In his Novum Organum, Francis Bacon classified the intellectual fallacies of his time under four headings which he called idols (Nova Organum 1620: 345:39):

There are four classes of Idols which beset men's minds. To these for distinction's sake I have assigned names, calling the first class Idols of the Tribe; the second, Idols of the Den; the third, Idols of the Market Place; the fourth, Idols of the Theater.

An idol is an image, in this case held in the mind, which receives veneration but is without substance in itself. Bacon did not regard idols as symbols, but rather as fixations. In this respect he anticipated modern psychology.

'Idols of the Tribe' are deceptive beliefs inherent in the mind of man, and therefore belonging to the whole of the human race. They are abstractions in error arising from common tendencies to exaggeration, distortion, and disproportion. First of all, our tendency to let emotions rule reason can give us false impressions of the truth based on our feelings at the time. Another common idol lies in our tendency to seek out evidence of that which we already believe to be true. Bacon suggests that we become affectionate to ideas we have found and carried with us for some time; we become attached to them and collect evidence that supports them while throwing out that which contradicts them. Interestingly, Castelfranchi (2001) proposes a similar mechanism to explain the emergence of social functions among agents through ‘learning without understanding’ processes, superseding the intentional cognitive processes. But, so far, emotional agents remain out of reach of the current developments in SAI.

'Idols of the Den' (also called 'Idols of the Cave' in some editions) are those which arise within the mind of the individual. Like in Plato’s allegory, thoughts of the individual roam about in this dark cave and are variously modified by temperament, education, habit, environment, and accident. Thus an individual who dedicates his mind to some particular branch of learning becomes possessed by his own peculiar interest, and interprets all other learning according to the colors of his own devotion. In our case, this idol may affect directly the designer rather than the artificial agent. Lissack and Richardson (2001: 101) give an illustration of this bias in their criticism of Wolfram’s claim that most complex systems can be accurately represented by rule-based atomistic models:

The act of interpreting differs from the act of observing, and both may differ significantly from the underlying phenomenon being observed. In their failure to respect this distinction, strong MSM proponents [Wolfram and colleagues] are implicitly suggesting that the interpretation is reality. However, while a good model of complex systems can be extremely useful, it does not allow us to escape the moment of interpretation and decision.

Obviously, deterministic and limited expert knowledge used to design FLN agents suffers the same type of criticism. Causal rules of behaviour inferred by an expert are merely subjective interpretations of a given reality. Putting it simply, there are as many realities as there are experts. Agar (2005) recently proposed a way forward by means of coupled emic/etic approaches to social simulations. The author advocates a constant feedback between what makes sense for the actual actors depicted in the model (emic) and what seems meaningful to the designer (etic).

'Idols of the Marketplace' are errors arising from the false significance bestowed upon words, and in this classification Bacon anticipated the modern science of semantics. The constant impact of words variously used without attention to their true meaning often betrays their purpose, obscuring the very thoughts they are designed to express. Acknowledging the volatility of the ‘true meaning’of words, Bacon just caught a glimpse of the active perception, theorised by Kant and Peirce later on. Words, as elementary percepts, do not carry any specific and intrinsic meaning when they are perceived. They have to be re-interpreted internally by the receiver, according to previous knowledge and environmental hints. Maturana and Varela (1980: 32) propose to drop the denotative understanding of language altogether in favour of a connotative approach:

So long as language is considered to be denotative it will be necessary to look at it as a means for the transmission of information, as if something were transmitted from organism to organism, in a manner such that the domain of uncertainties of the 'receiver' should be reduced according to the specifications of the 'sender'. However, when it is recognized that language is connotative and not denotative, and that its function is to orient the orientee within his cognitive domain without regard for the cognitive domain of the orienter, it becomes apparent that there is no transmission of information through language. It behooves the orientee, as a result of an independent internal operation upon his own state, to choose where to orient his cognitive domain; the choice is caused by the 'message', but the orientation thus produced is independent of what the 'message' represents for the orienter. In a strict sense then, there is no transfer of thought from the speaker to his interlocutor; the listener creates information by reducing his uncertainty through his interactions in his cognitive domain.

As stated earlier, mainstream SAI satisfies itself with passive perception processes. Hence, most FLC agents can receive intelligible and meaningful information from other agents, without having to engage into deciphering and re-interpretation stages. Somehow, it makes agents’ lives seem much easier than ours!

'Idols of the Theatre' occur due to sophistry and false learning. These idols are built up in the fields of theology, philosophy, and science, and, because they are defended by learned groups, are accepted without question by the masses. When theories have been cultivated and have reached a sufficient level of consensus they are no longer questioned. The long standing hegemony of the symbolico-cognitivist paradigm cannot hide the fact that relevant alternatives have challenged SAI’s dominion: connectionist and evolutionary theories perform better on learning processes (Kaplan 2001); the autopoietic theory, by refusing the conveyance of information through linguistic interaction, provides a unified and unchallenged approach to signaling interactions (verbal, non-verbal, or extra-verbal) through structural coupling between individuals (Maturana and Varela 1980).

