=Paper=
{{Paper
|id=Vol-2287/paper1
|storemode=property
|title=Evolution of Conscious AI in the Hive: Outline of a Rationale and Framework for Study
|pdfUrl=https://ceur-ws.org/Vol-2287/paper1.pdf
|volume=Vol-2287
|authors=David Sahner
|dblpUrl=https://dblp.org/rec/conf/aaaiss/Sahner19
}}
==Evolution of Conscious AI in the Hive: Outline of a Rationale and Framework for Study==
https://ceur-ws.org/Vol-2287/paper1.pdf
Evolution of Conscious AI in the Hive:
Outline of a Rationale and Framework for Study
David Sahner M.D.
Chief Scientific Officer, EigenMed
Abstract. The paper proposes a framework for investigating a potentially con-
scious AI system by considering an autonomous, multitask-capable, powerful,
highly intelligent and adaptive system (AMPHIA) as one more likely to acquire
consciousness and ethical principles through cooperative behavior, evolution, ex-
perience, training, and pedagogy.
1 Introduction
As artificial intelligence capabilities increase and machines start to exhibit characteris-
tics that we previously thought to reside within the unique domain of humanity, the
question of machine consciousness becomes pressing. For example, as machines be-
come more intelligent, will they naturally become sentient? If so, what might be the
implications for the safety and welfare of the inhabitants of future societies? How
would consciousness affect their ability or willingness to discharge tasks? And what
obligations might we owe to such creatures?
To begin to answer such questions in a secure and protected setting, prior to the
unintended onset of consciousness in highly intelligent, powerful and autonomous sys-
tems, it is reasonable to consider efforts to facilitate evolution of machine conscious-
ness under controlled and contained conditions, and to carefully study the effect it may
have on both functional competencies and moral behavior. Such an approach also en-
ables interventions at an early juncture in the emergence of consciousness that predis-
pose to artificial phronesis [10]; that is, the quick and reliable instantiation of learned
ethical codes in a social context under personal tutelage. In this way, two goals are
achieved: (1) a proactive understanding of the potential effects of machine conscious-
ness on the benefit-risk profile of highly intelligent autonomous systems, and (2) a
greater likelihood that a sound moral compass will guide emergent behavior in such
systems.
Although the ability of machines to "wake up" remains speculative, it can be hypoth-
esized that lessons from the human evolution of consciousness can be leveraged to in-
still consciousness in machines. Although it is not possible to definitively cite an exact
watershed moment marking the onset of modern human consciousness among Homo
sapiens or its ancestors, it is likely that human consciousness evolved over time in an
embodied, and cultural context, in which individual agents equipped with the necessary
“wetware” needed to both compete and cooperate in exceedingly complex societies.
Such an assumption is consistent with extensive observations made by Julian Jaynes in
2
his work, The Origin of Consciousness in the Breakdown of the Bicameral Mind, even
if the neuroscientific basis for consciousness adduced in his theory lacks current sup-
port. Mimicking such an evolutionary trajectory within a condensed time frame in silico
prior to the construction of physical robots that embody features of consciousness may
be expedient. Apart from evolution, either virtual or physical, it may also be useful to
initially leverage neuromorphic architecture, multimodal sensing, specific types of re-
inforcement learning, and current theories of human consciousness in our efforts to
generate and study emergent phenomena in machines.
2 Autonomous, multitask-capable, powerful, highly
intelligent, and adaptive system (AMPHIA)
The endowing of an autonomous, multitask-capable, powerful, highly (or super-) intel-
ligent, and adaptive system (hereafter referred to as AMPHIA) with phenomenal con-
sciousness may temper the risks to inhabitants of future societies. Several publications
have made the point (see, e.g., [11; 12]) that consciousness may be required for empathy
and moral-decision making.
As an alternative to fostering acquisition of cooperative behavior and ethical princi-
ples in a social framework through evolution, experience, training, and pedagogy,
moral codes can be built into AI “from the ground up” at the operating system level.
Formalized logic enabling automation of the doctrine of double-effect (i.e., allowing
for the necessity of producing unavoidable harm for the greater good) has been pio-
neered by Bringsjord [4] but has yet to be advanced to a level that enables the handling
of a multitude of various and more complex ethical quandaries. A machine that func-
tions well in the limited context of the well-known "trolley problem" cannot address
the infinite disjunction of thorny moral dilemmas that humans not infrequently face.
Implementation of general ethical strategies in a machine (e.g., utilitarianism) repre-
sents a profound technical challenge given computational demands. Interest has
mounted recently in verification techniques that may define a “safety envelope” for the
behaviors of systems driven by neural networks, but universal standards for verification
of the potentially infinite suite of behaviors of an AMPHIA-like system do not yet exist.
