I argue that neither explicit user evaluation nor self-directed exploration of pre-determined spaces of possibilities are viable approaches for implementing arti cial systems whose products are recognizable by humans as creative and valuable and, at the same time, as genuinely authored by the arti cial system (rather than by the original designer or by the human users): indeed, self-directed creativity requires critical engagement with a creative community, and for now arti cial systems are not capable of interacting with human creative communities as peers. I suggest a possible alternative, which might hypothetically be used to grant some degree of authorship to arti cial systems without having general arti cial intelligence as a prerequisite: in brief, arti cial systems could be trained, through the use of biometrics, to generate products that induce the same types of reactions in humans of recognized creative works. As this would not draw conscious human decision-making into the production process, there would then be grounds to argue that the products of such a system would have indeed non-human authorship.
the natures of their contents and the reasons why they are so well-regarded by the zots are truly anyone's guess.
The night before the start of the competition, Carol and Bob | two of the human participants | are discussing about their submissions over a glass of space ale: Carol: \I thought a lot about this, and I'm pretty sure I made the right choice in picking bluethings as the main theme of my work." Bob: \Bluethings? What are they?" Carol: \I'm not sure, to be honest, but they are blue. And spiky. Zots love them, that's for certain." Bob: \Uh. Seems di cult, painting something without knowing what it is or what signi cance it has..." Carol: \Yes, but I studied their works, and many of their most greatest pieces are all about bluethings. Well, bluethings and shinythings." Bob: \And the shinythings are..." Carol: \No idea, again, but they are shiny | shiny and round. The zots draw them a lot too, so they must be pretty important to them. Oh, and I noticed that whenever a composition contains bluethings, it never contains shinythings unless there is a jagged line separating them, so of course I did the same | it must be a cultural taboo of some sort, and it would not do to o end our guests. But enough about me, what is your piece about?" Bob: \Well, it is just a group of children playing board games..." Carol: \What? But we do not even know if the zots have any notion of childhood! And how are they supposed to recognize a board game? I really hope you won't lose us the competition..."
Mutatis mutandis, this would not be too out of place as a hypothetical dialogue between two computer programs for automatically generating creative works for human consumption. No arti cial system has yet been built which contains an acceptable model of the overall themes and motivations which drive human creative production and the appreciation thereof, largely because we ourselves are not entirely clear on what these themes and motivations are; but a good number of systems have been built which, much like Carol, generate automatically works in accordance to the \rules" typical of certain categories of human creations. There also exist Bob-like systems, which eschew human imitation entirely and operate through the iterative application of rules which have very little to do with human stylistic conventions; but then, very careful tuning of their parameters and algorithms is necessary if their products are to be appreciable by most humans beyond their novelty value.
If it is not easy to nd a common artistic ground between zots and humans, it cannot be much simpler to nd one between humans and non-sapient, nonbiological, arti cial processes; and the fact that such processes are human-made does not by itself imply that they are human-like in any meaningful sense, no matter the super cial similarities (or lack thereof) between their results and other forms of human expression.
But there is one approach to the problem of painting for zots that the artists of our story failed to consider: they could have attempted to show a variety of works to zots, measured accurately their physiological reactions, and compared them to those observed when they are shown well-regarded works of zot origin. If the zots' reactions in the two cases were similar, then | regardless of the the many ways in which their psychological makeup might di er from that of human beings | there would be reason to suspect that their psychological reactions would be also similar; and if instead they were di erent, it would not be unreasonable to conclude that their psychological reactions are also di erent.1
In what follows, I will argue that essentially the same process, with us taking the role of the zot subjects and our machines taking the role of the human examiners, provides us with a promising avenue for the exploration of creativity through arti cial means. -3.
