Page 426 -

17 Social Constraint 429

psychological theories. Behaviourism is a theory of learning developed principally
through experiments with animals. For instance, the conditioning experiments of
Ivan Pavlov are well known: he demonstrated that dogs can be trained to exhibit
a specific reaction such as salivation by presenting a specific stimulus such as the
sound of a bell together with food (Pavlov 1927). Bandura (1962, 1969) extended
the behaviouristic approach with a social dimension by developing a theory of social
learning through imitation. From a behaviouristic perspective, norms constitute a
learned behaviour and thus have to be explained using these theories. The dynamical
propensities of models inspired by game theoretical concepts are a straightforward
implementation of such a view on intra-agent processes. The propensity to cooperate
or defect is updated in proportion to the propensity of sanctions. The propensity of
sanctions, however, is a structural component resulting from inter-agent processes.
Hence, agents learn to modify their behaviour according to structural conditions.
Here we find the feedback loop between social and individual components that
are in fact essential for the concept of norms. However, the third component is
missing: this approach does not include a concept of obligations. Deontics are
out of the scope of this approach. This shortcoming can be traced back to the
psychological theory that is represented in the agents: behaviourism is not capable of
capturing mental processes. Indeed, it specifically avoids commenting on the mental
processes involved. Under the influence of positivism, reference to unobservable
entities such as the ‘mind’ has been regarded as not scientifically valid. Obligations
are such unobservable entities. Hence, they cannot be represented by the means of
behaviouristic learning theories that are applied in agent models.
In socialisation research, the complex cognitive processes necessary for grasping
the meaning of an obligation is denoted as internalisation. It has already been shown
that agent transformation is not the same as the internalisation of norms. This is also
behaviourally important because normative behaviour, guided by deontics, need not
be a statistical regularity, guided by propensities. In particular if moral reasoning is
involved, deviant behaviour is not explained by chance variation, leading to some
kind of normal distribution (where the mean value might be updated). There is a
difference between norms and the normal.
To represent a complex cognitive concept such as norm internalisation calls
for the cognitively rich agents of the AI tradition. However, the examination of
current models has revealed that a comprehension of the cognitive mechanisms
by which social behaviour regulation becomes effective in the individual mind is
still in its fledgling stages. It has been shown that a multiplicity of concepts is at
hand: while in the very beginning the agents were merely normative automata, there
exist conceptualisations of normative agent transformation ranging from updating
conditionals (of strategies) through knowledge to signalling and memetic contagion.
However, no consensus is reached what are the most relevant mechanisms. It can
be suspected that they remain effect generating rather than process representation
mechanisms. As an agenda for the next decade, a closer examination of the
processes by which normative obligation becomes accepted by humans might be
useful. For this purpose, it is necessary to recall socialisation theory. This will
help to clarify the problem situation. However, the results of socialisation research
are not unequivocal. Hence, in the development of a normative architecture, some
fundamental decisions have to be made.