All of the above suggests the potential importance of imparting ethical principles to a
phenomenally conscious machine through an experiential learning (and teaching) pro-
cess. Here, and in keeping with others cited above, we claim that if this moral education
is to stand the most excellent chance of success, a machine should be phenomenally
conscious.
Apart from the hypothesis outlined above, namely that phenomenal consciousness
in a machine may increase the likelihood of moral behavior, we must consider the po-
tential consequences of consciousness on other functional competencies:
The ability to adapt to changing environmental circumstances.
The capacity or “willingness” to discharge previously mastered tasks and responsi-
bilities, or execute useful pre-existing skills, either learned or pre-programmed.
The capacity to think with human-like imagination and creativity.
3
Creativity and imagination may also be linked to a propensity for flexible and gen-
eralized ethical behavior of the type we would hope to instill in a conscious machine.
For example, the capacity to imagine ourselves in another’s shoes forms the backbone
of empathy. Similarly, an optimally ethical solution to a complicated human or societal
dilemma may require creative thought, such as the concept of a “carbon tax.” Although
creativity is not restricted to the conscious domain, elements of phenomenal conscious-
ness influence creative output in humans, and full or adequate conscious awareness has
been experimentally shown to be essential to the creation of improvisational melody
[1]. Similarly, it is implausible that Proust would have been able to write In Search of
Lost Time if he was not phenomenally aware.
Creativity is, perhaps, most difficult to quantize in experiments but, as with the other
classes of endpoints, it behooves us to be as specific as possible in its measurement if
we are to assess the impact of phenomenal consciousness on creative prowess in ma-
chines. It is important to recognize that if creativity is evaluated as an endpoint, then it
should not be used as one of the operational criteria for consciousness, as that would
confound any analyses seeking to correlate consciousness with creativity.
3 Creation of a Substrate
If phenomenal consciousness is paramount to successful inculcation of an ethical com-
pass in machines, then we are faced with the question: How can we build or evolve
such a form of AI? As an initial substrate (i.e., Generation A) subjected to an evolu-
tionary paradigm, we can “jump start” the evolutionary process by starting with some
combination of neuromorphic architecture and formalisms that can be computationally
implemented. General examples are:
Cortical “sensory area” analogs;
“Claustrum analog” for possible amplification of salient cross-modal data and facil-
itation of the formation of integrated high-level “mental” constructs (see, e.g., [8]);
Models of self, environment, and conspecifics;
Memory – working, procedural, and long-term analogs;
Capacity for learning, including innovative forms of reinforcement learning that en-
able more “human-like” or imaginative/exploratory forms of learning (see, e.g., [6]),
and/or neural net architectures that support bidirectional processing that has been
hypothesized to underlie human perception (see, e.g., [13])
Natural language processing and means of communication –vocabulary relevant to
goals
Virtual embodiment, locomotion and capacity to directly interact with the environ-
ment and conspecifics in complex ways.
4 Evolution in a Robot Hive
If human history is to be taken as an example, it is likely that development in a social
context will be necessary for machine consciousness that bears any resemblance to
4
modern human consciousness. Human consciousness is an embodied phenomenon [14]
weaned on culture so, ultimately, physical embodiment may be required, although it
remains to be seen whether virtual embodiment might accomplish the same ends.
If emergent consciousness in machines takes root in a social milieu, it is essential to
bear in mind that, although evolution and cultural molding have been instrumental in
forging human phenomenological consciousness, an evolved notion of “self” and com-
plex social behavior does not guarantee a form of consciousness that buttresses ethics
remotely similar to our own. For example, among the highly socially organized Hy-
menoptera, within which selection pressure is exerted primarily at the group/colony
rather than the individual level, the division of labor results in a rigid caste system in
which profoundly altruistic female workers are distinguished from reproductively ac-
tive females, and males are expendable after mating [5]. Such "valuation" of life is
incompatible with human ethical conventions.
Differentiation in evolved robot societies or hives (i.e., into phenotypes fit for vari-
ous tasks) should be possible and may be expected based on characteristics of highly
evolved and evolutionarily successful hymenopteran organizations (certain species of
social ants and bees). Such differentiation might also enhance resilience in the face of
malicious attacks. For example, if one phenotype is obliterated during a malicious
cyberattack as a result of a unique vulnerability, other evolved phenotypes might re-
configure under group selection pressure to take on the task(s) of the functionally an-
nihilated subgroup/phenotype.
One might endow “Generation A” with a significant degree of potential functionality
(see Section 3 above) as a “jump start” in the evolutionary process to mitigate project
risk. In addition, one can envision a two-stage process whereby initial evolution in a
virtual environment is replaced, after initial gains have been made, by later evolution
in a brick and mortar world rife with challenges, opportunities, and conspecifics.
Finally, as suggested in the introduction section, there may be advantages to the
gradual inculcation of ethics through learning in a societal context, mentorship, and
pedagogy, much as children are taught the principles of moral behavior.
