Did you know ... Search Documentation:
Pack narsese -- jmc/freewill2.md


John McCarthy

from May 16, 2002 until November 6, 2005


A common feature of free will is that a person has choices among

alternative actions and chooses the action with the apparently most

preferred consequences. In a determinist theory, the mechanism that

makes the choice among the alternatives is determinist. The sensa-

tion of free will comes from the fact that the mechanism that gen-

erates the choices uses a non-determinist theory as a computational

device and that the stage in which the choices have been identified is

introspectable. The present formalism is based on work in artificial

intelligence (AI).

We present a theory of simple deterministic free will (SDFW) in a

deterministic world. The theory splits the mechanism that determines

action into two parts. The first part computes possible actions and

their consequences. Then the second part decides which action is most

preferable and does it.

We formalize SDFW by two sentences in situation calculus, a math-

ematical logical theory often used in AI. The situation calculus for-

malization makes the notion of free will technical. According to this

notion, almost no animal behavior exhibits free will, because exer-

cising free will involves considering the consequences of alternative

actions. A major advantage of our notion of free will is that whether

an animal does have free will may be determinable by experiment.

Some computer programs, e.g. chess programs, exhibit SDFW. Al-

most all do not. At least SDFW seems to be required for effective

chess performance and also for human-level AI.

Many features usually considered as properties of free will are omit-

ted in SDFW. That’s what makes it simple. The criterion for whether

an entity uses SDFW is not behavioristic but is expressed in terms of

the internal structure of the entity.

1 The Informal theory

Let the course of events, including events in my brain (or yours or his or its)

be deterministic. It seems to many people that there is no place for free will.

Even our thoughts are determined.

However, if we examine closely how a human brain (or chess program) de-

terministically makes decisions, free will (or imitation free will if your philos-

ophy forbids calling it real free will) must come back in. Some deterministic

processes consider alternative actions and their consequences and choose the

actions they think have the most preferred consequences. This deterministic

decision process uses a nondeterministic theory to present the set of available

actions and the consequences of each of them.

When a person, animal, or machine reacts directly to a situation rather

than comparing the consequences of alternative actions, free will is not in-

volved. So far as I can see, no animals consider the consequences of al-

ternative actions; hence they don’t have free will. Others think that apes

sometimes do compare consequences. A relevant experiment is suggested in

section 7. Using free will is too slow in many situations, and training and

practice often have the purpose of replacing comparison of consequences by

automatic reaction to a situation.

We believe this simple theory covers the most basic phenomenon of human

free will. We’ll call it simple deterministic free will and abbreviate it SDFW.

Robots with human-level intelligence will also require at least this much free

will in order to be useful.

Beyond having free will, some systems are conscious of having free will

and communicate about it. If asked to tell what it is doing, humans or some

machine will tell about their choices for action and say that they intend to de-

termine which action leads to the best consequence. Such a report, whether

given externally or contemplated internally, constitutes the human sensation

and the human report of free will. SDFW does not require consciousness of

having free will or the ability to communicate about it. That’s what’s sim-

ple about SDFW. Thinking about one’s free will requires theoretical struc-

ture beyond or above SDFW. So will considering actions as praiseworthy or

blameworthy. SDFW also doesn’t treat game theoretic situations in which

probabilistic mixed strategies are appropriate.

In AI research one must treat simple cases of phenomena, e.g.


tional behavior, because full generality is beyond the state of the art. Many

philosophers are inclined to only consider the general phenomenon, but this

limits what can be accomplished. I recommend to them the AI approach of

doing the simplest cases first.

2 Situation calculus formulas for SDFW

Artificial intelligence requires expressing this phenomenon formally, and we’ll

do it here in the mathematical logical language of situation calculus. Situ-

ation calculus is described in [MH69], [Sha97], [Rei01], and in the extended

form used here, in [McC02]. Richmond Thomason in [Tho03] compares situ-

ation calculus to theories of action in the philosophical literature. As usually

presented, situation calculus is a non-deterministic theory. The equation

s0 = Result(e, s)

asserts that s0 is the situation that results when event e occurs in the situation

s. Since there may be many different events that can occur in s, and the

theory of the function Result does not say which occurs, the theory is non-

deterministic. Some AI jargon refers to it as a theory with branching time

rather than linear time. Actions are a special case of events, but most AI

work discusses only actions.

