Revised expanded version of 1988 original.

Aaron Sloman School of Computer Science
The University of Birmingham

This was originally written while I was at the University of Sussex.
A slightly shorter version of the following text was broadcast via net news
("usenet") to the and news groups some time in 1988.

It began with this remark, which explains the closing question:
   "I wasn't going to contribute to this discussion, but a colleague encouraged me."

A slightly modified version of this appeared in
AISB Quarterly, Winter 1992/3, Issue 82, pp.31-2

NOTE 2 (8 Apr 2014):
The formatting of this version has been changed, e.g. with the list of transitions
now numbered, below, where reference is made to Stan Franklin's
book chapter which made use of these distinctions (with permission).
He also made a number of additional distinctions in the same spirit, summarised

NOTE 3 (9 Apr 2014):
A section from my 1978 Book discussing free will is appended below.



Much philosophical discussion concerning freedom of the will is based
on an assumption that there is a well-defined distinction between
systems whose choices are free and those whose choices are not. This
assumption is refuted by showing that when requirements for behaving
systems are considered there are very many design options which
correspond to a wide variety of distinctions more or less closely
associated with our naive ideas of individual freedom. Thus, instead
of one major distinction there are many different distinctions;
different combinations of design choices will produce different sorts
of agents, and the naive distinction is not capable of classifying
them. In this framework, the pre-theoretical concept of freedom of the
will needs to be abandoned and replaced with a host of different
technical concepts corresponding to the capabilities enabled by
different designs.



Philosophy done well can contribute to technical problems (as shown by the influence
of philosophy on logic, mathematics, and computing, e.g. via Aristotle, Leibniz,
Frege and Russell).

Conversely, technical developments can also help to solve or dissolve old
philosophical problems. I think we are now in a position to dissolve the problems of
free will as normally conceived, and in doing so we can make a contribution to AI as
well as philosophy.

The basic assumption behind much discussion of freedom of the will is:

(A): there is a well-defined distinction between systems whose choices are free
and those whose choices are not free.

However, if you start examining possible designs for intelligent systems in great
detail you find that there is no one such distinction. Instead there are many
'lesser' distinctions corresponding to design decisions that a robot engineer might
or might not take - and in many cases it is likely that biological evolution tried
both (or several) alternatives.

There are interesting, indeed fascinating, technical problems about the
implications of these design distinctions. For example, we can ask how individuals
with the different designs would fare in a variety of social settings, what they
would be like to interact with, which sorts of tasks they would be able to achieve
and which not. Exploring design details shows, I believe, that there is no longer any
interest in the question whether we have free will because among the real
distinctions between possible designs there is no one distinction that fits the
presuppositions of the philosophical uses of the term "free will". It does not map
directly onto any one of the many different interesting design distinctions.
So (A) is false.

"Free will" has plenty of ordinary uses to which most of the philosophical discussion
is irrelevant. E.g.

    "Did you go of your own free will or did she make you go?"
That question presupposes a well-understood distinction between two possible
explanations for someone's action. But the answer "I went of my own free will" does
not express a belief in any metaphysical truth about human freedom. It is merely a
denial that certain sorts of influences operated, such as threats or coercion by
another person. There is no implication that no causes, or no mechanisms were
involved. How could any lay person know that there are no causes, since we know very
little about how our brains work?

The claim to have done something of your own free will simply illustrates a
common-sense distinction between the existence or non-existence of particular sorts
of 'external' influences on a particular individual's action. We could all list types
of influences that might make us inclined to say that someone did not act of his own
free will, some of which would, for example, lead to exoneration in the courts. But
saying "I did not do it of my own free will because processes in my brain caused me
to do it" would not be accepted as an excuse, or a basis for requesting forgivenness.

However there are other deeper distinctions that relate to different sorts of designs
for behaving systems, but our ordinary language does not include terms for
distinguish behaviour flowing from such different designs. Before we can introduce
new theory-based distinctions, we need to answer the following technical question
that lurks behind much of the discussion of free will.

"What kinds of designs are possible for intelligent agents and what are the
implications of different designs as regards the determinants of their actions?"
I'll use "agent" as short for "behaving system with something like motives". What
that means is a topic for another day. Instead of *one* big division between things
(agents) with and things (agents) without free will we'll then come up with a host of
more or less significant divisions, expressing some aspect of the pre-theoretical
free/unfree distinction. E.g. here are some examples of design distinctions (some of
which would subdivide into smaller sub-distinctions on closer analysis):

Design distinctions
(Most of these were used in the discussion of Free Will in Franklin (1995), Chapter 2.
The relevant text is available, starting on page 31 on page 31, and adds additional
distinctions summarised below.)

  1. Compare (a) agents that are able simultaneously to store and compare different
    motives with
    (b) agents that have no mechanisms enabling this: i.e. they can have only one motive
    at a time.

