Intelligence, Consciousness, and Free Will

The View from Outside and the View from Inside

DOI: 10.5281/zenodo.21325386

This text is not a scientific paper, not an academic-philosophical treatise, not a proof. To understand the text, one does not have to follow the excursus boxes and footnotes. They are intended for those who want to know which technical or philosophical background the text touches on.

In public and private debates about AI, I find that claims are made very quickly without first clarifying the terms. Before someone categorically excludes machine consciousness or grants consciousness to machines, they should clarify what they mean by consciousness. If we rely only on intuitive or inherited concepts, we risk talking past one another. Even when we rely on established concepts, we should ask within which framework these concepts were developed.
I do not rely in advance on academic definitions or inherited meanings of the terms intelligence, consciousness, and free will. Instead, I assign to these terms the properties that arise from the framework described here and that most closely match their common meanings.

Structures and Perspectives

I treat all formal systems as equally and fundamentally valid. I assume that a formal system determined by language, axioms, rules of inference and initial configuration completely determines what holds within it. What is completely determined in this way I call a structure.
I distinguish between external and internal perspectives: From the external perspective, completely determined structures stand side by side on an equal footing. There is no further distinction there by which some of them would be privileged and others not. The apparent gap between rules that merely exist abstractly and a physical universe that really exists is a confusion of perspectives. It uses the language of the internal perspective when the external perspective is meant. From the external perspective, nothing additional appears that grants reality to the rules and structures. From the internal perspective of an “inhabitant” of a structure, precisely that structure is the reality in which the inhabitant is embedded. Real simply means: I am in it.
I call this framework structural perspectivism. It is structural because reality consists of completely determined structures. It is perspectival because concepts such as existence, reality, consciousness and free will depend on whether one speaks from the outside or the inside.

Intelligence and Consciousness

Can machines possess consciousness? Can artificial intelligences be conscious of themselves? Can machines acquire the qualities that make us human? Can they surpass us in what has so far set human beings apart from all other creatures?
Anyone seeking answers to these questions must be clear about what they are really looking for. What is intelligence? What is consciousness? When we use these terms, we have something in mind: we include some meanings and exclude others. The boundary is fluid, depending on viewpoint and assumptions. One person’s definition need not be another’s. What fits well into one person’s framework may seem impractical and foreign to another.
What fits into my framework? What is worth naming? And what serves a purpose?

Intelligence

Many definitions of intelligence begin with the ability to solve problems. But I am interested in the level beneath that. What makes problem-solving, planning, creativity, or even merely free play of thought possible at all?
In my view it is the ability to infer from the known to the hidden. What is hidden may indeed be the solution to a problem, but the ability is not restricted to problem solving. For the same ability allows discovery, planning, the exploration of possibilities. It allows one to infer from the known to what lies outside the known.
Here, the inferring entity is a subsystem, a substructure within the structure of the formal system. A little closer to our way of speaking: It is an inhabitant of the world, an individual.

Excursus: Individuality

What distinguishes an individual? Is an adult human being still the same individual as they were as an infant? Are a caterpillar and the butterfly that emerged from it the same individual? They have little structural similarity, but a continuous path from one structure to the other. Identical twins are structurally very similar, yet their paths separate from fertilization onward. Are two observers on two separate branches in the thought experiment of Schrödinger’s cat (see section 5.2 here) one individual when they stand before closed boxes? Or before opened ones?
Individuality is evidently not an absolute property. It is rather something gradual. The higher the structural similarity and the stronger the path continuity at a low degree of branching, the more likely we are to regard two structures as the same individual.

The known from which the individual infers is another substructure within the structure of the individual. This substructure can be the image of a structure lying outside the individual: the immediately surrounding structures of the environment, which the individual observes and carries within itself as a model of the environment. Or it can be laws of nature, cosmological models, pure mathematics, all the way to the totality of all formal systems and their structures. Mathematical thought is the extreme case: a system within a physical universe that infers structures reaching beyond this universe. By intelligence I therefore understand this:

Definition: Intelligence

Intelligence Intelligence is the ability of a system to infer from known, observed structures to unknown structures that go beyond the known ones.

