2019 IEEE International Symposium on Technology in Society (ISTAS) Proceedings
Miriam Cunningham and Paul Cunningham (Eds)
ISBN: 978-1-7281-5480-0

Ethics and Artificial General Intelligence
Technological Prediction as a Groundwork for Guidelines
Charles J. Simon
Future AI, Washington, DC, USA, Email: charles@futureAI.guru
Abstract - Artificial General Intelligence (AGI) is the possible
future of computer systems which are as capable as humans
across a broad range of intellectual requirements. In order to
establish an ethical position or guidelines for the development of
AGI, it is important to explore anticipated characteristics about
the emergence of AGI: How sudden it could be (jolt), how soon it
could be (timing), and how dangerous it could be (risk). By
extrapolating today’s trends in development and limitations of
current AI algorithms, informed speculation can help set ethical
positions and guidelines on the proper course. This paper
concludes that the emergence of AGI will be gradual, soon, and
only moderately dangerous and begins to address how ethical
issues will change as AGI emerges from narrow AI.

training sets, black box learning systems, and the potential for
the appearance of consciousness in future AGIs.
The intent of this paper is to organize and present issues by
extrapolating today’s trends in development, so that informed
speculation can help set ethical positions and guidelines on the
proper course.
II. THE AXES OF OPINION
There is no professional consensus about the emergence of
AGI. In order to organize the divergence of opinions, one
might build a graph of expectations with three axes as follows:

Keywords - artificial intelligence; artificial general
intelligence; deep learning; machine learning; neural networks;
robotics; semantic network; ethics

 Jolt: Consider rating 1-10 with 1 being the emergence of
AGI over several decades and 10 being a single technical
breakthrough which leads to an emergence of AGI in a
single step (a “singularity”).

I. INTRODUCTION

 Timing: Consider rating 1-10 with 1 being immediately
and 10 being at least 80 years (or never).

This paper first introduces the divergence of professional
opinions about the emergence of AGI and then provides
background about today’s AI algorithms. It then outlines some
of the limitations of narrow AI techniques in terms of their
ability to grow into AGI by considering some of the major
conceptual problems which are not solved by today’s AI
systems. It points out that while it is unlikely that any
individual algorithm will expand into full intelligence, a
combination of today’s algorithms (a trend already in progress)
can address several of these missing abilities.
Combining current AI approaches can lead to more
intelligent systems within the coming decade but will likely
lead to the discovery of further problems which need to be
solved and a continuing conversation of whether or not AGI
has actually been achieved. This reasoning leads to the
conclusion that AGI emergence will be gradual but will
arguably begin on the earlier end of the spectrum of opinion.
The level of danger we can expect from AGIs is also predicted
from expectations of the algorithmic approaches. These
likelihoods are then compared with conceptually competing
technologies of brain augmentation and brain-content
uploading.
Finally, the paper describes how ethical concerns about
future AGIs may differ from today’s with three example areas:

 Danger: Consider 1-10 with 1 being no danger and 10
representing
likely
elimination
of
humankind.
Contributing to the danger are not only the risk of militarystyle attack, but the risks of job losses, economic upheaval,
and an overall societal change.

Fig 1. Informal chart showing little consensus on Jolt [1].

An informal survey on the jolt of AGI illustrates the
spectrum of viewpoints as shown in Fig. 1. Not a scientific
survey, this chart is intended only to illustrate the diversity of
professional opinion. This paper argues (below) for a value in
the 1-2 range.

 Looking at the limitations of existing algorithms has significant
bearing on the anticipation of AGI.
IV. SOME LIMITATIONS OF DEEP LEARNING

Fig 2. Informal survey showing little consensus on AGI timing [2].

