Definitions:
Kevin Mitchell came out with a book called "Free Agents: How Evolution Gave us Free Will". After reading his book, I decided to list and critique his main positions.
THE ROBOT TOUCHSTONE
One of my main tests for free will is to see if the described decision making process could be performed by a robot. If it can, then this is not an argument for free will, but rather an argument for robotic will (i.e. determined will). Free will must transcend what a robot can do. This is the fundamental debate - are we robots or are we something more special than robots.
Compatibilist attempts to try to define free will as robotic are just missing the point of the entire debate. There is actually no point to defining "free" as "robotic". We can very much do without the adjective "free" and use more accurate adjectives instead.
CLASSIFYING KEVIN’S FREE WILL
Is he a Compatibilist or a Libertarian? Compatibilists believe that determinism is compatible with free will. Kevin situates himself against this position, going to extensive lengths to attack “predeterminism”. Libertarianism usually invokes indeterminism as a source of free will, but usually sources this indeterminism in something with organized causal powers, like a soul as opposed to appealing to randomness. Kevin also attacks this libertarian position by rejecting the existence of souls. Instead, Kevin mixes the two positions into a hybrid stance that could be reduced to “free will is compatible with a mixture of determinism and randomness”. His arguments rely heavily on randomness’s ability to free us from deterministic slavery.
TABLE OF CONTENTS
AGENCY
Bacteria can be agents. [2]
Agree!
Agents - [2]
Have real autonomy
Smuggling in “free will” in with the word “real”. “Have autonomy” would suffice without the unnecessary magical adjective.
Organized
Something deterministic robots can do
Dynamic patterns of activity
Something deterministic robots can do
Causally insulated
Nothing is completely causally insulated, just relatively.
Configured to maintain themselves
Something deterministic robots can do
Ability to learn (memory)
Something deterministic robots can do
Act on the world
Something deterministic robots can do
Goal-directed
Something deterministic robots can do
***I think mere goal-directedness is sufficient for the definition of an agent, but all of these additional items are fine additions to show the expanding robustness of an agent over evolutionary time.***
"Agents are loci of causation.” [3]
Metaphorically – yes. Physically – no.
Agency means behavior that is not completely determined by any given physical stimulus (not responding the same way to the stimulus every time)." [3]
No, agency means having the capacity for goal-directed action. Agency is compatible with predictably consistent responses to stimuli (such as pain avoidance). Free will is incompatible with “behavior being completely determined by any given stimulus”, as this complete determination precludes freedom (conflating agency with free will).
Agent causation transcends physical causation in that it is based on information not physical stimulus. [4]
Something deterministic robots can do
Agency includes the ability to learn [7]
Something deterministic robots can do
FEATURES
Perceptual systems must subtract out agentic biases from perception [11]
Something deterministic robots can do
Agent must include a model of itself in its model of the world [12]
Something deterministic robots can do
MEANING / REASONS
Meaning evolved from approach/avoid calculations (pleasure/pain) [1]
Something deterministic robots can do.
While it is unlikely that they can feel pleasure and pain, they can definitely calculate approach/avoidance in their value systems.
Cognition evolves to represent information in terms of meaning, not directly acting on perception, but instead acting upon meaning extracted from perception." [2]
Something deterministic robots can do
Robots incorporating AI algorithms can extract meaning and act upon it.
Learning and memory allow agents to act on their own reasons [8]
Something deterministic robots can do
GPT is very adept at learning from the user and adapting to them.
Hierarchy of neural processing is extracting meaning from perceptual data [10]
Something deterministic robots can do
Exactly what AI art is doing
Assertion: Reasons exist at the level of the whole organism, not its parts. [38]
Unjustified and unclear assertion. I would agree that reasons can exist for the whole organism. But I don’t think reasons exist ontologically across the whole organism.
CONFIGURATION
The stimulus may be the ultimate cause, but the configuration of the organism (that affords an interpretation) is the more proximate cause. [5]
Something deterministic robots can do
Free will means being free from external ultimate causes (like a stimulus). Kevin seems to grant that the will is not free from these ultimate causes.
Configuration Argument for Free Will [26]
P1) Structural organization can do causal work.
P2) We are our structural organization.
C1) We can do causal work.
P3) Doing causal work implies that the doer has free will.
C2) We have free will.
P1 is fine; structure does play a role in chains of cause and effect.
P2 is fine, our “self” can be considered the scope of reality related to our structural organization.
C1 is fine.
P3 is wrong causing C2 to fail. “Doing causal work” is just “being a cause”. Being a cause does not mean you are free from chains of cause and effect (i.e. free will). Rivers “do causal work” by virtue of their “structural organization”, yet they absolutely don’t have free will.
NEEDS, VALUES, GOALS, OPTIONS, SELECTION
The list of current needs to evaluate potential actions against is stored in the hypothalamus. [13]
Something deterministic robots can do
Hypothalamus signals intensity of needs, but not how to satisfy them. [14]
Something deterministic robots can do
Goals reside in the cortex (long-term goals in the prefrontal cortex). [15]
Something deterministic robots can do
Agents internally simulate possible actions, weigh the simulated outcomes against your goals, and select the action that should provide the highest utility. [16]
Something deterministic robots can do
All agentic actions are inhibited by default, and then a chosen action is allowed by overpowering the inhibition system. [17]
Something deterministic robots can do
Agency is located in nested circuit loops between the cortex, basal ganglia, and thalamus. [18]
Something deterministic robots can do
Agency involves assessing the situation, monitoring needs, prioritizing goals, conceive of possible actions, selecting an action, and learning from outcomes. [19]
Something deterministic robots can do
Options would be represented as patterns of neural activity in the cortex, basal ganglia, thalamus, and midbrain. [29]
Something deterministic robots can do
ALGORITHMS
We are not “step-by-step” algorithms [20]
True. We are complex algorithms not simple algorithms. Complexity does not afford free will.
We are neither spirit nor mechanical machine. Our self is a dynamic system, and that dynamic system is making decisions. [21]
True. But dynamic systems (i.e. complex algorithms) are not free from deterministic chains of causation. AI models are also dynamic systems that make decisions.