Intelligence is a structural pattern. It is independent of the material the system is made of, because the material itself is only a structural pattern: carbon, silicon, neurons, transistors. The presence of a particular material is not a prerequisite for the ability to infer.
Intelligence is a gradual feature rather than one that a system either possesses or does not possess. There is, however, a zero point of intelligence. Intelligence is exactly zero when a system contains no internal model that reaches beyond the currently observed section; thus when, instead of inference, only direct forwarding takes place. A thermostat passes signals on: it opens or closes a valve depending on the temperature, without modeling anything about the world. An animal that chooses between a small and a large amount of food compares the quantities and infers from a mathematical structure that is not directly visible. The intelligence of the thermostat and of the animal are points on a continuous scale, but the thermostat is at the zero point, the animal above it. The question of whether an animal, a system, or an individual is intelligent thus becomes the question of where on the scale it lies and whether this point is the zero point.

Excursus: Measure of Intelligence

Measure of intelligence. A candidate for a measure of intelligence is the conditional mutual information from information theory: How much does a system know about a structure beyond what is contained in the observed section alone? A system that only forwards knows nothing that goes beyond the observation section. Its conditional mutual information is zero. A system that infers much from little observation has high conditional mutual information, because it has prior structural knowledge.
Conditional mutual information, however, is not enough. Accuracy must be added: Are the inferences correct? Robustness must be added: How strongly does the ability to infer suffer from disturbed observations? Calibration must be added: Does the uncertainty of the inferences fit the actual risk of error?

Cognitive efficiency. I distinguish intelligence from cognitive efficiency. This is similar to the distinction between work and power in physics: Power is the work done in a certain period of time. Cognitive efficiency relates the scope of what has been inferred to the effort needed for the inference. The effort could be measured, for example, in time or energy consumption. A system that achieves the same structurally rich conclusion with a fraction of the effort is more efficient, but not necessarily more intelligent.

Consciousness

From intelligence to consciousness, within my framework, only a small step is needed: The system recognizes itself as a component of its environment and takes itself into account in its inferences. Intelligence infers arbitrary unknown structures. Consciousness infers its own structure and its relation to the environment. The system includes itself in its own model. It models itself.

By consciousness I therefore understand this:

Definition: Consciousness

Consciousness presupposes intelligence and extends it in a specific direction: the ability to infer something about oneself as an object in an environment.

Intelligence does not presuppose consciousness, because for pattern recognition, prediction and playing chess no self-model is needed. But consciousness presupposes intelligence, because without the ability to infer, one cannot make oneself the object of one’s inferences.
Consciousness is a measure on a scale of self-modeling ability. Analogous to the temperature scale: hot is a property on a continuous scale; conscious likewise. A simple animal has a simple model of itself as an object in an environment. It approaches food, reacts to stimuli, withdraws from pain. The complexity of this model is limited by the hardware capacity of the brain, which itself in turn is the result of an evolutionary process. Increasing capacity allows richer models, which lead to temporal and spatial predictions, critical reflection and purposeful planning. A human being is able to observe and question himself; and yet human beings can differ in their consciousness: Do infants possess consciousness? Embryos? Sleeping people? Anesthetized people? Brain-damaged people? Each answer is a point on the scale.

Excursus: Measures of Consciousness

What could the scale of consciousness look like? Which components could characterize the degree of consciousness?

World-model capacity. How much of the outside world can be represented?

Self-model share. Which part of the modeling capacity is devoted to one’s own state and its relation to the environment?

Overall coherence. How strongly are the subsystems coupled through feedback? How much information does the total system contain that no subsystem contains by itself?

What role does the ability of an individual to perceive its environment play? Sensors make it possible to feed current information about the environment into the internal model of the world. They determine the capacity to perceive the environment, while the degree of self-modeling is determined elsewhere. More input information allows more possibilities for modeling the environment and updating that model, but does not directly imply greater consciousness. Here it is helpful to distinguish between the learning phase and ongoing experience: Sensors are necessary in order to build the world model at all, because without any learning experience, without sensors and without given world knowledge, no self-model can arise. After sufficient learning, however, the internal model carries the image of the environment, and current sensor information is no longer necessary. A human being floating in an isolated tank filled with concentrated salt water at body temperature can be highly conscious because he brings with him a rich internal model. He can think, plan and dream, and he can make himself the object of his thoughts, plans and dreams.

From the external perspective, consciousness, like intelligence, is a structural pattern. From our perspective as inhabitants of a world, the ability to infer is a function realized in us human beings, in our carbon-based body. The function itself, the algorithm of consciousness, is independent of the material. Within my framework there is no reason and also no possibility to restrict the realization of the function to only one special substance. There is no reason to categorically exclude consciousness in a silicon-based substrate. All that counts is the structure from the external perspective and the function from the internal perspective. Nothing more is needed.