  2. Compare (a) agents all of whose motives are generated by a single top level goal
    (e.g. "win this game") with
    (b) agents that have several independent sources of motivation (motive generators
    implemented in hardware or software), e.g. thirst, sex, curiosity, political
    ambition, aesthetic preferences, etc.

  3. Contrast (a) an agent whose development includes modification of its motive
    generators and motive comparators in the light of experience, with
    (b) an agent whose generators and comparators are fixed for life (presumably the case
    for many animals).

  4. Contrast (a) an agent whose motive generators and comparators change partly under the
    influence of genetically determined factors (e.g. puberty), with
    (b) an agent for whom they can change only in the light of interactions with the
    environment and inferences drawn therefrom.

  5. Contrast (a) an agent whose motive generators and comparators (and higher
    order motivators) are themselves accessible to explicit internal scrutiny, analysis
    and change, with
    (b) an agent for which all the changes in motive generators and comparators are
    merely uncontrolled side effects of other processes (as in addictions, habituation,
    etc.) A similar distinction can be made as to whether motives themselves are or are
    not accessible to explicit internal scrutiny, analysis and change.

  6. Contrast (a) an agent pre-programmed to have motive generators and comparators change
    under the influence of likes and dislikes, or approval and disapproval, of other
    agents, and
    (b) an agent that is only influenced by how things affect it. The former will be more
    likely than the latter to absorb the values of its culture.

  7. Compare (a) agents that are able to extend the formalisms they use for thinking about
    the environment and their methods of dealing with it (like human beings) and
    (b) agents that are not (most other animals?)

  8. Compare (a) agents whose motives are never inconsistent, e.g. because the latest
    motive always removes all others, and
    (b) agents that can simultaneously have incompatible motives (e.g. wanting to drink
    at the water hole and wanting not to go near that hungry looking lion crouching
    beside the water).

  9. Compare (a) agents that are able to assess the merits of different inconsistent
    motives (desires, wishes, ideals, etc.) and then decide which (if any) to act on with
    (b) agents for which motivator conflicts are always resolved using some automatic
    reaction, e.g. having a measure of 'strength' associated with each motive and always
    being driven by the strongest motive, or agents that are always controlled by the
    most recently generated motive (like very young children and perhaps some other

  10. Compare (a) agents with a monolithic hierarchical computational architecture where
    sub-processes cannot acquire any motives (goals) except via their 'superiors', with
    only one top level executive process generating all the goals driving lower level
    systems with
    (b) agents where individual sub-systems can generate independent goals.
    In case (b) we can distinguish many sub-cases, for instance:
    (b1) the system is hierarchical and sub-systems can pursue
    their independent goals if they don't conflict with the goals
    of their superiors

    (b2) there are procedures whereby sub-systems can (sometimes?)
    override their superiors (e.g. trained reflexes?)

  11. Compare (a) a system in which all the decisions among competing goals and
    sub-goals are taken on some kind of 'democratic' voting basis or a numerical
    summation or comparison of some kind (a kind of vector addition perhaps) with
    (b) a system in which conflicts are resolved on the basis of qualitative rules, some
    of which are determined genetically (from birth) and some of which are products of a
    complex high level learning system.

  12. Compare (a) a system designed entirely to take decisions that are optimal for its own
    well-being and long term survival with
    (b) a system that has built-in mechanisms to ensure that the well-being of others is
    also taken into account. (Human beings and many other animals seem to have some
    biologically determined mechanisms of the second sort - e.g. maternal/paternal
    reactions to offspring, sympathy, etc.).

  13. Compare (a) a system that includes some kind of random generator that determines some
    of its major decisions and
    (b) a system all of whose decisions are based on its motives, beliefs, preferences,
    etc. which in turn are produced by totally deterministic processes including long
    term learning. There are many distinctions that can be made between systems according
    to how much knowledge they have about their own states, and how much they can or
    cannot change because they do or do not have appropriate mechanisms. (As usual there
    are many different sub-cases. Having something in a write-protected area is different
    from not having any mechanism for changing stored information at all.)
NOTE (9 Apr 2014)
Franklin's chapter adds several more distinctions in the same spirit,
available here. The additional distinctions are concerned with differences in
sense modalities (which he labels S1, S2), memory mechanisms (M1, M2,
M3, M4), and differences in planning, visualising and creating mental
models (T1, T2, T3).
There are some overlaps between these distinctions, and many of them are relatively
imprecise, but all are capable of refinement and can be mapped onto real design
decisions for a robot-designer (or evolution).

They are just some of the many interesting design distinctions whose implications can
be explored both theoretically and experimentally, though building models
illustrating most of the alternatives will require significant advances in AI e.g. in
perception, memory, learning, reasoning, motor control, etc.

When we explore the fascinating space of possible designs for agents, the question
which of the various systems has free will loses interest: the pre-theoretic
free/unfree contrast totally fails to produce any one interesting demarcation among
the many possible designs -- though it can be loosely mapped on to several of them.
However, different mappings will imply different implications for classifying an
agent as free, or as unfree.