NECESSITY, RANDOMNESS, & WILL
Necessity isn’t everything. Some things are accidental (i.e. random). Some things are based on will. (quoting Epicurus) [27]
It would seem that determinism and indeterminism (logical opposites) would contain all of reality. If necessity is based on determinism, and accidents are based on indeterminism, most of the gamut is captured in that dichotomy. Will must exist within that paradigm. If that paradigm doesn’t afford the self freedom from cause and effect, then there is no free will.
“It would be better to follow the myth about the gods than to be a slave to the [destiny] of the physicists." (quoting Epicurus) [27]
Facts don’t care about your feelings. Pretending you are not a slave to the laws of physics doesn’t give you freedom from the laws of physics. The boy who wants to believe he can fly will only get himself hurt with his false beliefs.
William James’s free will is based on random creativity presenting the agent with options, and agentic selection producing the choice. [29]
Randomness does not afford free will because you can't apply willpower to that which is random. Selections are based on deterministic algorithms.
Randomness in the brain’s decision algorithm breaks the bonds of determinism. As James put it, "Our thoughts come to us freely. Our actions go from us willfully." [30]
Where does the randomness come from? Is there no chain of cause and effect that leads to the affordance of randomness? Did you control your randomness? You can’t control your randomness by definition. Hence, that randomness cannot be under the control of your will. Or else, why weren’t you free to choose to be more random as opposed to less random? How will you process that unfree randomness? The algorithmic process is not free.
Randomness is helpful for exploratory learning. [31]
I would agree with that!
Randomness Free Will of the Gaps <fallacy>: There is a gap of randomness in the brain that we don’t understand, so we will assert that free will is to be found in this randomness! ***Summary of Kevin’s way of thinking*** [32]
PREDETERMINISM
Predeterminism and reductionism preclude agency. (Predeterminism - the idea that only one possible timeline exists). [22]
False. The minimum standard for agency is the ability to act (such that you move towards a goal). Robots can act in a predeterministic way and still satisfy the definition of having agency.
Predeterminism is absurd. (Predeterminism - the idea that only one possible timeline exists). [23]
Just because something seems absurd to you doesn’t mean it is false. Quantum physics seems absurd, but it is substantiated by the evidence extremely reliably.
Predeterminism precludes choice. (Predeterminism - the idea that only one possible timeline exists). [23]
False. Choice inevitably requires a single selection. The fact that a timeline has only one selection is compatible with a choice that creates one selection.
Predeterminism precludes “could have done otherwise”. (Predeterminism - the idea that only one possible timeline exists). [24]
True! If we assume predeterminism is true, this would be a great argument for why there would be no free will!
Predeterminism precludes possibilities. (Predeterminism - the idea that only one possible timeline exists). [25]
Predeterminism precludes “actual alternative possibilities”, but it doesn’t preclude “potential possibilities” (algorithms simulate potential possibilities to make decisions).
Predeterminism precludes personal existence (personhood). (Predeterminism - the idea that only one possible timeline exists). [25]
False. Predeterminism does not preclude algorithms. Algorithms are all that is needed to fulfill the existence of personhood. Algorithms can calculate decisions based on the goals of a certain scope of reality. That scope of reality that can have goals can be termed a “self”. If an algorithm makes decisions based on the interests of 37 trillion organized human cells, then the aggregate of those 37 trillion organized human cells can be termed a “biological self”. If an algorithm makes decisions based on the interests of 80 billion neurons in a brain, then the aggregate of those psychological forces can be termed a “psychological self”.
Predeterminism precludes self-control. (Predeterminism - the idea that only one possible timeline exists). [25]
False. Control comes from deterministic algorithms (like thermostats). Selfhood comes from scope of application for the algorithm (like robots). Neither self, control, nor the combination (self-control) is incompatible with predeterminism.
RANDOMNESS OF THE INTERNEURONS
"Degree of autonomy (free will) is based on the loosening of the coupling between perception and action.” [2]
Kevin wants to argue that if there is a 1 to 1 pathway between perception and action, then this is robotic no free will. Kevin then argues that our brain is not a 1 to 1 pathway, and hence we have free will. Despite a variety of pathways, those pathways are still determined by the laws of physics operating on the electricity to move in the direction of the path of least resistance.
The determinism between perception and action is loosened by the indeterminism of the middle layers of interneurons. [6]
Quantum “free will of the gaps” argument. Doesn’t make scientific sense – brain signals are too macroscopic to afford quantum indeterminism. Assumes indeterminism without proving it. Even if indeterminism was proven here, it would still not follow that this indeterminism afforded free will.
Software (and AI models) has many “middle layers”. In AI models, the middle layers afford unpredictability of output based on inputs (the so called “loosening”). Yet, the existence of “middle layers” in software by no means makes software indeterministic and full of free will.
Perception is decoupled from action by the production of meaning [9]
The word “decoupled” is a bit strong. It’s not like there is no relationship between perception and actions because of the intermediary of meaning. Robots can take in data via perception, process it, generate meaning, and make decisions. None of this implies free will. It implies algorithmic deterministic agency.
Agency provides the decoupling of perception from obligatory action, allowing for free will. [19]
Free will means freedom from deterministic chains of causation such that you are more than just a meat robot with consciousness. Decoupling perception and action does not free you from chains of causation, it just makes those chains of causation more complex. Kevin is just assuming freedom exists within the obscurity of complexity without justification.
Quantum randomness defeats predeterminism by opening up multiple possible pathways. (Predeterminism - the idea that only one possible timeline exists). [26]
True! But “defeating predeterminism” does not give you free will.
Quantum randomness makes lower-level features causally insufficient, allowing higher-level features to play a causal role. [26]
False. Under predeterminism paradigm, lower-level features (particle physics) are deterministic. With quantum randomness, lower-level features (particle physics) is a combination of determinism and indeterminism. Including quantum randomness does not magically make lower-level features causally insufficient. More precisely, causal sufficiency is the combination of both the deterministic and indeterministic elements of these lower-level features. When we incorporate all of these lower-level features, we can explain higher-level features. Lower-level randomness is not a get out of jail free card for jumping to the conclusion that higher-level features can reach down with their metaphysical hand and causally alter the quantum randomness, such that they can involve themselves at this causal level.