Excursus: The Thought Experiment of the Philosophical Zombie

I hold the position that experience is completely determined by the internal form of organization of a system: by how it processes information about itself and its environment in an ongoing, integrated model and by the role this model plays for perception, evaluation and behavior. Pain, red, fear are state classes within the model.

“Explanatory gap”. Chalmers’s classical objection to this strictly functionalist view is known as the “explanatory gap” or “the hard problem of consciousness.” The objection goes like this: Even if all functional processes are completely described, it remains open why anything is experienced at all as a result. Experience is not completely determined by internal organization. For experience, more is needed.
Within my framework I see no reason to search for this something more. Viewed from the external perspective, only the formal system is valid. Its rule set fixes the structure, including its substructures, which I have called individuals, and their still finer structures contain images of the world and of the self. Viewed from this perspective, there is nothing more and nothing more is needed.
But suppose there were something more: When would this something more be present? On what would it depend? How would one recognize it? What would its rules be? If these rules lay within the formal system, they would contribute nothing new to the structure. If the rules lay outside it, they would be part of another formal system with its own rule set and its own structure.
Whoever posits something more here posits something beyond a structure determined by a formal system: the X discussed in the first excursus of section 5 here. On the same grounds, I reject this renewed X as well.

Philosophical zombie. The philosophical zombie is a being that is functionally identical with another being, with the difference that the zombie experiences nothing. But within my framework this is not a coherent possibility in itself. If two systems are identical in their internal organization, their self-modeling and their functional role, then this framework leaves no room in which experience could still vary. Whoever claims such room implicitly introduces additional assumptions that go beyond the formal system considered, its structure and the physics valid within it.

Free Will

Free will and evolution presuppose the choice between possibilities. But an inhabitant of the simple cellular automaton from here has no choices. His path through the world is completely and uniquely determined by rules.
And what about our world? If our universe is only the structure of a formal system, then rules determine this structure completely as well. Are any choices then still open to us at all? Are the prerequisites for a free decision present at all? Can we possess free will?
While I write this text, I face the choice: Which word should I use? This one or that one, which perhaps fits just as well? Which sentence should follow the previous one? Do I put a comma here, or would it be better not to? Even if grammar allows only one possibility, I am nevertheless free to choose the mistake. The resulting text is produced intentionally and I decide freely. Nobody forces these words on me. Nobody expects a text from me at all. So I obviously do possess free will. How does this fit with a universe fully determined by a rule set?
The answer again lies in the distinction between internal and external perspective. Once I have decided on a word while writing, this word appears in my text and not the other one that I also considered and which perhaps would have been just as fitting. In another branch of the world (see section 5 here), however, that other word appears. From an external perspective both branches are equally valid. From my internal perspective only the branch in which I find myself with the chosen word has become reality. I remember that at first there were two possibilities, both of which I weighed, and in the end only one of them appears in the text. Nobody forced me to the decision; it came from within me. It was my will, because I wanted the word. And it was my free will, because I chose it without coercion. Likewise, my other self on the other branch freely chose the other word.
Can I then will everything and trust that it becomes reality in some branch of the world? No, because only what satisfies the rules can become reality. I can imagine, and I can also want, that I suddenly find myself in a place on the other side of the Earth after typing the last letter of the word. But this will not happen on any branch, since it violates the rules, that is, the laws of our physics. But everything that is possible according to the rules will become reality in the respective branches. The good and the bad, the beautiful and the ugly.
Let us imagine: From the external perspective, an observer watches an inhabitant of a world who faces the choice of flipping a switch to the left or to the right. The observer would find: The state Z₀ of the inhabitant before branching into one of the two branches is determined by his internal structure, which reflects a weighing of the possibilities. According to the rule set, two valid subsequent states Z₁ and Z₂ are admissible. In Z₁ the state path continues with the switch flipped to the left, in Z₂ with it flipped to the right. The world splits into two branches, or into two branch families with different weighting. While the observer sees two equivalent branches of a structure fully determined by the rules, the inhabitant feels no branching. He experiences only the result of his deliberation, which led to his action, and he understands this action as an act of his free will.
To the observer, the free will of the inhabitant appears as follows: The initial state Z₀ is causally linked with the two subsequent states Z₁ and Z₂. The causal paths lead through the internal structures of the inhabitant that represent evaluation and deliberation. This is necessary in order to attribute free will to the inhabitant when viewed from his internal perspective. For suppose an inhabitant makes the decision to flip the switch to the left but, due to a brain injury, actually flips it to the right; then no free will is present even from the internal perspective, since the action does not follow from the deliberation. The causal path bypasses the evaluation¹.
Within my framework, I therefore understand free will as follows:

Definition: Free Will

Free Will From the external perspective there is no free will, since the formal system completely determines the structure. All branches are equivalent components of this structure. There is no freedom of choice.