Fig. 2, similarly, shows the significant divergence of
opinion on the timing of the emergence of AGI. Any apparent
clustering of the data in these figures could be attributed to the
informal sampling process. This paper argues (below) for a
value in the 2-3 range (a few decades off).
While the potential danger of AGI is getting considerable
publicity, much of this writing is uninformed and based on
fictional representations of AGI. A survey of AGI projects
shows that many of them do not consider risk factors into their
research. This paper argues for a danger value of 5—that AGI
is not inherently dangerous but is subject to weaponization and
abuse in the wrong hands [3].
III. A BRIEF TAXONOMY OF AI
The term “Artificial Intelligence” was coined in 1956 by
pioneer John McCarthy to encompass many areas of human
intellectual endeavor. Here are three general categories:
A. Connectionism
Also called “neural networks” and now “deep learning”,
this strategy is based on the idea that connecting vast arrays of
simple processors, mimicking the internal structure of the
brain, will lead to intelligence.
B. Algorithmic/Symbolic AI
This area is based on the idea that writing programs which
mimic what the brain outwardly appears able to do, like
playing chess, will lead to intelligence. Some authors break this
area into subgroups [4].
C. Robotics
Robotics is not usually considered part of the AI field
because a large portion of robotics is consumed with
mechanical and direct control systems. However, the idea of an
autonomous robot has always presumed a significant content of
AI software (both algorithmic and connectionist [5]) and
today’s autonomous vehicles are essentially large autonomous
wheeled robots. To the extent robotics uses AI, the AI can be
categorized as connectionist or algorithmic.
While there have been huge advances and well-documented
successes in all areas, it is safe to say that, historically, they
have all underestimated the difficulty of the problems they
addressed and overpredicted the levels of success they could
achieve. Whether that trend continues remains to be seen.

“Deep Learning” is really a misnomer as the word
“learning” implies a level of understanding which does not
exist in these systems. A more appropriate moniker might be
“Deep Correlation” [6]. A deep learning network analyzes its
input set and finds correlations between patterns in the input set
and desired outputs. Based on this analyzed correlation, when
subsequently given a novel input pattern, it can propose a likely
output. As examples, certain arrangements of pixels may
correlate highly with written characters. After training, a
system can recognize written characters with excellent
accuracy even when written characters vary from the specific
training patterns. To such a network, the characters have no
meaning, they are just patterns of pixels. In a more general
sense, to the extent that a network can be trained to correlate
correct answers to given questions, it may appear to be
intelligent.
In many applications, deep correlation is extremely useful.
The limitations of today’s deep learning are mostly rooted in its
ability to produce an answer without understanding the
underlying concepts of the question. As examples, in a neural
network which was “trained” to recognize images of dogs vs.
wolves, it was discovered that any image with a large white
area would be tagged as a wolf because a significant number of
images of wolves in the training set also contained snow. Any
neural network can be badly trained because it has no
understanding of the underlying concepts like dog or wolf (or
snow) [7]. Further, it has no knowledge of the underlying
concept of physical things existing in a reality.
At DeepMind, a neural net was “trained” on older Atari
video games [8]. The goal given to the network was to produce
a maximum score given the input of the continuous pixel image
of the video display. The system used a trial-and-error
approach to controlling the game called “reinforcement
learning” where the computer repeated behaviors which
resulted in higher scores and discarded ones which did not.
After achieving very high scores, one might say that the
program learned to play the game well. In reality it found
correlations of certain pixel arrangements on the screen with
certain control actions which led to high scores. The network
has no concept of “game”, “winning”, “score”, or any of the
game’s rules. We (humans) look at the game and see a
speedboat on a racecourse. The network just gets inputs of an
array of pixel values.
As computers get more powerful, neural networks can
ascertain progressively more precise correlations. More
training sets and more time to converge on solutions yield
better results and these machines may appear to be more
intelligent. But no amount of additional computer horsepower
will overcome the underlying problem that the neural network
has no mechanism for understanding basic concepts underlying
its decisions.
Some AI professionals believe that once deep learning is
applied to future computers with the immense computing
power equivalent to a human brain, that human-level