Indeterminism Free Will of the Gaps <fallacy>: Indeterminism proves that lower-level details do not explain everything, hence free will is in the gap where randomness gives flexibility to higher-order organization. ***Summary of Kevin’s way of thinking*** [39]
BONUS: YouTube short contrasting Kevin's stance with Robert Sapolsky's:
TOP-DOWN CAUSATION
Higher-level configurations can constrain the lower-level parts and select among patterns of those components. [26]
True! Lower-level parts respond to their context. Higher-level configurations create that context that they will respond to. The problem with this argument is that it fails to incorporate an understanding that the higher-level configurations are caused by prior lower-level configurations, meaning that higher-level configurations are not immune to chains of cause and effect.
Decision-making is based on neurological signals, and we can hack these signals to control decisions. [33]
If the control mechanisms for decisions can be hacked, that means that there is a “cause” behind the decision that can be altered by a different causal chain – admitting that decisions are not free from chains of cause and effect.
BOTTOM-UP CAUSATION
Personality traits don’t explain everything. [34]
True. Personality traits are only one part of the web of causal chains that leads to a decision.
Reinforcement algorithmically controls behavior probabilities. [35]
True! Another reason to believe that decisions are not free from what came before (i.e. reinforcement).
Habit Argument against Free Will [35]
P1) Most of our behavior is habitual.
P2) Habitual behavior preempts decision making.
C1) Most of our behavior preempts decision making.
P3) If decision making is preempted, then the will is not involved.
P4) If the will is not involved, then free will is not involved.
C2) Free will is not involved in most of our behavior.
Kevin does a better job arguing against free will than he does for it!
We have some control over the configuration of our psychology, because we were involved with developing our psychology. [36]
Development is not a free process. If our current psychology is a slave to its developmental history, and the developmental history is a slave to the causes that came before it, then nothing about our current psychology is free.
We are not slaves to our psychological predispositions. [40]
We are not slaves to just our psychological predispositions. We are slaves to a variety of things; and all of those things, when added together, are causally comprehensive. We can calculate causation in biology – percentages between nature and nurture. There is no leftover percentage to give to free will.
Readiness potential is only discovered on non-consequential choices (pressing a button). [28]
I don’t think readiness potential is that important for the free will argument. If conscious choices are not caused by readiness potentials, then they must be caused by something else. Nothing in the observable universe (aside from maybe quantum physics) is free from cause and effect.
MORALITY & RESPONSIBILITY
Moral and legal responsibility need not change because we do have free will. [41]
I am not convinced by Kevin's argumentation that we have free will. I tend to think that moral and legal responsibility can largely be shifted to the concept of agency (not free will). For example, I think it is perfectly reasonable to lock up murderous robots in jail. This is because robots can have agency, despite not having free will.
CONCLUSION
“A puppet is free as long as he loves his strings.” – Sam Harris. Selfhood includes and requires “strings”. [37]
Sam Harris astutely critiques the compatibilist position as absurdly attributing freedom to puppets. Kevin bites the bullet and agrees that we are puppets with strings.
"If free will is the capacity for conscious, rational control of our actions, then I am happy in saying we have it.” [40]
Kevin’s book has done very little by way of addressing consciousness nor rationality. To just sneak in these two complex ideas at the end of the book and just assert that our willful control has them seems lazy.
We don’t care if Kevin is happy to say we have free will. We want to see good reasons for or against such a proposition. Kevin has not given them.
Consciousness has also been proven to be a slave to chains of cause and effect. Kevin's book did not talk about neural stimulation much, but there is extensive evidence that almost any type of conscious experience can be causally produced by neural stimulation of some sort. This means that we can use artificial means to cause consciousness. Consciousness is not the ultimate cause – the things that caused consciousness are the cause.
Rationality is just logic and logic is based in algorithm. Rationality does not give you free will.
Free will is not “conscious rational control”. Free will is a decision-making process whereby the self is free to do what it wants without being controlled by chains of cause and effect. Conscious rational control is something that meat robots can do (meat gives you the consciousness; algorithms give you the rationality). If a robot can do it, it's not free will.
REFERENCES (Quotes from the book):
1 A new trick was invented: action, the ability to move or affect things out in the environment. Information became a valuable commodity, and mechanisms evolved to gather it from the environment. With that came the crude beginnings of value and meaning. Movement toward or away from various things out in the world became good or bad for the persistence of the organism. These responses were selected for and became wired into the biochemical circuitry of simple creatures.
2 Even these humble unicellular creatures thus have real autonomy and agency, as organized dynamic patterns of activity, causally insulated from their environment, and configured to maintain themselves through time. It is not merely that they hold themselves apart from the world outside: they act in the world and on the world in a goal-directed manner. They are causal agents in their own right. As evolution proceeds, the degree of autonomy increases—at least along some lineages, like the ones leading to humans. The tight coupling of perception and action is loosened. With the advent of multicellularity and especially the invention of nervous systems, additional layers of processing emerge. Organisms evolve the means to represent sensory information internally without directly acting on it. More sophisticated control systems emerge for guiding action over longer timeframes. Organisms develop internal systems of evaluation that free them from the brutal, life-or-death judgment of natural selection. Crucially, all these systems are informational. Meaning becomes the currency of cognition.
3 "Is This Action without an Actor?
Perhaps you think I’ve gone too far. Is it really correct to say that these simple organisms are selecting an action? Is the organism—as an entity—really doing something? Or is it just that it contains some mechanisms that, when triggered by some external stimulus, result in the movement of the whole structure? Is the organism just a complicated machine, driven by physical forces impinging on it and playing out within it, or is a different kind of causation at play?
First, let’s ask ourselves what would justify our characterizing any of these single-celled organisms as an agent—a locus of causation—in the scenarios outlined earlier. A primary condition might be that its behavior is not completely determined by any given physical stimulus. If every time it encountered some object, it always behaved in exactly the same way, we would be correct in perceiving a simple stimulus-response machine at work. There would be little justification for granting any causal responsibility to the organism as a whole in determining the outcome.