5 Measurement of Cognitive and Phenomenal
Consciousness in an Experimental Setting
At multiple points during evolution, artificial agents can be followed for sociogenesis,
types of communication, and subjected to a battery of tests to assess for the presence of
emergent consciousness. These might include assessments of both cognitive and phe-
nomenal facets of consciousness. The ability to administer some specific tests would
depend upon a natural language interface. The latter would also be desirable in case a
form of consciousness is generated equivalent to the human "locked in" syndrome (we
would thereby be more easily alerted to the presence of underlying consciousness). The
list below is not meant to be final.
5
5.1 Cognitive/Functional Consciousness
We are well on the road to a "cognitively" or "functionally" conscious robot that can
model self and environment, reason, sense (in a sterile experience-free way), learn, and
behaviorally mimic aspects of consciousness. In essence, a functionally conscious ro-
bot would come close to what we consider conscious, in that it passes the mirror test
(this has already occurred), can model its interaction with the environment (this has
already been accomplished), behave intelligently in an autonomous fashion, form ab-
stractions, plan, learn from mistakes, and even, potentially, follow some basic ethical
rules (with the proviso that it would be extraordinarily difficult if not impossible to
enable a "functionally conscious" robot to deploy the flexibility of humans in dealing
with an infinite number of potential moral quandaries).
5.2 Phenomenal facets of consciousness and self-awareness
This has been traditionally defined as the appreciation of the “redness red,” although,
truthfully, consciousness is built from multimodal sensory input, integrated, rich, and,
frequently, emotionally laden. The tests below may interrogate phenomenal conscious-
ness and self-awareness, and we include a number of them because no one test is 100%
sensitive and specific. The manner in which we consolidate the results of such a battery
of tests would require deep thought. A Boolean concatenation is possible, or some
weighted average of the components. Validation of the measure in humans and select
animal species will be necessary, although the problem is thorny as, ultimately, the
assured attribution of both phenomenal consciousness and self-awareness technically
requires a (currently non-existent) objective and agreed-upon “gold standard” against
which test results are validated. Unfortunately, from a scientific standpoint at least,
phenomenal consciousness is uniquely subjective. With this caveat in mind, here are
listed some potential metrics:
Quantitative measures of “information integration”
Emotional intelligence;
Theory of mind;
“First-machine” accounts of its phenomenal experience [9];
The mirror test;
Behavioral capabilities suggestive of consciousness, akin to the “sensorimotor” ver-
sion of the Turing test: for example, the ability to spontaneously “imagine” what it
is like to be another creature and behave like such a creature;
A machine that is apparently "interested in" or "wishes to" explore altered states (by
altering its own hyperparameters or injecting noise) for no functional reason apart
from what seems like "curiosity" and "desire."
6 Proposed experimental design
It is critical to compare the functional capacities and moral/ethical behavior and dispo-
sitions of groups/subgroups. This consists of both a historical comparison within the
6
context of the study and parallel prospectively followed groups and subgroups as out-
lined below:
1. Control Group: Intelligent autonomous systems lacking any form of consciousness
(base case or “Generation A” with no subsequent evolution).
2. Evolved intelligent autonomous systems with or without efforts to engender artificial
phronesis
As a possible second control group, we may consider neurally controlled virtual Ani-
mats [2]. Virtual evolution of Animats has been shown to correlate with increasing
levels of information integration [3], a postulated marker of consciousness.
The base system (intelligent autonomous system without consciousness) would be
capable of accomplishing a set of pre-specified tasks/goals. The proposed framework
suggests the step-wise introduction of elements or groups of factors listed in Section 3
above. It is possible but doubtful, that phenomenal consciousness will arise spontane-
ously from the substrate alone in the absence of evolution in a social context. Once all
planned elements of the substrate have been introduced over time, the next phase of the
experiment would commence (evolutionary phase). We hypothesize that phenomenal
consciousness may evolve/emerge during successive generations using such a strategy.
The type of evolutionary programming would need to be carefully selected.
In parallel with monitoring for potential evidence of the development of conscious-
ness, one could observe for instances of selfless or altruistic behavior. If it is discovered
that emergence of proto-consciousness (e.g., primitive forms of phenomenal conscious-
ness) unexpectedly correlates with undesirable attributes (e.g., extreme behavioral vol-
atility with the destruction of other robots), then appropriate steps can be taken. Vari-
ants of the trolley problem have been evaluated recently as means of probing the per-
ceived relative moral value of self- vs. other-sacrifice [7]. Conceivably, machine-
adapted versions of such a test could be administered in the future.