Usually, there are some preconditions for the event to occur, and then we

have the formula

P recond(e, s) → s0 = Result(e, s).

[McC02] proposes adding a formula Occurs(e, s) to the language that can

be used to assert that the event e occurs in situation s. We have

Occurs(e, s) → (N ext(s) = Result(e, s)).

Adding occurrence axioms, which assert that certain actions occur, makes

a theory more deterministic by specifying that certain events occur in situa-

tions satisfying specified conditions. In general the theory will remain partly

non-deterministic, but if there are occurrence axioms specifying what events

occur in all situations, then the theory becomes deterministic, i.e. has linear


We can now give a situation calculus theory for SDFW illustrating the

role of a non-deterministic theory in determining what will deterministically

happen, i.e. by saying what choice a person or machine will make.

In these formulas, lower case terms denote variables and capitalized terms

denote constants. Suppose that actor has a choice of just two actions a1

and a2 that he may perform in situation s. We want to say that the

event Does(actor, a1) or Does(actor, a2) occurs in s according to which of

Result(Does(actor, a1), s) or Result(Does(actor, a2), s) actor prefers.

The formulas that assert that a person (actor) will do the action that he,

she or it thinks results in the better situation for him are

Occurs(Does(actor, Choose(actor, a1, a2, s), s), s),




Choose(actor, a1, a2, s) =

if P ref ers(actor, Result(a1, s), Result(a2, s))

then a1 else a2.

Adding (2) makes the theory determinist by specifying which choice us

Here Prefers(actor, s1, s2) is to be understood as asserting that actor


prefers s1 to s2.

Here’s a non-deterministic theory of greedy John.

Result(A1, S0) = S1,

Result(A2, S0) = S2,

W ealth(J ohn, S1) = $2.0 × 106,

W ealth(J ohn, S2) = $1.0 × 106,

(∀s s0)(W ealth(J ohn, s) > W ealth(J ohn, s0)

→ Prefers(J ohn, s, s0).

As we see, greedy John has a choice of at least two actions in situation S0

and prefers a situation in which he has greater wealth to one in which he has

lesser wealth.

From equations 1-3 we can infer

Occurs(Does(J ohn, A1, S0)).


1(2) uses a conditional expression. if p then a else b has the value a if the proposition

p is true and otherwise has the value b. The theory of conditional expressions is discussed

in [McC63]. Conditional expressions are used in the Lisp, Algol 60, Algol 68, and Scheme

programming languages.



For simplicity, we have omitted the axioms asserting that A1 and A2 are

exactly the actions available and the nonmonotonic reasoning used to derive

the conclusion.

Here Prefers(actor, s1, s2) is to be understood as asserting that actor

prefers s1 to s2.

I used just two actions to keep the formula for Choose

short. Having more actions or even making Result probabilistic or quantum

would not change the nature of SDFW. A substantial theory of Prefers is

beyond the scope of this article.

This illustrates the role of the non-deterministic theory of Result within

a deterministic theory of what occurs. (1) includes the non-deterministic of

Result used to compute which action leads to the better situation. (2) is the

deterministic part that tells which action occurs.

We make four claims.

  1. Effective AI systems, e.g. robots, will require identifying and reasoning about their choices once they get beyond what can be achieved with situation-

    action rules. Chess programs always have.

  2. The above theory captures the most basic feature of human free will.
  3. Result(a1, s) and Result(a2, s), as they are computed by the agent, are not full states of the world but elements of some theoretical space of

    approximate situations the agent uses in making its decisions. [McC00] has

    a discussion of approximate entities. Part of the problem of building human-

    level AI lies in inventing what kind of entity Result(a, s) shall be taken to


  4. Whether a human or an animal uses simple free will in a type of situation is subject to experimental investigation—as discussed in section 7.