As it happens, there is substantial variability in the behavioral outcomes arising from any given stimulus. This is primarily caused by the nature of the machinery, which is made from wetware, not hardware. All those tiny components are jittering about in the cell all the time, binding other molecules and letting them go, with the precise number of protein molecules fluctuating as some degrade and new ones are produced. The cell is never the same from moment to moment: it is a complex dynamic system, and its response to even the exact same stimulus may be affected by the precise state it happens to be in at the time."
4 "Although their behaviors appear simple from the outside, these single-celled creatures are thus far from being passive stimulus-response machines. Their response to a given signal depends on what other signals are around and on the cell's internal state at the time. These organisms infer what is out in the world, where it is, and how it is changing. They process this information in the context of their own internal state and recent experience, and they actively make holistic decisions to adapt their internal dynamics and select appropriate actions.
This represents a wholly different type of causation from anything seen before in the universe. The behavior of the organism is not purely driven or determined by the playing out of physical forces acting on it or in it. Clearly, a physical mechanism underpins the behavior, which explains how the system works. But thinking of what it is doing—and why it is doing it—in terms of the resolution of instantaneous physical forces is simply the wrong framing. The causation is not physical in that sense—it is informational."
5 "The stimulus may be a trigger, but it is the particular configuration of the organism that causes that signal to cause that behavior. And that configuration is the outcome of eons of natural selection, which has pragmatically wired reasons for doing things into the structure of the living system. Evolution packs causal potential into life: like potential energy, this causal potential can be used to do work in the sense of directing the behavior of the organism.
These simple organisms are not aware of those reasons. But it is still correct to say that the organism is doing something because it increases its chances of persistence. Or, at a finer level, that it is moving in a certain direction to get food or to escape a predator. It’s right to think of various components and subsystems as having functions. And it’s right to say the organism is acting on the basis of inferences about what is out in the world, rather than simply being triggered by external stimuli. The mechanisms are simply the means by which those goals are accomplished."
6 "associated with its food sources, and some aversive, including various toxic chemicals or signals of unfavorable conditions. Accordingly, they approach or avoid such molecules, using strategies similar to bacteria to move up or down a perceived concentration gradient.
As mentioned, the valence of these signals—whether they are positive or negative from the point of view of the animal—is innately configured into their neural circuitry. Sensory neurons that detect a variety of toxic or noxious chemicals (or aversive stimuli like a bump on the nose) are wired to command neurons that cause the animal to reverse direction and go somewhere else. Conversely, neurons that detect chemicals associated with food or potential mates are wired to command neurons that promote approach. Note that the meaning and value of these signals, for the worm, are still pragmatic, tied up with its response and not yet independently represented. It is not that the worm moves away from something because it senses that it’s aversive; it’s aversive because the animal’s ancestors moved away from it, and natural selection said that was a good idea.
But the coupling between perception and action is at least loosened a bit. There are now some intermediate stages of processing—carried out by the middle layers of interneurons—during which multiple signals are integrated to allow the animal to respond to the situation as a whole, as opposed to independent stimuli. Specific interneurons collect signals from multiple sensory neurons responding to diverse aversive stimuli, while other interneurons sum the activity of a different set of sensory neurons responsive to diverse attractive stimuli. The relative activity of these interneurons is then itself integrated at another stage to determine whether the sum of attraction outweighs the sum of aversion. All of this is dependent on the context: responses to those integrated external sensory signals differ depending on the current internal state of the animal."
7 "As we proceed in our evolutionary journey, we will see that these same signals are used in an even more sophisticated way: to evaluate the outcomes of actions. When an animal performs a certain action in a given situation, these signals about the resulting change in internal state will feed back to the decision-making systems and either reinforce or inhibit the choice of that action when the animal finds itself in a similar situation again. This goes beyond associative learning (“this external signal is predictive of conditions that are good for me”) to reinforcement learning, giving feedback about the agent’s own behavior (“in this context this action had a good outcome—if I encounter that situation again, I should do that again”).
C. elegans, which have the ability to develop knowledge about their environment based on their own experience, highlight an increase in the degree of agency over what we encountered so far. We saw that simple unicellular creatures are biochemically configured to behave in"
8 "These organisms have a repertoire of possible actions and choose between them for reasons.
But it could be argued that they are natural selection’s reasons, not those of the individual organisms themselves. They come pre-wired, thanks to the life-or-death feedback of natural selection across preceding generations. What we see in C. elegans is a major step beyond that. Individual worms can learn from their own experience and develop their own reasons for choosing one action over another in any given situation. An individual worm is no longer just an instance of an evolutionary lineage—a preprogrammed drone rolling off the factory conveyor belt. It goes out into the world and develops its own agency, through the history of its own actions and its own experiences.
Summary
Life got big and it got complicated. The acquisition of mitochondria gave new eukaryotic life-forms the luxury to increase in complexity. This opened the door to multicellularity, with a division of labor among diversified cell types. Coordinating this new type of body required muscles and neurons to control them. Neurons proved to be the perfect vehicles to link sensory information to action, and the evolution of a hierarchical architecture enabled greater integration of internal and external signals to guide behavior; in turn, this enabled the colonization of more complex and changeable environments. Finally, the emergence of associative learning and long-lasting memory let individuals transcend their pre-wired instincts and be able to make decisions based on their own reasons. The next step was to give them more to reason about."
9 "The Perceiving Self
It's worth pausing here to consider how far we have come from simple systems that directly, or at least proximally, couple particular sensory signals to particular actions (like coupling detection of a shadow to taking evasive action). All the visual information processing discussed here is decoupled from action. There are now levels and levels of internal processing in which information is being processed, parsed, and transformed from each cortical area to the next. The outcome of all this processing will, of course, eventually be used to inform action, but not until the organism knows what the incoming signals mean."
