familiar). Sometimes, you would even jump to the next sentence because the end of the current one is quite obvious.

3. Predicted outcome often matches the real outcome, but not all the time. The prediction error is the difference between the generated outcome and the real outcome. Continuing example. You may start reading the sentence expecting the subject and action to be in the first couple of words. In case they are not there, you surprise or confusion will feel. Or as Yoda said, "The greatest teacher, failure is”.

4. If you live in a country where everyone speaks like Yoda, you would need to wait till the end of the sentence to actually hear the verb that carries the meaning. It is precision-weighting. Based on the environment, predictions may be favored or discouraged, and sensory evidence correspondingly may be suppressed or amplified. Example (in German, this is exactly the case). *Ich habe beschlossen, das Büro früher zu verlassen, weil die Besprechung so langweilig war, dass ich mich nicht mehr konzentrieren konnte.* [I have decided, the office earlier to leave, because the meeting so boring was, that I myself not more concentrate could.]

My example was about the language, but predictions, according to this theory, are built for all types of sensory input. Sometimes predictions are useful and sometimes detrimental. The placebo effect (the belief in the healing power of a substance that has no effect) is a type of prediction that can be effectively used. The nocebo effect (the belief in the harm of a substance that is not harmful) is quite the opposite. 

The extreme example of unconscious predictions is described in the book. The case of a teenage girl was observed by the neurologist Jon Stone from the University of Edinburgh. The girl’s vision progressively deteriorated, and one day she became completely blind. However, there were no structural deficits in her eyes, along the visual tracts, or in the visual cortex. She was diagnosed with functional neurological disorder. In this lucky case, the girl has recovered her vision. It turned out that she had migraines in the past that were triggered by light, and she spent a long time in dark rooms. Her brain learned the association. With behavioral therapy and transcranial magnetic therapy (TMS), the doctor managed to reverse the false predictions. The author stresses that it doesn’t mean that the patient has simply “imagined” the disease, and they should not be treated as such if it is all in their imagination. The disease is real, even if there is no physical damage, and those patients deserve to be treated (but maybe with different methods).

If predictions indeed change perceptions, this can be used for therapy. Examples are virtual reality of winter scenery for patients dealing with pain from burns, conscious reframing of chronic pain, and self-affirmations in mental health.

The theory of predictive processing seems to be able to explain everything. That is why I’m reluctant to jump on this train. I believe prediction error can explain some things in certain circumstances, but I don’t believe (yet) that it is a theory of everything. The author treats “prediction error” quite liberally, referring to computation at the neuronal level as well as our daily experiences related to expectation. It could be, of course, that a common principle governs every level of hierarchy. However, it could also be that this explanation just fits our human experiences (our understanding) and is simply being stretched to other levels. Also, it could only appear as if the system is computing prediction errors while, in reality, no errors are being computed. I’m curious to see whether, in the future, the theory will be supported by empirical research.

Favorite quote:

“False expectations, once they get a grip on us, become increasingly resistant to change”

Ocotber, 2025