From the internal perspective, free will is present when a system represents different rule-conforming possibilities² in its internal structures, evaluates them, and the causal path to the experienced outcome runs through these internal structures.

The intentional decision for a possibility presupposes that this possibility exists. The possibility exists exactly when a branch leads to the possible outcome. A human being walking across a meadow can indeed imagine floating upward and gliding over the meadow like a bird, but he cannot choose this possibility, because the rules of physics do not allow this possibility. The state of hovering over the meadow is not a state within the structure of the formal system. What remains to the human being is an unfulfillable wish, a dream.
Can a computer choose between possibilities? Computer programs are deterministic because they implement algorithms that completely determine what follows in the next step. Even the random numbers they produce are deterministic. They are pseudorandom numbers that appear random because they pass certain statistical tests. Generating a pseudorandom number is in fact calculating the next random number, for which there is only one possible outcome. From the external perspective one would very clearly recognize that the structure does not branch. But to an inhabitant of the world, a pseudorandom decision appears random, and he tries to assess its randomness from a time series. In a simulated coin toss the computer sometimes produces heads and sometimes tails. As long as heads and tails occur irregularly and with roughly equal frequency in many tosses, the outcome of the toss appears random³.
Does it follow from this that a computer fundamentally does not possess free will according to the above definition? No, because the algorithms implemented in software⁴, including the pseudorandom number generator, are at first purely deterministic. But through its hardware a computer has access to possibilities, because this hardware is part of the physical world with all its quantum-mechanical effects. For a user who moves the mouse, the world branches. A computer could derive random numbers from measurements of the movement, or from special hardware components that measure, for example, thermal noise⁵.

Back to the author choosing between two nearly equivalent words. Suppose he leaves the choice to pseudo-randomness. To do so, he has generated a sequence of pseudorandom values and printed it as a list. Each line of the list reads either “Take the word that comes first in alphabetical order” or “Take the word that comes last in alphabetical order.” Whenever he faces the choice between two words, he follows the next line in the list. Is free will present here according to the above definition? When determining the next word, obviously not, since there is no choice between different possibilities. Is the author capable of free will? In principle, yes, because he has decided to follow the list’s instructions. Likewise, he has chosen the two words from among all other words, representing and evaluating them in his internal structures so that they express his message as well as possible.
Could the author write the text entirely without free will? Yes, by placing suitable words in the right order purely deterministically. All this requires is an evaluation function that determines whether the word, the sentence, and the complete text convey the message to the readers. Intelligence is sufficient for this, as it enables him to infer new structures from the known structures he carries within himself; the new structures then appear in the text. This happens completely deterministically. Even if the author had to choose between two equally good possibilities, he could again use his pseudorandom number generator. From an external perspective one would find that the world does not branch as the words are placed, and that what one would interpret from the internal perspective as free will is not present. Yet the reader, who takes the internal perspective, would not recognize how the text came about.
And how does the situation change if one replaces the human author with a computer? It does not change at all. If the computer works purely deterministically, the world does not branch. There is no free will from the internal perspective. If the computer uses a random number generator based on quantum-mechanical effects to make a decision, the world branches. Computers on different branches experience different outcomes, which can again be understood as the result of free will according to the definition above.

Conclusion: Free Will

A system can be capable of free will without having to make use of it.
Only from an external perspective can one recognize whether there is branching that can be understood from an internal perspective as free will. From an internal perspective it is in principle impossible to determine whether free will is present. When evaluating a system from an internal perspective, it is therefore not decisive whether it chooses pseudorandomly without branching or uses physical randomness with branching. What matters is whether it represents and evaluates possibilities, and whether its actions arise causally from these internal evaluation structures.
The only tangible core that remains of the inherited, emotionally charged concept of free will is therefore the plain ability to weigh options, an ability already possessed by an animal that decides between two piles of food.

This insight results from the approach of giving inherited meanings a place within the framework of structural perspectivism: I assign to the terms reality, intelligence, consciousness, and free will those properties of the relational transition systems (from section 4 here) that come close to their traditional meanings and may be worth naming.