One of the main challenges faced in the study of creativity | and, at the same time, one of its most intriguing aspects | is the fact that recognizing the products of creative endeavors is a very context-dependent task. To mention a rather infamous example, a urinal may or may not be considered a creative work depending on whether it is placed in a museum or in its more typical habitat; and, in the former case, its possible recognition as such presupposes the appreciation of a complex set of cultural and historical circumstances. A hypothetical alien, bereft of any knowledge of the history of European artistic movements and of human biological functions, would presumably nd the same object just as confusing in both circumstances; and, while they may notice that the placement of such an object in a museum is less common than its placement in a bathroom, they would have little way to infer the signi cance attributed to the urinal in the second case.
More prosaically, cross-cultural recognition and evaluation of creative works is a notoriously di cult enterprise; and it would not be unreasonable to expect this di culty to only worsen in the case of works generated by nonhuman, articial processes whose structures and behaviours are, after all, vastly dissimilar from the ones of human beings.
For many products of arti cial processes, this is not at all the case. Instead,
the processes they originate from are tailor-built to imitate known forms of
human creative activity, be it musical composition
It is worth emphasizing here that relaxing the constraints of such a system, for example by allowing it to arbitrarily break out from some of the boundaries of the target conceptual space, would not in itself make it more creative. For instance, the fact that the numbering of the sections of the present work starts from minus four rather than from one is a largely pointless deviation from the common conventions of writing which does not increase in any sense its quality. Further changes to this work along similar lines could only make the experience of reading it more and more frustrating: the subversion of expectations can at times be used to convey interesting e ects, but breaking them at random is unlikely to improve the quality of the product.
On the other hand, there are also works of generative creativity that cannot
be reasonably described as mere attempts to imitate the super cial features of
human products. These works emerge from processes, often based on notions
from arti cial life, which develop and unfold freely and without any expectation
of compliance with the conventions of human genres: as an example of this kind
of approach, we can mention here Reas' Process Compendium
Many theoretical models of the human creative process have been produced
Thus, without making any strong commitment to any theory of creativity, we can reasonably recognize at least two distinct feedback loops as having a role in the creative process: 1. The internal loop, through which the author evaluates the products of their own activity and uses the results of such evaluations to direct their future activity; 2. The external loop, through which the very criteria used by the author to evaluate their own work are updated by such means as the observation of other people's reactions to the author's works or the author's own evaluation of others' works.
These two loops are necessary components of any genuinely creative process: indeed, the activity of an author bereft of the internal loop would be nothing but mere automatic reaction to external stimuli, whereas an author bereft of the external loop would lack any means for updating their own evaluation criteria, and would thus have no way out of their solipsistic cycle of self-evaluation and optimization according to a pre-de ned, static ideals.
But this poses a serious problem for arti cial creativity. Building an arti cial agent with an external loop towards a human creative community of richness comparable to that of the loops linking humans authors to it would appear to require endowing the arti cial agent with human-comparable cognitive and linguistic abilities; and this, despite the advances of arti cial intelligence, still remains an unsolved task. Do we have to give up, and accept that arti cial creativity requires general-purpose arti cial intelligence?
Maybe not. A potential alternative solution, perhaps, might be to piggyback
on the evaluation functions of human users, for instance by having them rate
the quality of the arti cial agent's products and attempting to incrementally
improve the average quality of its solutions. This is the approach exempli ed by
much interactive evolutionary art, from Dawkins' Biomorphs
Indeed, the locus of agency | and, therefore, of creativity | of such systems does not lie within the arti cial process, but instead solidly within the human beings evaluating its products. If the human users wanted the process' results to become dull and uninteresting, they could easily succeed; and if they wanted them to improve, they could only succeed up to the degree to which they can correctly evaluate the quality of partial improvements on the current product. The machine takes part to the internal loop, but not to the external loop, whereas the human evaluators take part to the external and to the internal loop; thus, the arti cial agent appears to be taking a role more akin to that of a tool (if a somewhat unpredictable one) than to that of a creator. Nevertheless, interactive evolutionary algorithms show much promise: by providing human authors with a vast repertoire of themes and possibilities that they might not necessarily have developed autonomously, these algorithms are very well-suited for performing the role of creativity-enhancing tools.
