Can AI predict your next move?

We are on the brink of another Copernican revolution: Artificial Super Intelligence that surpasses humans in virtually every respect. What will this mean for humanity and the human condition? Will AI be conscious? What if AI is able to foresee every move you are going to make? What will life be like when AI is simply going to manipulate you in doing what it wants you to do? Well, the good news is that there are promising theories saying that we all have free will. The bad news; we should be more afraid of the people controlling AI than the AI itself. The most explosive and most profound innovation of humanity to do date is resurfacing an age-old debate: determinism, free will, and the nature of our consciousness.

Materialism vs. Idealism

The world in which AI is capable of foreseeing your very thoughts and actions, is rooted in two paradigms: materialism and determinism. Materialism supposes that everything is physical. Everything we experience and everything that is, is constituted by interactions of physical particles. Somehow, our consciousness emerges from those interactions. There is no universally accepted theory for how this happens. This is what is called the 'hard problem of consciousness'.

On the opposite side sits idealism. It posits that consciousness is a quantum process and a fundamental unit of the universe. Depending on who you ask, consciousness actually gives rise to matter, space, and time. Mind is a precursor to matter. The goal of this essay is not to provide evidence for these views. Instead, it aims to show how various aspects of the conventional view on reality are untenable within the existing paradigm. This should prepare our thinking to move beyond. The scientific paradigm is shifting. The further we venture, the smaller the gap between science and spirituality seems to become.

Determinism

First, we have to talk about determinism. It basically says that, given full knowledge over a system and a complete model for understanding it, we will always be able to predict its outcome. It is the billiard-ball view of nature: knowing the initial force and direction of the white ball, you know the exact positions of all balls at the end of the turn. Extending this view to all the atoms in the universe, we get Laplace's Demon: the idea that the complete past, present, and future are predetermined and knowable for an all knowing entity.

Probabilistic determinism

Now a variation to this view is one in which we are working with probability. The idea is that we can not predict the precise outcome of a system. Reality is not deterministic, but can reliably be understood through probability distributions. I would call this view probabilistic determinism, because it carries over the flavour of determinism. The requirement for single outcome predictions has simply been replaced with a probability distribution.

Chaos theory

The first blow to determinism is dealt by chaos theory. It turns out that, for chaotic systems, tiny variations in the initial conditions have an unrecognisable effect on the outcome. These systems are characterised by the 'butterfly effect'. To accurately predict the weather in Eindhoven next year on any particular date, we have to know exactly what every butterfly in Australia is up to. We simply do not have the measurement instruments to get to this level of accuracy, and it is not conceivable that we will be able to circumvent this problem. Something similar is going on with the Three-Body-Problem. The only way to get reasonable accuracy of the celestial movements of three bodies circling around each other, is to regularly sample the data and make predictions on a short time horizon.

The point is that, even in a purely deterministic world, there are limits to what can be known. We can invent better measurement instruments to get the precise data that we need in principle. What would we measure though? Would we measure the momentum of each cell in the butterfly's wing? How would this be used in a grand formula to predict the weather? I think what this shows is that the world is extremely complicated. The universe is governed by strongly interconnecting processes that are hard at least and impossible at worst to disentangle. In practice, the information we need to make accurate predictions is inaccessible.

Computational irreducibility

The second big idea is computational irreducibility as coined by Stephen Wolfram. It says that, "the only way to determine the answer to a computationally irreducible question is to perform, or simulate, the computation". This is a problem. The point of a model is to pre-create or re-create reality in order to predict the likelihoods of certain outcomes in the system without having to go through the actual motions.

I like to think of a model as a formula or set of computational steps that we follow in the right order to get a fitting probability distribution of the outcomes. This is much like following the exact steps of a recipe (the algorithm) with the right ingredients (inputs) to get the output we expected (a delicious Indian dahl). Without a model to describe the movements of billiard balls, we would have to track the minuscule forces and interactions of all particles of each ball. Fortunately, the lower level mechanics reliably generate a pattern of behaviour that we can understand on a higher-level. We can formulate a model to understand the system on the level of individual billiard balls.

According to Wolfram, this is the exception rather than the rule. Not all processes can be understood in fewer computational steps than the ones constituting the process. Rule 110 is a famous example. It is a cellular automaton with a simple set of rules that generates complex behaviour. The only way to predict the state of the system at a particular time step, is by calculating the evolution of the system at each iteration until that time step. The model is the system.

The nature of models

I would argue that a model is anything that aims to represent a process in fewer computational steps than the actual process itself. Models can then be evaluated based on two essential metrics: effectiveness, and efficiency. Effectiveness is the degree to which stated goals are reached. In the case of forecasting, we can equate effectiveness with accuracy. Efficiency is the ratio between goals achieved and resources spent. In the light of prediction, we can finetune this definition to mean the difference between the number of computational steps in the process at hand and the steps that the model takes to make a prediction. A deterministic model with a 100% accuracy for predicting the next state in Rule 110, is maximally inefficient because it performs the exact number of steps that constitute the process itself.

So far we have talked about physical systems only. With chaos theory we have seen that some of the information we need to make accurate predictions is inaccessible. Through computational (ir)reducibility we have understood the nature of models and how they are intended to map to reality. When humans and agents with some form of autonomy come into play, we enter a different domain. Here, inaccessibility of information is a property of the system.