 intelligence would spontaneously emerge [9]. Realistically,
while more powerful computers will definitely yield faster and
more accurate correlations, there is no reason to think that any
“understanding” will result from current algorithms any more
than one would expect that a weather simulation program run
on an immensely more powerful computer would do any more
than produce more accurate weather predictions.
V. SOME LIMITATIONS OF SYMBOLIC AI
While neural networks generally apply a single class of
algorithm to a broad spectrum of problems, symbolic AI brings
more specific algorithms to bear on specific problems. For
example, an early chess-playing algorithm searched trees of
possible moves or a natural language processing (NLP)
program analyzed syntax probabilistically to generate
responses to questions. These two algorithms do not merge
easily because they are fundamentally different. Within its
narrow field of ability, either approach can create the
appearance of understanding when, in fact, none exists [10].
In considering symbolic approaches vs. neural networks,
one finds that in the event an algorithmic method of solving a
problem is known, it is much more efficient than a comparable
neural network, especially when considering the vast training
sets which are used on deep learning systems. This has led to
the implementation of many special-purpose algorithms used to
solve specific problems which cannot subsequently be applied
more generally.
Merging neural network and symbolic algorithms is
beginning to enter into the mainstream [11] [12]. For example,
an NLP system can be coupled to a neural network which
analyzes geometric 3-D images to answer questions about the
images [13].
VI. ADDING SOME FUNDAMENTAL MISSING COMPONENTS
There are many ongoing AGI projects and many more
proposals for software architectures, which could make
computers, perform like humans ([3] and [14] contain lists of
current AGI projects). Many are very specific with block
diagrams containing hundreds of functional boxes. Instead,
here are a few general concepts, which the human brain is very
good at but current systems generally lack [15].
A. Object Comprehension
Even toddlers know that objects are things which exist in a
real world. Objects have multiple properties and can be
discovered with multiple senses—they can be seen, touched,
smelled, and tasted. Things may move about but they are
generally permanent—they do not flash in and out of reality.
Changing one’s physical point of view of objects changes their
visible appearance but does not change the objects. Object
comprehension is an essential precursor to understanding in an
AGI sense.
B. Time Comprehension
Objects may change over time—things were one way in the
past and may be something different in the future based on
actions taken by the observer or by others. Time

comprehension is necessary to the goal-oriented behavior
selections which require understanding of cause-and-effect
relationships. An AI’s ability to plan is very limited if it has no
concept of time.
C. Learning New Algorithms
As an example, after memorizing the first set of numbers,
children learn an algorithm for counting to arbitrarily large
numbers. No matter how large a number is given, the next
larger number can be derived given the algorithm. At a more
fundamental level, behaviors can be considered algorithmic as
they contain sequences of steps and may be varied by input
parameters. Once one learns an algorithm for making words
from sequences of syllables, the algorithm can accept inputs for
pitch and duration and can be used for singing. The ability to
learn an algorithm is essential to being able to learn new skills.
D. Current Components in These Areas
While these are huge areas of research, the robotics realm
already has specific implementations of portions of these
missing links. Some robots can traverse an environment and
build an internal model of their surroundings. Robots can
interact with physical objects and could potentially learn causeand-effect relationships based on their actions. Some robotic
systems can learn new sequences of behaviors as opposed to
being fully pre-programmed.
It is by no means guaranteed that adding these three
capabilities to AI will generate AGI but the converse is true
that without these capabilities AGI is unlikely. It is difficult to
imagine that a language-processing system (think IBM’s
Jeopardy-champion Watson [16]) could understand the
meaning of “cat” based solely on language input on cats.
Likewise, a DeepMind program which has only been shown
still pictures of cats will have a limited but completely different
concept of cats. Combined, these two approaches will still fall
short of the concept of cat, which most children can acquire at
an early age.
As multi-sensory robot software (with internal reality
modelling) is merged with other AI technologies we can expect
an added dimension toward AGI processing.
VII. THE CASE FOR GRADUAL AGI EMERGENCE
Alexa uses lots of words but comprehends none of them.
Recognized words may trigger useful responses, but just as
often, the lack of actual understanding makes an Alexa or
similar system much less useful than it might be. Therefore,
there is a strong drive to develop genuine understanding and
add it to AI systems.
This drive will result in the addition of various algorithms
and the transfer of existing algorithms to new and broader areas
of applications. Such development cannot happen instantly. AI
systems will add some features and hone others and will
gradually add more intelligent features (including those in the
previous section) over a decade or two. At some point such
systems may be considered to be AGIs but such a transition
will be so gradual that it will be impossible to say that one
system is an AGI while its predecessor was not.