Prospect theory focuses on cognitive and psychological factors that determine the value of risky prospects (Kahnemann and Tversky 2000). Its initial assumption is that subjective values attached to gambling are carried by expected changes of wealth (gains or losses) rather than ultimate states of wealth. More importantly, prospect theory replaces the traditional concept of risk aversion by a more intuitive one, called loss aversion, by which people tend to consider that a loss of $X is more averse than a gain of $X is attractive. This assumption explains why people might be risk seeking, and no longer risk averse, in the domain of losses.

Though trivial from an espistemological viewpoint, Prospect theory tends to reconcile theoretical development with empirical facts. According to Kahnmann and Tversky (2000:1):

The study of decision addresses both normative and descriptive questions. The normative analysis is concerned with the nature of rationality and the logic of the decision making. The descriptive analysis, in contrast, is concerned with people’s beliefs and preferences as they are, not as they should be.

Although based on formal logic predicates, Prospect theory recognises the fact that part of the knowledge necessary to complete the theory is inherently elusive. It is possible to design FLC agents founding their decisions on this theory, but the axiomatic predicates that will tell us about discrepancies between gains and losses, or the way different items will be affected by these discrepancies remain, by far, out of reach.

Furthermore, prospect theory threatens directly two logical pillars of decision theory traditionally used by SAI: preference invariance and value coincidence. Invariance requires that the preference order between prospects should not depend on the manner in which they are described. Kahneman and Tversky (2000:5) have demonstrated that invariance cannot generally be satisfied: ‘invariance is normatively essential, intuitively compelling, and psychologically unfeasible’ due to the framing of outcomes through formulation effects. The framing effect also affects the relation between experience and decision values. But rational agents seldom make a difference between experience values (direct outcomes of actual actions) and decision values (expected outcomes of an anticipated choice). These two values tacitly coincide, despite Kahnemann’s and Tversky’s (2000: 16) warning:

Some factors that affect experience are not easily anticipated, and some factors that affect decisions do not have a comparable impact on the experience of outcomes.

For Bacon, knowledge is intimately mixed with the idols, hence prefiguring our modern concept of belief. More importantly, Bacon draws visionary consequences from the presence of the idols, in terms of communication (Nova Organum 1620: 346:35):

enter quietly into the minds that are fit and capable of receiving it; for confutations cannot be employed, when the difference is upon first principles and very notions and even upon forms of demonstration.

Individuals and groups exhibit varied responses when faced with new information. If such information is consistent with extant behaviours and beliefs, it can be readily accepted and integrated. However, if the new information conflicts with behaviour and belief, the resulting state is described as ‘cognitive dissonance’ (Bradshaw and Borchers 2000). According to the theory, the inconsistency and psychological discomfort of cognitive dissonance can be reduced by changing one's beliefs, values, or behaviour. Dissonance can also be avoided by rejecting or avoiding information that challenges belief systems or by interpreting dissonant information in a biased way. In this regard, most SAI structures force agents to discard new information conflicting with a given set of consistent predicates. It is only through reinforcement (punishment or reward) due to experience that contrasted set of predicates can be established. Elaborating on the conflicting views upon ‘uncertainty’ between scientists and policy-makers to explain the science-policy gap, Bradshaw and Borchers (2000: 3) outline the complexity of cognitive dissonance:

Dissonance between existing beliefs and new information may be shaped by a host of factors, all of which inhibit the rate at which scientific findings are assimilated into policy. In what we have called the ‘volition’ phase of the science-policy gap, public debate around an emerging scientific consensus may derive from a combination of cultural, psychological, and economic interests threatened by the policy inferences of dissonant scientific findings.

The authors particularly point at the contrasted rhetorical figures used by scientists when they are in charge of policy-making compared with their usual handling of scientific uncertainties. Designers in DAI have tried to encapsulate these internal cognitive conflicts by implementing Agent-Group-Role structures in which one agent belongs simultaneously to several socio-cultural groups and plays different roles accordingly (Ferber, 1999). But the internal conflict resolution mechanisms—for example, between friendship engagement and professional commitment—rely on formal and individualistic logic once more.

Overall, despite the increasing interest in, and use of Multi Agent Systems to represent human ecosystems, the SAI paradigm remains mostly unchallenged as a theoretical framework used to develop cognitive agents. But looking at the real world through the lenses of the social sciences, we have to acknowledge the fact that a meta-theory of human behaviour doesn’t exist yet. Rational decision theory, prospect theory, social learning theory, and others give us partial clues about human behaviour, and none can stand as an overarching and unified framework.

Hence, we must handle cautiously Cederman’s assertion about the ‘New Deal’ offered by Multi Agent Systems to social scientists. These tools and their current states of application generally rely upon reductionist views of the world: symbolico-cognitive hypothesis, formal predicate logic, rational decision, and autonomy. As a matter of fact, the SAI paradigm stands out as a nomothetic meta-model for agent’s behaviour while its foundation doesn’t represent a meta-model of human behaviour but merely a partial interpretation of it.