10 When configured in this way, perceptual systems are not just processing information—they are extracting meaning. The patterns of neural activity across different areas in the visual hierarchy represent the system's best guesses of what is out in the world, focused on what is most relevant and important for the survival of the organism. Those guesses are not merely passively computed through successive levels of information processing. The organism is actively, subjectively interpreting this information, bringing its prior experience and expectations to bear. Indeed, it also brings the prior experience of all its ancestors to bear through the evolutionarily selected genetic preconfiguration of
11 "I Move; Therefore I Am
There is one final, important implication of this predictive, interactive, inferential view of perception. The goal of all this perceptual work is to build up an internal model of what is out in the world—especially to track what is moving, how fast, and in which directions—so the organism can anticipate things and not just react to them. But one of the things that may be moving is the organism itself. This creates an opportunity—to make inferences about the world by active exploration. But it also creates a problem: How can you tell whether changes in the visual image impinging on the retina are due to something out in the world moving or to you moving? If you’ve ever sat on a train at a station and had an adjacent train start to change position relative to you, you may have had the disorienting experience of momentarily not knowing whether you were the one moving or not. That kind of confusion would not do at all for an organism trying to track prey or avoid predators.
There are a number of solutions to this problem. The first is to incorporate expectations of how the world should change depending on how you are moving into the interpretation of the visual image. For example, if you are walking or running forward, everything in the visual scene should loom toward you, with closer things looming at a faster rate. Anything that is moving differently from that overall pattern is probably doing so under its own devices.
A more sophisticated approach is for the action system to send a message—a kind of internal copy of the action command—to the visual system every time an action is intended. The expected sensory consequences of this action are then actively subtracted from even the low-level sensory areas. This happens all the time and explains why when you move your head or shift your gaze you do not perceive the whole world as moving, even though the image on your retina shifts massively. Indeed, we make little subconscious eye movements—called saccades—several times a second, without perceiving movement in the world. This"
12 "is possible because the visual system is forewarned about the impending movement and is factoring it into its predictions. By contrast, if you use your finger to (gently!) press your eyeball from the side to make it move a little, you will see the world shift in a disconcerting fashion. You might think that we could predict that too, but clearly our visual systems are not prepared to deal with the consequences of such an unusual action.
The upshot of having to distinguish self-caused from non-self-caused movements in this way is that the organism must include itself in its model of the world. The only way to productively make sense of things is for the organism to infer its own existence. This has been proposed by Fred Keijzer, Peter Godfrey-Smith, and others as the origin of subjectivity: having not only a point of view but also the experience of being a self, experiencing the world. Here, we can start to see the first glimmers of self-awareness that will ultimately be so important in understanding free will in humans, as we explore in later chapters.
But our next step is to see how all this subjectively extracted meaning is integrated with systems of planning and evaluation to inform action selection in increasingly sophisticated ways across evolution."
13 "These are the most immediate and proximal needs of the animal, and they powerfully influence behavior. But other imperatives are equally important over longer timeframes. The need to reproduce, for example, and the drive to engage in mating and parenting behaviors are also controlled by circuits in the hypothalamus. In social animals like humans, social behaviors—staying with a herd or family group, maintaining social relationships, achieving more dominant status, and so on—are just as essential to both survival and reproduction. More complex emotions like loneliness, jealousy, and anger can thus also be strong drivers of behavior in many species. And in many situations, the right thing to do is to wait for or to actively seek more information, and that can become a goal as well.
The overall profile of current needs is conveyed by the hypothalamus and connected regions to the parts of the brain involved in selecting actions."
14 "..this information is also conveyed to the cortex and other parts of the forebrain. These signals help determine what type of behavior the animal should engage in. But they don't direct the specific actions that should be taken to mediate that behavior—those are context specific.
For example, if an animal has low nutrient status, it will send a hunger signal to the decision-making regions. How this is acted on depends on the situation. If there is food available, the animal should eat it. If there is no food, the animal should seek it. Doing so may involve all kinds of behaviors—crying, begging, trading, stealing, grazing, foraging, scavenging, hunting—and each of those activities involves a variety of possible actions. The same principles apply for other behaviors, which can be divided into consummatory or appetitive components: if you've got what you need, then complete the action (feeding, drinking, mating, sleeping); if you don't have what you need (food, water, a willing member of the opposite sex, a safe place to sleep), then seek it out. Essentially, the animal makes a decision to exploit or explore, depending on what is currently directly available. The signals from the hypothalamus convey the relative urgency of these needs—broadly speaking, what to do—but not how to do it."
15 Goals are thought to be represented in the cortex, likely in a hierarchical fashion proceeding from immediate, short-term goals (or action plans), which may be represented in the motor cortex, to longer-term goals that rely on more frontal regions, known as the premotor and prefrontal cortex. We will return to the prefrontal cortex in later chapters.
16 "You may have noticed that some of the possible actions on that list, like screeching or eating dirt or sticking a pebble up its nose, are things that infants—monkeys and humans alike—do actually engage in. They explore all kinds of options, without much apparent discrimination, but they learn over time which ones tend to pay off and which ones do not.
In other words, animals develop, through experience, habits of thought. We will consider such habits in more detail when we discuss free will in humans in later chapters—especially the notion that if such ideas just spring to mind, then we are not really in control of them. But for the moment, the important thing is that this shortcut dramatically narrows the search space of possible actions.
The next step is to choose among those actions. In theory, an animal could just try different ones out and see what happens. The trouble with that trial-and-error approach in the real world is that ""error"" often implies death. At the very least, in a competitive world, it means losing out to rivals on possible gains and opportunities by spending time and effort on actions with low payouts. A far better strategy is to internally simulate a set of possible actions and predict and evaluate their likely outcomes. These actions can then be weighed against each other to enable selection of the one with the highest utility relative to the organism’s current goals. There is an obvious benefit to evaluating potential actions offline without actually engaging with the world, providing it can be done with enough efficiency that the organism is not caught dithering.
These processes are similar to those we engage in when we play a game of chess. At any point in the game, there may be ten or more pieces that we could move. When selecting among these options, we go through in our minds what would happen if we moved the queen to this square or that square or if we moved the bishop or the knight or this pawn or that piece. We simulate in our minds the consequences of each possible move and evaluate whether they are good or bad."
17 "... simulation allows us to ""let our hypotheses die in our stead."" And, of course, the more experience we gain—in life, as in chess—the better we are at making these decisions, recognizing situations, narrowing our search space of options, predicting outcomes, and accurately evaluating them over a longer horizon into the future.