    Formulas (1) and (2) illustrate person making a choice. They don’t say

    anything about person knowing it has choices or preferring situations in

    which it has more choices. SDFW is therefore a partial theory that requires

    extension when we need to account for these phenomena.

    3 A generalization of SDFW

    We can generalize SDFW by applying preferences to actions rather than to

    the situations resulting from actions. The formulas then become

    Occurs(Does(actor, Choose-action(actor, a1, a2, s), s), s)


    (1) and

    Choose-action(actor, a1, a2, s) =

    if P ref ers-action(actor, a1, a2, s)

    then a1 else a2.


    (5) and (6) obviously generalize (1) and (2), because the earlier case is

    obtained by writing

    P ref ers-action(a1, a2, s) ≡ P ref ers(Result(a, s), Result(a2, s)).


    I am doubtful about the generalization, because I don’t see how to repre-

    sent commonsense preferences between actions except in terms of preferring

    one resulting situation to another.

    4 Knowledge of one’s free will and wanting

    more or fewer choices

    This section is less worked out than basic SDFW and not axiomatized. That’s

    why it was best to start simple.

    Here are some examples of it being good to have more choices.

    “I’ll take my car to work today rather than bicycling so I can shop on the

    way home if I want to.”

    “If you learn mathematics, you will more choice of scientific occupations”.

    “The more money I have, the more models of car I can choose from.”

    “If I escape from Cuba, I will have more choice of what to read, what I

    can say or write, and where to travel.”

    We want to say that situation s1 is at least as free as situation s2, written

    s1 ≥f reedom s2, if every fluent achievable by a single action from s2 is achiev-

    able from s1. Just as with equation (1), we can say that person chooses an

    action that leads to more freedom at the next situation.

    s1 ≥f reedom s2

    (∀f )((∃a)(Holds(f, Result(Does(person, a), s2)))


    (∃a)(Holds(f, Result(Does(person, a), s1)))).

    Here f ranges over fluents. Having more choices is usually preferred.

    However, one sometimes wants fewer choices. Burning one’s bridges, nailing

    the flag to the mast, and promising to love until death do us part are examples

    of actions that reduce choices. The conditions under which this occurs are too

    difficult for me to formalize at present. They can involve fearing that one’s

    preferences in the future might be different from one’s present preferences

    for future actions or that making a commitment about one’s future actions

    confers a present benefit.

    5 Philosophical issues

    The formalism of this paper takes sides in several philosophical controversies.

  5. It considers determinism and free will, as experienced and observed by humans, as compatible. This is in accordance with the views of Locke

    and Hume.

  6. It takes a third person point of view, i.e. considers the free will of others and not just the free will of the observer.
  7. It breaks the phenomenon of free will into parts and considers the simplest systems first—in contrast to approaches that demand that

    all complications be understood before anything can be said. In this it

    resembles the approaches to belief and other intentional states discussed

    in [Den71], [Den78], and [McC79]. Starting with simple systems is the

    practice in AI, because only what is understood can be implemented

    in computer programs.

    It seems to me that formulas (1) and (2) expressing the use of the branch-

    ing time Result(e, s) function in determining what events occur make the

    philosophical ideas definite. Thus we can see which modifications of the

    notions are compatible with (1) and (2), and which require different axioms.

    The process of deciding what to do often involves considering a pruned

    set of actions which eliminate those that have obviously bad consequences.

    The remaining actions are those that one can do. When someone refers to

    a pruned action, one sometimes gets the reply, “Oh, I could do that, but I

    really can’t, because . . . .”

    6 Praise and blame

    We have maintained that the basic notion of free will is the same for humans,

    animals and robots. Praising or blaming agents for their actions is an ad-

    vanced notion requiring more structure, e.g. including good or bad actions or

    outcomes. Blaming or praising humans requires taking human peculiarities,

    not shared with agents in general, e.g. robots, into account.