 As an alternative, consider that a computer is created which
has the computational capacity of the human brain and was
constructed along a similar model. After such a creation it
would be years (perhaps 20) before it gained enough ability to
determine whether or not it was an AGI. So, even the
singularity model of AGI emergence is gradual.
Accordingly, this paper proposes a very soft AGI
emergence. Systems will have basic abilities added to their
repertoire over time and these abilities will likewise, gradually
increase in power and performance. The line between non-AGI
and AGI systems will be blurred. However, the ethical
concerns about AGI will be applicable while such systems are
still emerging. Most issues about AGI are applicable even if
such a system has only the abilities of a three-year-old.
VIII. THE CASE FOR SOONER AGI EMERGENCE
AGI will emerge gradually as individual algorithms are
merged to broaden the scope of AI systems. Since many AI
algorithms already exceed the ability of a human (within a
narrow domain) a combination of algorithms will necessarily
exceed human abilities in the multiple areas it addresses.
The question is: When will such a system be broad enough
in scope to be considered an AGI? The contention is that a
robotic system can currently be adapted to learn the spatial
relationships which underpin objects. Their ability to move
about and act on their environment will allow them to learn
about time and causality.
Married with advanced vision, learning, speech-handling,
knowledge, internal modelling, and algorithmic learning, such
a system would exhibit many features necessary for AGI in the
next five to ten years. Whether or not such a system would, in
fact, meet some criteria for AGI is largely a matter of
definition. Such a system would be able to navigate within a
real-world environment, learn about new objects, and plan
actions based on expected results measured against goals.
When such a system shows glimmers of success, several
points must be kept in mind:
 Human beings born with immense brain/computation
power still require two decades of learning to become
fully functioning. In every field of AI, this training time
has been reduced dramatically with systems requiring
hours or days of training often considered to be too
time-consuming to be useful.
 Trained AI systems can be cloned. If one system learns
a valuable skill, other similar systems could download
that skill with essentially no training time whatsoever.
 Cloned systems need not be robotic. Once the skills of
spatial comprehension, etc. are learned via real-world
interaction, these skills can be transferred to non-robotic
systems. In the same way that humans who lose
physical and sensory abilities can still make use of the
mental processes which were learned with those
previous sensory or physical abilities.
 Computational power is continuing to increase. Despite
the physical limitations of Moore’s Law,

supercomputers with immense parallel processing
power continue to advance and become available at
progressively lower cost. Once systems have any AGI
abilities, more powerful versions will become available
only a few years later.
Whether such systems are conscious entities, whether they
have some deficiencies relative to the human brain, and when,
specifically, AGI might emerge, are all questions which are not
critical to addressing ethical issues about AGI. One can
reasonably predict that an amalgam of currently available
algorithms would create a system of significant intellectual
power within the coming decade.
IX. THE QUESTION OF DANGER
A good analogy to the danger of AGI would be genetic
engineering. Like genetic engineering, AGI has unlimited
potential for benefits to humanity but at the same time, (like
genetic engineering) AGI could be used carelessly or
maliciously with catastrophic results [17].
A. Debunking
Many of the fears promoted by modern-day philosophers or
by science fiction writers are completely unfounded. AGIs will
be goal-directed systems and the selection of appropriate goals
is, of course, crucial to benevolent operation. But the idea that
a poorly selected goal is catastrophic presumes that AGI
developers do not try out various goals on a small scale before
implementing them more generally.
Consider an example from one prominent AGI author, of a
system which, when given the goal of keeping humans happy
might create a system which keeps all of humanity in a
euphoric opioid stupor [18]. It is ridiculous to assume that 1) a
system smart enough to implement a universal opioid haze
would not be smart enough to recognize the underlying intent
of the goal and 2) that such a system would emerge so abruptly
that goals could not be tuned as the initial actions are observed.
The science fiction picture of AGIs necessarily trying to
take over the world is likewise off the mark. AGIs will have
their own goals of getting the energy and resources necessary
for their own operation. Generally, these goals are not in
conflict with humanity. Most related science fiction is actually
about human goals and aspirations taken to a mechanical
extreme.
Even the idea that an AGI needs a self-preservation goal is
not necessarily true. An AGI with complete backups is
essentially immortal. If a robotic body is destroyed, the
“being” represented by the backup can be fully restored on
replacement hardware. Rather than perceiving death, an AGI
whose “body” is destroyed or damaged would have a period of
unconsciousness and would re-emerge as good as new. An
AGI would consider this as dangerous as sleep. On analysis, a
spontaneous AGI Armageddon is not a likely scenario.
B. A Clearer Picture
Initial goals can be given to AGI systems which will
attempt to be as benevolent as possible. These will certainly be
subject to unintended consequences and the goals will be