In vertebrates, these processes of simulation, evaluation, and eventual choice are mediated by an extended set of nested circuit loops between the cortex, basal ganglia, and thalamus, with inputs from midbrain centers carrying utility signals and outputs to motor command centers in the tectum and other parts of the midbrain. The exact details of what all these elements do is a matter of active debate and research, but a broad sketch of our current understanding is as follows.
First, some action plans are conceived of in the cortex. Doing so may entail the activation of different sets of neurons, with each specific pattern corresponding to a particular action plan. At this point, these patterns represent the idea of doing something, not a commitment to it. These cortical patterns of activity are conveyed through a massive fan of parallel nerve fibers to the input region of the basal ganglia, called the striatum (so named because this fan of incoming fibers gives it a striped or striated appearance). Neurons in this region are hard to activate. By contrast, the output regions of the basal ganglia (called the GPi and SNr) contain inhibitory neurons that are active nearly all the time. They project to the midbrain motor centers and keep the brakes on all the motor command neurons in those regions. The baseline function of the basal ganglia is thus to inhibit all actions.
18 In vertebrates, these processes of simulation, evaluation, and eventual choice are mediated by an extended set of nested circuit loops between the cortex, basal ganglia, and thalamus, with inputs from midbrain centers carrying utility signals and outputs to motor command centers in the tectum and other parts of the midbrain.
19 To summarize, these extended brain systems, involving circuits within and between the cortex, hippocampus, hypothalamus, thalamus, basal ganglia, midbrain nuclei, motor command centers, and other regions, collectively mediate these diverse and integrated processes of behavioral control. These processes enable the animal to identify and assess the situation, monitor current needs and prioritize different goals, conceive of possible actions, select among them, and learn from their outcomes to inform the process in the future. These capabilities let animals further decouple perception from obligatory action, test out possibilities in their heads before risking them in the real world, and recruit their whole historical selves in the service of making optimal decisions.
20 "Neither Ghost Nor Machine
First, organisms do not passively wait for external stimuli to respond to. Their brains, when awake, are constantly cycling through possible actions, and this stream of behavior accommodates to new information and the changing environment. Second, this is not a one-way relationship from environment to organism: it is a recursive loop of mutual interaction. The activity of the organism changes the environment and the organism’s relation to it. The apparently linear chain of causation is really a loop or a series of loops—you can think of it as a spiral stretched through time. If we ignore these reciprocal effects, we are left studying only half of the overall system. Third, the processes of decision making and action selection are just that—processes: they have duration through time. They are not instantaneous transitions from one physical state of the system to the next. This point is crucial when we consider some philosophical challenges to the idea that choices can be made at all.
However, the idea of an algorithm—a series of steps being completed methodically and sequentially—is not an accurate conception of what is happening."
21 "Chapter 6
In a holistic sense, the organism’s neural circuits are not deciding—the organism is deciding. It’s not a machine computing inputs to produce outputs. It’s an integrated self deciding what to do, based on its own reasons. Those reasons are derived from the meaning of all the various kinds of information that the organism has at hand, which is grounded in its past experience and used to imagine possible futures. The process relies on physical mechanisms but it’s not correct to think it can be reduced to those mechanisms. What the system is doing should not be identified with how the system is doing it. Those mechanisms collectively comprise a self, and it’s the self that decides. If we break them apart, even conceptually, we lose sight of the thing we’re trying to explain.
Our minds are not an extra layer sitting above our physical brains, somehow directing the flow of electrical activity. The activity distributed across all those neural circuits produces or entails our mental experience (and similarly for whatever kinds of mental experience other animals have). The meaning of those patterns for the organism has causal power based on how the system is physically configured. We can thus build a holistic physical conception of agency without either reducing it or mystifying it."
22 "Chapter 7
... them physical predeterminism (the idea that only one possible timeline exists) and causal determinism (the idea that every event is necessarily caused by preceding events—usually seen as the same thing as physical predeterminism but subtly distinct). And I will add a third flavor that we will need to tackle too: biological determinism (the idea that an organism’s apparent choices are really internally necessitated by its own physical configuration: its biochemical state or nervous system wiring).
I deferred discussion of these challenges until now but tackle the first two types of determinism head-on in this chapter. These ideas are not just problems for free will in humans: they present an equally stern challenge to the question of whether any living organisms can really be said to have agency at all. We will see that the real issues stem not only from deterministic thinking but also from a reductionist viewpoint, especially the idea that all the causal influences arise at the lowest levels of reality: that of subatomic particles and fundamental physical forces.
The idea of physical predeterminism is both ancient and widespread. The Greek philosopher Democritus, who first proposed the existence of atoms—the tiniest indivisible units of matter—also argued that their trajectories through time do not diverge from predestined linear paths. Fast-forward two thousand years, and Newton’s laws of motion seemed to fully support this view. These laws of what is now known as ""classical"" mechanics are (supposedly) completely deterministic. If you know the position and momentum of every component of a system (for simple systems at least), these laws can be used to predict the next state of the system with high degrees of accuracy—enough, for example, to predict solar eclipses hundreds of years in the future.
Several hundred years later, Einstein’s work on relativity also seemed most consistent with physical predeterminism. He saw time not as distinct from space but as a fourth dimension of space-time, building up a picture of a static block universe in which all space and time points are essentially given at once. In this picture, the future is as determined as the past. We happen to occupy a certain slice of space-time at any given moment, but the trajectory of how the universe unfolds is already fixed. Note that, in this model, it is not clear what determines which moment..."
23 "Chapter 7
... everything in the universe would have been predetermined from the dawn of time, including me writing this sentence and you reading it.
In my view, this claim is absurd on its face. That is not an argument for whether it’s true or not but just an observation that this claim jars so bracingly with our actual experience of the world that we should be strongly suspicious of it. (And, indeed, we will see later that it is undermined by quantum physics itself.) But if we take it as a premise, it is clearly a problem not only for free will as variously conceived in humans but also for the more fundamental concept of agency itself. How can you say an organism is doing something—that it itself is a cause—if what’s happening is just the inevitable manifestation of the interactions of physical particles within it? That claim certainly doesn’t leave any room for the organism to choose what to do. In that kind of predeterministic scenario, there is nothing to choose between. There is only one future open, with no possibilities, no deciding or acting, no purposiveness, no mattering, no trying, no goals, or functions—it would be literally a meaningless universe."