    Consider the verdict: “Not guilty by reason of insanity” as applied to a

    person with schizophrenia. Schizophrenia is basically a disease of the chem-

    istry of the blood and nervous system. At a higher level of abstraction, it is

    regarded as a disease in which certain kinds of thoughts enter and dominate

    consciousness. A patient’s belief that the CIA has planted a radio in his brain

    is relieved by medicines that change blood chemistry. If the patient’s belief

    caused him to kill someone whom he imagined to be a CIA agent, he would

    be found not guilty by reason of insanity. If we wanted robots susceptible

    to schizophrenia, we would have to program in something like schizophrenia,

    and it would be a complicated and unmotivated undertaking—unmotivated

    by anything but the goal of imitating human schizophrenia. The older Mc-

    Nachten criterion, “unable to understand the nature and consequences of

    his acts”, uses essentially the criteria of the present article for assessing the

    presence or absence of free will.

    I don’t know if all praise or blame for robots is artificial; the matter

    requires more thought. Verbally one might praise a robot as way of getting

    it to do more of the same.

    7 A possible experiment with apes

    Here’s a gedanken experiment aimed at determining whether apes (or other

    animals) have free will in the sense of this article. The criterion is whether

    they consider the consequences of alternate actions.

    The ape can move a lever either to the left or the right. The lever causes

    a prize to be pushed off a shelf, either to the left or the right. The goody

    then hits a baffle and is deflected either to the ape in control of the lever or

    to a rival ape. On each trial, the baffle is set by the experimenter. The whole

    apparatus is visible to the ape, so it can see the consequences of each choice.

    The free will involves the ape having two choices and being able to de-

    termine the consequences of each choice.

    There is a possibility that the ape can win without determining the conse-

    quences of the possible actions. It may just learn a rule relating the position

    of the baffle and the action that will get the prize. Maybe we wouldn’t be

    able to tell whether the ape predicted the consequences or not.

    We can elaborate the experiment to obviate this difficulty. Let there be a

    sequence of (say) six baffles that are put in a randomly selected configuration

    by the experimenter or his program at each trial. Each baffle deflects the

    prize one way or the other according to how it is set. If the ape can mentally

    follow the prize as it would bounce from baffle to baffle, it will succeed.

    However, there are 64 combinations of baffle positions. If a training set of

    (say) 32 combinations permits the ape to do the remaining 32 without further

    trial and error, it would be reasonable to conclude that the ape can predict

    the effects of the successive bounces.

    I hope someone who works with apes will try this or a similar experiment.

    Frogs are simpler than apes. Suppose a frog sees two flies and can stick out

    its tongue to capture one or the other. My prejudice is that the frog doesn’t

    consider the consequences of capturing each of the two flies but reacts directly

    to its sensory inputs. My prejudice might be refuted by a physiological


    Suppose first that frogs can taste flies, i.e. when a frog has a fly in its

    mouth, an area of the frog’s brain becomes active in a way that depends

    on the kind of fly. Suppose further that when a frog sees a fly, this area

    becomes active, perhaps weakly, in the same way as when the frog has the

    fly in its mouth. We can interpret this as the frog imagining the taste of

    the fly that it sees. Now further suppose that when the frog sees two flies, it

    successively imagines their tastes and chooses one or the other in a consistent

    way depending on the taste. If all this were demonstrated, I would give up

    my prejudice that frogs don’t have SDFW.

    8 Comparison with Dennett’s ideas

    Daniel Dennett [Den03] writes about The evolution of freedom. I agree with

    him that free will is a result of evolution. It may be based on a more basic

    ability to predict something about what future will result from the occurrence

    of certain events including actions. He compares determinism and inevitabil-

    ity, and makes definitions so that in a deterministic world, not all events that

    occur are inevitable. He considers that freedom evolves in such a way as to

    make more and more events evitable, especially events that are bad for the


    Dennett’s ideas and those of this paper are in the same direction and

    somewhat overlap. I think SDFW is simpler, catches the intuitive concepts

    of freedom and free will better, and are of more potential utility in AI.