In the previous section, I suggested that interactive evolutionary art is not usually quite an example of arti cial creativity because the onus of evaluating the results of the algorithm lies entirely with the human user. But a natural question at this point would be: how does this di er anyway from the way in which the creations of a human author are evaluated by their community? The answer, in short, is that peer evaluations do not generally rob human authors of their autonomy, because human authors are not typically in thrall of their peers. They instead listen to their opinions and motivations, arguing | occasionally quite vociferously | about their validity or lack thereof, and more than occasionally rejecting them altogether. It might not be incorrect to claim that, nonetheless, this implies that authorship is in some sense shared between the direct creator and the community whose teachings and overall opinions a ected their choices; but, in any case, it is clear that while human authors do generally seek the approval of their public, their activity is up to some not insigni cant degree self-directed | they face their critics as peers, not as slaves. This is not the case for generative programs, whose activity is entirely determined by their users' evaluations, by random or pseudo-random events, and by the possibilities and preferences inherent to their programming. The rst two kinds of in uences mentioned cannot be reasonably compared to a human author's own autonomous decisions; and as for the third one, it constitutes a true case of arti cial authorship only insofar as such possibilities and preference have not been designed by the programmer with the express purpose of causing the program's products to be of a certain speci c quality.
Our imaginary critic is, thus, not entirely wrong, and we have to retract to some degree the overly strong claim made in the previous section: an interactive evolutionary system can indeed be granted authorship, but only up to the measure in which its behaviour is not a function of the user's evaluations, of random events, or of the designs of its own author. This, however, does not resolve all our problems: indeed, what we would ideally want to have would be an arti cial creative system whose products were \human-like" enough for us to evaluate and appreciate. This puts us in the same situation of the zots mentioned in Section -4: we want works of value to us, and we want them out of entities which are almost entirely unlike ourselves, whose choices we do not want to have to micromanage, which lack any even vaguely accurate model of ourselves, and which are not capable of communicating with us as equals. How can they possibly succeed?
One possible answer, the one this work has been driving towards, is the
following: if an arti cial process could measure human involuntary reactions to
high-quality works of human creativity, it could then | at least in principle |
iteratively update its products3 while attempting to increase the degree of
simi3 Far more ambitiously, if we wished for the machine to eventually become capable of
generating valuable works without constant biological feedback, we might attempt
to make it optimize its production algorithm through some form of genetic
programming, rather than its individual products. In essence, this would be equivalent to
having the machine attempt to iteratively build a model of the features of humankind
larity between our responses to them and our responses to the above-mentioned
human works.4 This would put the agency in the creative process rmly out
of the reach of human consciousness, thus leading to the situation depicted in
Figure 3: the machine is now taking part in the internal loop of creation and
self-evaluation, but also | through the monitoring of the involuntary reactions
of its human users, which now are not making decisions in its place | in the
external loop of interaction with the larger human community.
In this work, I discussed some of the di culties inherent to the problem of
creating non-human agents which are capable of creating novel and
valuableto-humans works without relying exclusively on human decisions (made either
at implementation time or at execution time); and I suggested a possible
approach based on the use of biometrics. There already exist works of generative
which are relevant to our appreciation, or lack thereof, of its products. Starting from
the perspective of Figure 3, this would amount to pushing the human users entirely
out of the internal loop of creation and self-evaluation.
4 In order to prevent plagiarism, we might also want to add the additional requirement
that the algorithm's products should be su ciently di erent from all the
humanmade examples used to nd the human physiological responses to high-quality work.
This might or might not be necessary, depending on the speci c characteristics of
the algorithm; it could be also be the case that such a program would not be likely
to produce any given example anyway.
art which rely on biometrics, such as for example