24 "The Future is Not Written
... we can only hold someone accountable for an action if they could have done otherwise in a given situation, then this is ruled out by physical predeterminism. ""Otherwise"" doesn’t apply in such a universe because nothing other than what was going to happen ever happens.
Dennett argues instead that what is important for moral responsibility is that individuals themselves were the source of the determining causes—that they did some act (let’s call it A) because they wanted to do A. Had they wanted to do B, they would have done B. In fact, Dennett sees the question of whether they could have done otherwise if we ""re-wound the tape"" and put them back in precisely the same physical conditions as nonsensical and irrelevant. He argues that of course they could not: the physical circumstances at any instant determine the next state of the system (of the whole universe, in fact), which necessarily encompasses the action that these individuals will do. What is more important for him is the ability to do different things if the conditions were slightly different. He argues that ""the general capacity to respond flexibly ... does not at all require that one could have done otherwise in ... any particular case, but only that under some variations in the circumstances—the variations that matter—one would do otherwise."""
25 "Causal Slack
So, where does that leave agency and free will? On the face of it, it doesn’t really help. If all of physics were really deterministic, then possibilities and choices would not exist, and you would not be in control of your actions. (I think you would not exist at all, in fact.) But just adding some randomness does not obviously solve the problem. If my actions are controlled by random physical events at the level of subatomic particles in my brain, then I am no more in charge of them than if they were fully physically predetermined. That argument is perfectly valid, but it misses the wider point.
The idea is not that some decisions are determined (driven by necessity) and others are driven by chance. No, the really crucial point is that the introduction of chance undercuts necessity’s monopoly on causation. The low-level physical details and forces are..."
26 "Chapter 7
... they are not sufficient to determine how a system will evolve from state to state. This opens the door for higher-level features to have some causal influence in determining which way the physical system will evolve. This influence is exerted by establishing contextual constraints: in other words, the way the system is organized can also do some causal work. In the brain, that organization embodies knowledge, beliefs, goals, and motivations—our reasons for doing things. This means some things are driven neither by necessity nor by chance; instead, they are up to us.
As Epicurus put it, ""Necessity, introduced by some as the absolute ruler, does not exist, but some things are accidental, others depend on our arbitrary will. Necessity cannot be persuaded, but chance is unstable. It would be better to follow the myth about the gods than to be a slave to the [destiny] of the physicists.""
Epicurus’s real target was thus reductionism, not merely determinism. His view was necessary to give the will some room in which to operate. The contemporary philosopher and mathematician George Ellis similarly argues that physical indeterminacy creates causal slack in physical systems, which opens the door for what is known as ""top-down causation."" Put simply, this is the principle that the way a system behaves depends on the way it is configured, which can constrain the lower-level components and functionally select among patterns of those components. In living organisms, that configuration itself is the outcome of selection over millennia and over the lifetime of the organism itself, on a timescale of seconds to hours to years. This highlights another key principle, which also diverges from a reductive, comprehensively bottom-up view: causation is not wholly instantaneous."
27 As Epicurus put it, "Necessity, introduced by some as the absolute ruler, does not exist, but some things are accidental, others depend on our arbitrary will. Necessity cannot be persuaded, but chance is unstable. It would be better to follow the myth about the gods than to be a slave to the [destiny] of the physicists."
28 "Page 186
... the results were very clear: in the trials where the decision was arbitrary, a readiness potential was detected. That is, the movement of the left or right hand was associated with random fluctuations that raised the activity in the premotor cortex above the threshold for initiating a movement. But in the trials where the subjects made a deliberative, consequential choice, no such association was found. Presumably, the subjects were processing information elsewhere in the brain as they were considering the option, and then movement was triggered on that basis."
29 "Page 187
... known as the two-stage model of free will, proposed by the American psychologist and philosopher William James in 1884. This model incorporates a degree of indeterminism in our cognition while protecting the causal role of the agent in actually deciding what to do. In the first stage, in any given situation, some set of possible actions occur to the organism. In this process, James proposed that some degree of randomness is at play, but the randomness does not decide the outcome—the organism does. The options are presented for consideration, and the organism selects the one that is most congruent with its current goals and beliefs and that has the highest predicted utility. That is, the organism selects from the range of presented options based on its current reasons.
In terms of our current understanding of the neuroscience of action selection as discussed in chapter six, the possible actions would be represented by patterns of activity arising in cortical areas. These patterns would then be evaluated through extended interlocking circuit loops among the cortex, basal ganglia, thalamus, and midbrain. This evaluation..."
30 "Chapter 8
Figure 8.5. Two-stage model of action selection. If a situation is familiar, a habitual response may be ""suggested"" by the cortex, taken as optimal, and executed. In more novel situations, a range of options may become primed, with some degree of randomness in which ones happen to be activated (reflecting some ideas just ""occurring to you""). They are then subjected to evaluation of the expected outcomes through the action selection circuitry, with one eventually being selected and released. In both familiar and novel situations, the outcome is monitored and evaluated with respect to the goals to see whether the behavior has been successful or remains optimal. When selected actions fail to achieve goals, signals may be sent back to the cortex to request additional options. The degree of randomness in that process may itself be modulated to increase the ""search space"" and find more creative solutions.
This model thus powerfully breaks the bonds of determinism, incorporating true randomness into our cognitive processes while protecting the causal role of the agent itself in deciding what to do. As James put it, ""Our thoughts come to us freely. Our actions go from us willfully."""
31 ... learned template of the tutor's song. This takes some exploration of motor patterns, similar to babies babbling as they learn to talk: they figure out how to make different sounds through trial and error. Some randomness is valuable in this process as a means of tweaking the output and thus exploring the space of possible motor patterns. In this case, the degree of randomness is under the active control of a particular brain area, which is independent of the ones that actually do the learning. Neurons in this area fire in a highly variable way and seem to take a "copy" of the current motor pattern, add noise to it, and transmit it back to the circuits that evaluate how good a match it now produces to the template. When this area is inactivated, juvenile birds do not explore the space of motor patterns in the normal way, and their song crystallizes before it is a good match to the template.