    Consider a species of animal with eyes but without a blink reflex. Every

    so often the animal will be hit in the eye and suffer an injured cornea. Now

    suppose the species evolves a blink reflex. Getting hit in the eye is now often

    evitable in Dennett’s sense. However, it is not an exercise of free will in

    my sense.2 On the other hand, deciding whether or not to go through some

    bushes where there was a danger of getting hit in the eye on the basis of

    weighing the advantages against the dangers would be an exercise of free will

    in my sense. It would also be an evitability in Dennett’s sense.

    Evitability assumes that there is a normal course of events some of which

    may be avoided, e.g. that getting hit in they eye is normal and is avoided

    by the blink reflex. My notion of free will does not involve this, because the

    choice between actions a1 and a2 is symmetric. It is interesting to ask when

    there are normal events that can sometimes be avoided.

    The converse of an evitability is an opportunity. Both depend on a dis-

    tinction between an action and non-action. In the case of non-action, nature

    takes its course.

    9 Summary and remarks

    A system operating only with situation-action rules in which an action in

    a situation is determined directly from the characteristics of the situation

    does not involve free will. Much human action and almost all animal action

    reacts directly to the present situation and does not involve anticipating the

    consequences of alternative actions.

    One of the effects of practicing an action is to remove deliberate choice

    from the computation and to respond immediately to the stimulus. This is

    often, but not always, appropriate.

    Human free will, i.e. considering the consequences of action, is surely the

    product of evolution.

    free will in his sense

    2Dennett (email of 2003 Feb 27) tells me that the blink reflex involves no significant

    Do animals, even apes, ever make decisions based on comparing antici-

    pated consequences? Almost always no. Thus when a frog sees a fly and

    flicks out its tongue to catch it, the frog is not comparing the consequences

    of catching the fly with the consequences of not catching the fly.

    One computer scientist claims that dogs (at least his dog) consider the

    consequences of alternate actions. I’ll bet the proposition can be tested, but

    I don’t yet see how.

    According to Dennett (phone conversation), some recent experiments sug-

    gest that apes sometimes consider the consequences of alternate actions. If

    so, they have free will in the sense of this article.

    If not even apes ordinarily compare consequences, maybe apes can be

    trained to do it.

    Chess programs do compare the consequences of various moves, and so

    have free will in the sense of this article. Present programs are not conscious

    of their free will, however. [McC96] discusses what consciousness computer

    programs need.

    People and chess programs carry thinking about choice beyond the first

    level. Thus “If I make this move, my opponent (or nature regarded as an

    opponent) will have the following choices, each of which will give me further

    choices.” Examining such trees of possibilities is an aspect of free will in the

    world, but the simplest form of free will in a deterministic world does not

    involve branching more than once.

    Daniel Dennett [Den78] and [Den03] argue that a system having free will

    depends on it being complex. I don’t agree, and it would be interesting to

    design the simplest possible system exhibiting deterministic free will. A pro-

    gram for tic-tac-toe is simpler than a chess program, but the usual program

    does consider choices.

    However, the number of possible tic-tac-toe positions is small enough so

    that one could make a program with the same external behavior that just

    looked up each position in a table to determine its move. Such a program

    would not have SDFW. Likewise, Ken Thompson has built chess programs

    for end games with five or fewer pieces on the board that use table lookup

    rather than look-ahead. See [Tho86]. Thus whether a system has SDFW

    depends on its structure and not just on its behavior. Beyond 5 pieces,

    direct lookup in chess is infeasible, and all present chess programs for the

    full game use look-ahead, i.e. they consider alternatives for themselves and

    their opponents. I’ll conjecture that successful chess programs must have at

    least SDFW. This is not the only matter in which quantitative considerations

    make a philosophical difference. Thus whether the translation of a text is

    indeterminate depends on the length of the text.

    Simpler systems than tic-tac-toe programs with SDFW are readily con-

    structed. The theory of greedy John formalized by (3) may be about as

    simple as possible and still involves free will.