After detailed analysis of design options we may be able to define many different
notions of freedom, with corresponding predicates:- free(1), free(2), free(3), ....
However, if an object is free(i) but not free(j) (for i /= j) then the question "But
is it really FREE?" has no answer.

It's like asking: What's the difference between things that have life and things that

The question whether something is living or not is (perhaps) acceptable if you are
contrasting trees, mice and people with stones, rivers and clouds. But when you start
looking at a larger class of cases, including viruses, complex molecules of various
kinds, and other theoretically possible cases, the question loses its point because
it uses a pre-theoretic concept ("life") that doesn't have a sufficiently rich and
precise meaning to distinguish all the cases that can occur. (This need not stop
biologists introducing a new precise and technical concept and using the word "life"
for it. But that doesn't answer the unanswerable pre-theoretical question about
precisely where the boundary lies.)

Similarly "What's the difference between things with and things without free will?"
may have an answer if you are contrasting on the one hand, thermostats, trees and the
solar system with, on the other hand, people, chimpanzees and intelligent robots. But
if the question is asked on the presumption that all behaving systems can be divided,
then it makes the false assumption (A).

So, to ask whether we are free is to ask which side of a boundary we are on when
there is no particular boundary in question, only an ill-defined collection of very
different boundaries. This is one reason why it is that so many people are tempted to
say "What I mean by 'free' is..." and they then produce different incompatible

In other words, the problem of free will is a non-issue. So let's examine the more
interesting detailed technical questions in depth.

It is sometimes thought that the success of computational models of the human mind
would carry the implication that we lack freedom because computers have no freedom.
However, as I argued in section 10.13 of Sloman (1978) (below), on the contrary, such
models may, at last, enable us to see how it is possible for agents to have an
architecture in which their own desires, beliefs, preferences, tastes and the
like determine what they do rather than external forces or blind physical and
chemical processes. This line of thinking is elaborated in the books and papers cited
in the bibliography. Dennett (1984), in particular, analyses in considerable depth
the confusions that lead people to worry about whether we are free or not.

Now, shall I or shan't I submit this.........????
[The question with which the original usenet posting ended is explained above.]

Originally written over a decade before the above:
Inserted 9 Apr 2014:
Extract from Sloman (1978) Chapter 10 Section 13:

10.13. Problems about free will and determinism

A common reaction to the suggestion that human beings are like computers running
complex programs is to object that that would mean that we are not free, that all our
acts and decisions are based not on deliberation and choice but on blind
deterministic processes. There is a very tangled set of issues here, but I think that
the study of computational models of decision-making processes may actually give us
better insights into what it is to be free and responsible. This is because people
are increasingly designing programs which, instead of blindly doing what they are
told, build up representations of alternative possibilities and study them in some
detail before choosing. This is just the first step towards real deliberation and
freedom of choice.

NOTE (Added 9 Apr 2014)
For a discussion of subdivisions between proto-deliberative systems and
various other increasingly sophisticated kinds of deliberative systems,
see Sloman(2006).

In due course, it should be possible to design systems which, instead of always
taking decisions on the basis of criteria explicitly programmed in to them (or
specified in the task), try to construct their own goals, criteria and principles,
for instance by exploring alternatives and finding which are most satisfactory to
live with. Thus, having decided between alternative decision-making strategies, the
program may use them in taking other decisions.

For all this to work the program must of course have some desires, goals, strategies
built into it initially. But that presumably is true of people also. A creature with
no wants, aims, preferences, dislikes, decision-making strategies, etc., would have
no basis for doing any deliberating or acting. But the initial collection of programs
need not survive for long, as the individual interacts with the physical world and
other agents over a long period of time, and through a lengthy and unique history
extends, modifies, and rejects the initial program. Thus a robot, like a person,
could have built into it mechanisms which succeed in altering themselves beyond
recognition, partly under the influence of experiences of many sorts.
Self-modification could apply not only to goals but also to the mechanisms or rules
for generating and for comparing goals, and even, recursively, to the mechanisms for

This is a long way from the popular mythology of computers as simple-minded
mechanisms which always do exactly what they are programmed to do. A self-modifying
program, of the sort described in chapter 6, interacting with
many people in many situations, could develop so as to be quite unrecognisable by its
initial designer(s). It could acquire not only new facts and new skills, but also new
motivations; that is desires, dislikes, principles, and so on. Its actions would be
determined by its own motives, not those of its designers.

If this is not having freedom and being responsible for one's own development
and actions, then it is not at all clear what else could be desired under the name of

As people become increasingly aware of the enormous differences between these new
sorts of mechanisms, and the sorts of things which have been called mechanisms in the
past (clocks, typewriters, telephone exchanges, and even simple computers with simple
programs), they will also become less worried about the mechanistic overtones of
computer models of mind. (See also my 1974 paper on determinism. Sloman (1974))

Some possibly useful references
(Expanded 9 Apr 2014)