32 ""It is thus wrong to think of behavior as simply reflecting the inevitable transition of defined physical states of the brain from one instant of time to the next. No such states exist; no such instants exist. And there is nothing inevitable about the trajectory of neural states that the brain will follow through time. It is neither predetermined nor simply a passive response to stimuli from the environment. As Epicurus said, some things are caused by necessity, some are due to chance, and some are up to us. It’s that up-to-usness that is the key element of agency. The elements of randomness at work in our brains give some leeway for us to have the final say in settling the matter."""
33 "the sense that neuroscientists have identified at least some of the circuits that seem to mediate these different signals and feed into the processes of belief formation, goal prioritization, and action selection. Indeed, new technologies in animals are allowing researchers to observe these separable functions and to correlate levels of activity in specific circuits with, say, confidence levels, assessed threat levels, or anticipated reward; scientists are even able to experimentally tweak the tuning of these parameters to alter decision making and behavior in real time."
34 """The profile of variation across all these parameters, under this model, feeds into an overall level of extraversion or neuroticism or conscientiousness. Two people might thus end up with similar scores on any of these traits for quite different underlying reasons. Because variation in all the individual parameters is heritable, the higher-level constructs are too.
So, yes, we really are all tuned a little differently. And these differences affect our behavioral tendencies in any given situation. But here’s the thing: we are never actually in “any given situation”—we’re always in some particular situation. That’s what life is, after all, one damn thing after another. What we actually do in any particular situation depends on all kinds of factors, including our evaluation of the current scenario in the context of the knowledge we built up from our own personal history, along with our current goals and motivations, the predicted utility of the outcomes of various actions, and so on. Within this complex context, our personality traits are a contributory factor, but they are far from a determining one in any moment. However, they do have an influence on the development of our characteristic adaptations and habits over time."""
35 "As discussed in chapter six, most of our behavior is habitual in nature. Reinforcement learning mechanisms evaluate the outcome of an action and increase or decrease the likelihood that you will repeat the same action the next time you are in the same or a similar situation. If a given action continues to have a positive outcome, then that reinforcement will eventually lead to it becoming habitual. This means that the processes of decision making and action selection are preempted: there's"
36 """Habits, attitudes, policies, and mature tendencies that make us all who we are. These views can be reconciled by recognizing that there will be an interplay between our innate traits and the trajectory of development of our character. Cicero was at pains to say that the choices each of us makes for how we are going to live our lives should align with our individual natures: 'Everyone has the obligation to ponder well his [sic] own specific traits of character. He must also regulate them adequately and not wonder whether someone else’s traits might suit him better. The more definitely his own a man’s character is, the better it fits him.'
The idea that our current psychology was shaped by prior causes over which we had no control is thus not accurate. Each of us has been very actively involved in the process of shaping the person we’ve become."""
37 "Becoming Ourselves - Page 247
Free will skeptic Sam Harris has argued that the compatibilist view put forward by philosophers like Daniel Dennett—where we are constrained to act according to our own reasons—amounts to saying that “a puppet is free as long as he loves his strings.” To me, this misconstrues the nature of the self. Selfhood entails constraints: they are what selves are made of. The puppet is made out of strings—if you removed them, there would be no puppet left."
38 "Chapter 12 - Page 280
And we, like other animals, have a set of neural resources designed specifically to allow us to select our actions in the service of our goals; that is, to act for our own reasons. Those reasons inhere at the level of the whole organism, not its parts. The system acts as a unified whole (at least under nonpathological circumstances). Although it relies on subsystems and the workings of its physical components, its function cannot be deconstructed or reduced to those workings. The agent as a whole is deciding what to do. In humans, who have an added layer of conscious cognitive control, this seems to meet any reasonable, realistic criteria for free will."
39 "Page 280
The counterargument—that indeterminacy or randomness doesn’t get you free will—also misses the mark. The idea is not that some events are predetermined and others are random, with neither providing agential control. It’s that a pervasive degree of indefiniteness loosens the bonds of fate and creates some room for agents to decide which way things go. The low-level details of physical systems plus the equations governing the evolution of quantum fields do not completely determine the evolution of the whole system. They are not causally comprehensive: other factors—such as constraints imposed by the higher-order organization of the system—can play a causal role in settling how things go.
"
40 "Page 282, Chapter 12
for action. We can and do reason about our reasons prior to and in the process of making a decision. We can and do exercise the capacity for conscious, rational control. Thus, although we cannot necessarily change our basal psychological predispositions, we are not slaves to them.
If free will is the capacity for conscious, rational control of our actions, then I am happy in saying we have it. Indeed, the existence of this capacity is brought home by the realization that some people have a greater capacity for this control than others. Babies, for example, are not born with this capacity, which entails a set of cognitive skills that must be learned and practiced. There is also inevitable variation across individuals in the diverse domains that underpin these cognitive faculties, contributing (along with social, cultural, and experiential factors) to the fact that some people achieve higher levels of rational control than others. In addition, these capacities can be reduced or even severely impaired in some individuals such as those suffering from compulsions, addiction, psychosis, depression, dementia, or many other forms of mental illness.
Moreover, just because we have this capacity does not mean we can always exercise it to the same extent. It can also be impaired by alcohol or drugs; by strong emotions like desire, rage, pain, jealousy, or grief; or even just by being tired or harried or distracted. Our free will is thus not some nebulous, spooky, mystical property granted to us by the gods. It is an evolved biological function that depends on the proper functioning of a distributed set of neural resources."
41 "Moral and Legal Responsibility
What are the implications of this position for our views on moral and legal responsibility? Can we still hold people responsible for their actions? Are they still deserving of praise or blame or reward or punishment? I believe that our views on these issues do not need to change. Despite the headlines proclaiming the death of free will, it remains stubbornly alive and well. Nothing in philosophy or physics or neuroscience or genetics or psychology or neurology or any other science undermines the idea that we do have the capacity for conscious, rational control of our actions."
Comments