    Essential to having any kind of free will is knowledge of one’s choices of

    action and choosing among them. In many environments, animals with at

    least SDFW are more likely to survive than those without it. This seems to

    be why human free will evolved. When and how it evolved, as with other

    questions about evolution, won’t be easy to answer.

    Gary Drescher [Dre91] contrasts situation-action laws with what he calls

    the prediction-value paradigm. His prediction-value paradigm corresponds

    approximately to the deterministic free will discussed in this article.

    I thank Drescher for showing me his forthcoming [Dre06]. His notion of

    choice system corresponds pretty well to SDFW, although it is imbedded in

    a more elaborate context.

    This article benefited from discussions with Johan van Benthem, Daniel

    Dennett, Gary Drescher, and Jon Perry. The work was partly supported by

    the Defense Advanced Research Projects Agency (DARPA).


    [Den71] Daniel C. Dennett. Intentional systems. The Journal of Philosophy,

    68(4):87–106, 1971.

    [Den78] Daniel Dennett. Brainstorms: Philosophical Essays on Mind and

    Psychology. Bradford Books/MIT Press, Cambridge, 1978.

    [Den03] Daniel Dennett. Freedom Evolves. Viking, 2003.

    [Dre91] Gary Drescher. Made up minds: a constructivist approach to arti-

    ficial intelligence. MIT Press, 1991. Q335D724.

    [Dre06] Gary Drescher. Good and real: Paradoxes from physics to ethics.

    M.I.T. Press, forthcoming, 2006.

    [McC63] John McCarthy. A Basis for a Mathematical Theory of Compu-

    tation3. In P. Braffort and D. Hirschberg, editors, Computer Pro-


    gramming and Formal Systems, pages 33–70. North-Holland, Ams-

    terdam, 1963.

    [McC79] John McCarthy. Ascribing mental qualities to machines4. In Martin

    Ringle, editor, Philosophical Perspectives in Artificial Intelligence.

    Harvester Press, 1979. Reprinted in [McC90].

    [McC90] John McCarthy. Formalizing Common Sense: Papers by John Mc-

    Carthy. Ablex Publishing Corporation, 1990.

    [McC96] John McCarthy. Making Robots Conscious of their Mental States5.

    In Stephen Muggleton, editor, Machine Intelligence 15. Oxford Uni-

    versity Press, 1996. Appeared in 2000. The web version is improved

    from that presented at Machine Intelligence 15 in 1995.

    [McC00] John McCarthy. Approximate objects and approximate theories6.

    In Anthony G. Cohn, Fausto Giunchiglia, and Bart Selman, edi-

    tors, KR2000: Principles of Knowledge Representation and Rea-

    soning,Proceedings of the Seventh International conference, pages

    519–526. Morgan-Kaufman, 2000.

    [McC02] John McCarthy. Actions and other events in situation calculus7. In

    B. Selman A.G. Cohn, F. Giunchiglia, editor, Principles of knowl-

    edge representation and reasoning: Proceedings of the eighth inter-

    national conference (KR2002). Morgan-Kaufmann, 2002.

    [MH69] John McCarthy and Patrick J. Hayes. Some Philosophical Prob-

    lems from the Standpoint of Artificial Intelligence8. In B. Meltzer

    and D. Michie, editors, Machine Intelligence 4, pages 463–502. Ed-

    inburgh University Press, 1969. Reprinted in [McC90].

    [Rei01] Raymond Reiter. Knowledge in Action. M.I.T. Press, 2001.

    [Sha97] Murray Shanahan. Solving the Frame Problem, a mathematical

    investigation of the common sense law of inertia. M.I.T. Press,







    [Tho86] K. Thompson. Retrograde analysis of certain endgames.


    (International Computer Chess Association) Journal, 9(3):131–139,


    [Tho03] Richmond Thomason. Logic and artificial intelligence.

    In Ed-

    ward N. Zalta, editor, The Stanford Encyclopedia of Philosophy.




    /@steam.stanford.edu:/u/ftp/jmc/freewill2.tex: begun Thu May 16, 2002, latexed November 6, 2005