Taleb's turkey

Consider the story painted by Nassim Taleb about a turkey. Throughout the whole year, it is reliably fed by its human caretakers. This dramatically changes when Thanksgiving arrives and the turkey is slaughtered. From the perspective of the turkey, it is entirely rational to become increasingly confident in a regular food supply. Yet the caretakers have obvious intent in their actions. As a principle, this is a real threat to one of the foundational principles of modern science: empiricism. It states that if a model fits the data, it is an accurate representation of reality and we should increase our confidence in the model.

I believe this has to do with the fact that all autonomous agents are embedded in a larger environment. They occupy a certain place in information space. All information is accessible in principle, but one will have to move through information space to get it; they can only acquire information by leveraging the informational affordances in their environment.

Computing climate change

I would like to conduct a thought experiment to extend this idea. Suppose you want to predict the global temperatures by the year 2100. We can use the latest climate models developed by the International Panel for Climate Change. They defined various scenarios for global temperature rise. But instead of those scenarios, we are interested in how much global temperatures will actually rise. That is, the predictions of the model itself presumably inform the policies and actions of institutions, organisations, and our fellow human beings. What and at what level do we need to measure in order to predict the course of the planetary ecosystem?

A few ways of doing that would be to look at actual policies currently in place, the (contractual) commitments from various organisations, or the use of detailed economic data and projections. We can hardly expect this to be accurate however: policies change, priorities shift, people change their mind. To get a (probabilistically) deterministic answer, I believe we have to go lower level analogous to the way forces of molecules constitute the movement of a billiard ball. We need to model human decision making based on the electrochemical processes in human brains. How would we do that?

Modelling human decision making

This is nothing short of subjecting the whole human population to brain scans in machines more fine-grained than those that exist today. Possibly, we would have to have a snapshot of the states of all particles in our solar system. The principle of chaos theory already breaks our efforts, because the level of detail needed in our measurements will - as far as I can see - exceed our capacity for measurement. Then we could ask ourselves whether, and to what extent, the computations involved in determining the state of the climate are reducible. Suppose that this is the case, and suppose also that chaos theory is not bothering us, can we determine the actual state of affairs of the climate by the year 2100?

Games of information

There are two reasons I think we can not do this: 1) the observer problem comes into play when trying to measure the state of all particles in the human brain at a particular point in time (let alone that it is a hard sell to get all humans to consent to being fully modelled by a machine), and 2) the predictions of the model will have an unpredictable effect on the decision making of humans. The only situation in which the second case does not apply, is when the prediction is a self-fulfilling prophecy. But if it is, why bother to model the evolution of the climate in the first place? We are entering the realm of paradoxes here.

I think there are two related principles from the Active Inference framework that directly map onto these problems. The first we have already encountered: embeddedness. The second is the principle of being enactive: any agent will have to model its own effect on its environment in order to be effective.

In my head, enactive is the information-theoretical equivalent of time-space curvature. Masses and interactions between them make up the fabric of the universe; intelligences are large masses that attract and alter the course of other intelligences and informational objects. (My personal suspicion is that there exists an information-theoretical description of reality that explains and generalises physical theories of gravitation, though I would be hard-pressed to formulate that description myself). Familiar situations that prove this point are: success attracts success, large financial asset managers need to be careful with asset purchases and sales because their actions affect the market, platforms have network effects because of the value of the information that they have already managed to attract others.

Complexity

Games of information are very common in real life. Others that are more adversarial in nature include: intelligence agencies outwitting each other in hiding their real plan of action, trying to infer the recent internal developments of a competitor in the market, negotiating a partnership deal, or the strategies of your opponent in a game of Go. In all these situations, we are dealing with intelligences that are modelling each other's perceptions, intentions, and behaviours, and modelling the other modelling theirs, and.. This is an infinite regress in which we encounter the Halting problem. Here, inaccessibility of information is a property of the system.

I would argue that complexity is an epistemological perspective on reality that describes these kinds of situations. They are characterised by dynamics that: 1) make the situation change before you are able to analyse it in its entirety (special case of computational irreducibility), 2) prevent you from acquiring crucial information without affecting the state of the system (observer problem), 3) make predictions have an unpredictable effect on the outcome (reflexivity). This feels like meeting Gödel's incompleteness theorem in the streets of Vienna.

The quantum-reality of consciousness

The concepts discussed so far were intended to make you sceptical of the idea that any human can be fully 'reduced' or 'out-predicted' by Artificial Super Intelligence. Now, what if consciousness is a quantum process? This is the speculative view that Roger Penrose took in the 1990s. Federico Faggin now takes this one step further, saying that "consciousness is a quantum system in a quantum state experiencing its own state" - a self-referential loop. Quantum states are fundamentally private and can not be copied nor shared. If this is true, then this effectively provides a basis for free will. Quantum fields are the fundamental unit of consciousness.

Where does this leave us?

Will Artificial Super Intelligence be able to predict your every move? There are several reasons to doubt that: 1) chaos theory means that the level of detail needed to predict your next action might not be attainable, 2) processes in the brain are quite possibly computationally irreducible, and 3) active resistance on the part of humans might help them conceal crucial information. Science is moving beyond. In a new paradigm spearheaded by members of the Essentia Foundation and others, the quantum nature of consciousness provides a fundamental basis for the existence of free will.

I would personally not bet on any of these reasons though. The fate of Taleb's Turkey should warn us. We have already seen the addictive, predictive, and manipulative power of large scale (social) media platforms because of their ability to collect relevant data. Their capability is vastly increasing with the development of generative AI. The only effective route to curb the power of AI-driven platforms is through privacy, regulation, and a continued commitment to science and democracy.