7 Conclusions
This paper presented a structure for the study of potentially conscious AI systems. The
main idea at the basis of the framework is to facilitate the evolution of an AMPHIA
under controlled conditions to study the effect it may have on both functional compe-
tencies and moral behavior, as well as the potential benefit of artificial phronesis. Some
of that evolution occurs in silico (virtual), with the option of transitioning to evolvable
hardware within which functionally useful software, identified in earlier phases of
study, has been embedded.
During the evolution of the system, it is of great importance to consider the ethical
obligations that might be owed to any system itself that appears to demonstrate evi-
dence of phenomenal consciousness and, especially sentience.
While a fundamental tenet of this paper is that social interaction is one element nec-
essary to the evolution of consciousness resembling that of (neurotypical) contempo-
rary humans – or, perhaps, that of certain other social animals – the existence of con-
sciousness among less social creatures, such as polar bears, some larger feline species,
7
opossums, armadillos, and other species cannot be excluded. Yet, even animals such
as these (for example, bears) are not completely asocial. A discussion of solitary animal
consciousness is beyond the scope of this essay but such forms of consciousness are
likely to be quite different from that of humans. Yet, the same could be said of machine
consciousness, even if it does come to pass through evolutionary, social, and cultural
mechanisms analogous to those that have imbued contemporary humans with our brand
of consciousness and selfhood. The extent to which social interaction underwrites ma-
chine consciousness could be studied through social privation, but the ethical justifica-
tion for such an experiment in the context of “valid machine consciousness” would be,
in the author’s opinion, unsupportable. Identifying valid machine consciousness may,
however, be challenging, particularly if that conscious self is endowed with high-level
mental constructs foreign to our own, informed by raw phenomenal experience tethered
to sense modalities we can only imagine. The extent of the overlap in the Venn diagram
(human versus machine consciousness) remains to be seen, but the degree of that inter-
section will likely be related to the scope of interaction between humans and sentient
machines, embodied similarities, and the ethical tutelage we provide.
Acknowledgments
Author would like to thank the participants at the series of workshops on Technology
& Consciousness organized in 2017 by SRI International, and Antonio Chella for his
editorial feedback. He also thanks the reviewers for their feedback and valuable sug-
gestions.
References
1. Baumeister R., Schmeichel B., and Dewall N. Creativity and Consciousness. In: P. Kaufman
(Ed.), The Philosophy of Creativity (Oxford University Press, 2014).
2. DeMarse T., Wagenaar D., Blau A., Potter S. “The Neurally Controlled Animat: Biological
Brains Acting with Simulated Bodies.” Auton Robots: 11 (2001): 305–310.
3. Edlund J., et al. Integrated Information Increases with Fitness in the Evolution of Animats.
PLOS Computational Biology: 7 (2011).
4. Govindarajulu N., Bringsjord S. “On Automating the Doctrine of Double Effect.” IJCAI
(2017)
5. Hölldobler B. and Wilson E.O. The Superorganism: The Beauty, Elegance and Strangeness
of Insect Societies. Norton, 2008.
6. Machado M., Bellemare M., and Bowling N. “A Laplacian Framework for Option Discov-
ery in Reinforcement Learning.” Proceedings of the 34th International Conference on Ma-
chine Learning, Sydney, Australia, PMLR 70 (2017).
7. Sachdeva S., Iliev R., Ekhtiari H., Dehghani M. “The Role of Self-Sacrifice in Moral Di-
lemmas.” PLoS ONE: 10(6): e0127409.doi:10.1371/journal.pone.0127409 (2015). Availa-
ble from: https://www.researchgate.net/publication/279448498_The_Role_of_Self-Sacri-
fice_in_Moral_Dilemmas [accessed Dec 04 2018].
8. Sahner D., “Whither the Self: The Foundation of Consciousness and its Implications for
Poetics.” Journal of Consciousness Exploration and Research: 4 (2013): 856-873.
8
9. Schneider S. “Is Anyone Home: A Way to Find Out if AI Has Become Self-Aware.” Sci-
entific American blog (2017): https://blogs.scientificamerican.com/observations/is-anyone-
home-a-way-to-find-out-if-ai-has-become-self-aware/
10. Sullins J. “Artificial Phronesis and the Social Robot." In: J. Seibt et al. (Eds.), What Social
Robots Can and Should Do (IOS Press, 2016).
11. Torrance S. “Ethics and Consciousness in Artificial Agents.” AI & Soc: 22 (2008): 495–
521
12. Wallach W., Allen C., Franklin S. “Consciousness and Ethics: Artificially Conscious Moral
Agents.” International Journal of Machine Consciousness: 3 (2011): 177-192.
13. Xu D., Clappison A., Seth C., Orchard, J. “Symmetric Predictive Estimator for Biologically
Plausible Neural Learning.” IEEE Transactions on Neural Networks and Learning Systems:
29 (2018): 4140-4151
14. Zebrowski, R. (2010). In Dialogue with the World. JCS: 17 (2010): 156-172.