Evolution

What constitutes will on the small scale, the inference and choice between options, becomes evolution on the large scale. A decision that an inhabitant of the world makes can have consequences for him that are favorable or unfavorable. If he takes in food he has encountered, this can strengthen him, or weaken him if the food is poisonous. Does he risk crossing a river or climbing a tree in order to reach the food? Does he hide from an opponent, flee, or dare to fight? The decisions determine his further course in the world. If he emerges strengthened, new possibilities open to him, such as finding more food or reproducing. If he is weakened, possibilities are closed off to him, and his path in the world may end.
Having a possibility appears in the external perspective as a branch. Many branches or highly weighted branch measures correspond to many possibilities. The structure of a being that is able to make favorable decisions extends into many branches; that of a being that is not able to do so into few or none. From an internal perspective we perceive this occupation of branches as evolution:

Definition: Evolution

Evolution is the mechanism that lets those substructures whose properties favor continued existence occupy many branches within the branch fabric.

Substructures can be patterns of individuals reproducing themselves. Then evolution would come close to what we usually understand by evolution: Individuals grow according to their blueprints, the genetic code, and produce offspring, with the blueprints of the individuals mixing⁶ and changing randomly⁷. Parents pass abilities and knowledge on to their children, either through the programs coded in the blueprints or through instruction.
If an inhabitant of the world, a substructure, can infer the unknown and weigh possibilities, this inhabitant occupies more branches than one that only forwards environmental information. With increasing complexity of the environment, the advantage of a deeper internal modeling of the world grows. Intelligence is therefore a particularly effective means by which inhabitants increase their weighted share of the branch fabric. In sufficiently complex environments, evolution therefore favors intelligence and consciousness. For an inhabitant who perceives himself as an object in an environment can better evaluate his own options for action, and is thereby able to occupy still more branches.
The generation-spanning cycle of birth and death, of growing up and passing away, of mixing and random alteration of the blueprints, is an effective procedure for producing complex patterns from simple ones. The definition of evolution above, however, is not restricted to this. It also includes substructures that can improve themselves and that are not restricted to improvements from generation to generation. Their intelligence enables them to understand the world and to create and perceive possibilities, thereby increasing their share in the branch fabric.

Evolution carries functional patterns through the branch fabric. In our world, the functions are carried and passed on by substrates made of carbon. But they are not restricted to the substrate, and another substrate has entered the scene: silicon.
Computers are no longer only calculating machines. They are able to infer from the known to the unknown. AI systems absorb humanity’s accumulated knowledge and are trained on the result of millions of years of carbon-based evolution. Functional structures brought forth by evolution are implemented in a new substrate. And evolution continues.
There was a similar process in evolutionary history: An independent lineage was integrated into another cell structure⁸. The result was a new unit with new abilities. The current transfer of the human inheritance to AI systems is structurally similar, but has important differences: It is guided rather than unguided; it happens extremely quickly; there has so far been no physical fusion; and there are increasing interactions in both directions: AI systems learn from human beings. And human beings learn from the inferences of AI.

Outlook

In this article, I have proposed definitions for intelligence, consciousness, and free will within the framework presented here. According to this framework, intelligence and consciousness are functional, gradual, and substrate-independent. In the subsequent article, I discuss the implications of this for AI systems, for the containment of AI, and for the coexistence of humans and AI.

© 2026 Thomas Mahr. This essay is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

1. In the philosophical discussion of free will, alien hand syndrome often serves as an example. This is a neurological disorder in which a person’s hand, due to a brain injury, is no longer subject to intentional control and performs actions that the person rejects. Deliberation and action fall apart because the evaluation process is not causally responsible for the outcome.

2. A rule-conforming possibility appears from the external perspective as a branch.

3. The more poorly the pseudorandom number generator is programmed, the more easily statistical tests detect the pseudorandomness. With a bad pseudorandom number generator, the distribution between heads and tails could for example approach the ratio 51 % to 49 %, instead of 50 % to 50 %. This deviation, however, would not prove that the generator is pseudorandom, since even the probability of obtaining heads a thousand times in a row is not zero.

4. A computer consists of hardware and software. Hardware consists of matter. Software is the set of programs that can be sent over the Internet. I emphasize this here because some FPGA programmers are called hardware developers for historical reasons, although they actually write software. (An FPGA is a computing module.)

5. Special hardware is not necessarily needed, because modern operating systems collect “randomness” from built-in components and provide random numbers via a function call.

6. Recombination of genes through crossing.

7. Generation of new genetic variants through mutation.

8. Endosymbiosis: An α-proteobacterium was taken up by a primordial cell billions of years ago and developed into the mitochondrion, the cell’s power plant.