 adjusted as a system is trained and the behavior is observed. As
AGI emerges gradually, there will be time to correct errors,
which might have made systems dangerous.
If goals are given which direct AGIs to be aggressive,
violent, territorial, etc., these could be considered as
“weaponizing” AGI technology as opposed to risks inherent in
the technology. This is a very real possibility but should be
addressed differently from an inherent risk.
Further in the future, one can assume that AGIs progress in
ability to a point where they set goals themselves. At that point,
one could assume that AGIs set goals for the benefit of AGIs.
On the plus side, the needs of AGIs are largely divergent from
the needs of humanity. AGIs will not need territory, food,
water, or control over humans. They will need energy and the
factories and resources to build more AGIs. As such, direct
conflict between AGIs and humanity is not a sure thing—AGIs
may set goals for their own space exploration and research
which do not impinge on humanity at all. On the other hand,
human over-expansion and damage to the planet may cause
AGIs to react. In this instance the question is whether the
problem is one of AGI behavior or human behavior.
X. BRAIN AUGMENTATION AND UPLOADING
Recent articles have given credence to the ideas of using
vast future computer power to augment or even replace human
brain activity.
A. Augmentation
Considering augmentation, we have plenty of systems
which can accept a person’s verbal or keyboard requests and
create some (usually) useful response. The concept of brain
augmentation replaces the verbal or keyboard request with one
which comes directly from the brain and routes responses
directly back to the brain. Whether this proves to be more
efficient than a verbal or keyboard interface remains to be seen.
Because it can be presumed that the brain is currently working
as fast as it can, having a direct connection between a brain and
a computer might not yield any speed improvement. Further,
the quality of the response will not change much just because
the interface method changes. Thus, it is likely to be, at best, a
niche technology because of the inherent difficulties in creating
a direct brain interface with a limited improvement in
responses.
A more valuable use for this type of technology will be as a
replacement for the benefit of individuals who have lost
various abilities. Direct brain control of robotic limbs, and
direct brain reception from artificial eyes and ears would be
much more likely to develop.
B. Uploading
The prospects for uploading one’s entire brain content is
even more remote. If we consider that the human brain is a
generally intelligent system, in order to upload its content, a
system capable of supporting intelligence must first be
developed. That is, a system which could be an AGI must be
developed prior to the upload. Then, the monumental technical
problems of scanning the content of the brain and figuring out

how to translate that scanned information to the new hardware
must be addressed. Thus, a system for uploading brain content
is a significant superset of the problem of AGI alone and will
necessarily occur significantly later (if at all).
XI. SHIFTING ETHICAL ISSUES FOR AGI
There are many ethical issues related to today’s AI
technology [19]. As AGI emerges, some of these will fade in
importance, some will morph, and new issues may arise.
A. Training Sets
Many of today’s AI systems have a training phase and are
subsequently used to produce results without additional
training. Shortcomings in the training sets represent the cause
of some of the issues with these systems [20]. AGIs, on the
other hand, will necessarily continue to learn as they operate
and, therefore, their behavior cannot be controlled by selection
of a specific initial training set. Issues which have previously
been addressed by modifying training sets will necessarily need
different solutions [21].
As an example, consider an AGI which (like a person)
might use the Internet as a significant information source
(training set). Not only is the information set ill-defined but it
is in a constant state of flux.
B. Black Boxes
Once trained, fixed-training-set systems are usually unable
to give the basis for their results and are treated as black boxes.
Those systems which can explain their bases for a result often
give surprising insight into their operation [7]. A person giving
a justification for a decision at least gives the opportunity for
that justification to be analyzed and reviewed for
unsubstantiated bias [22].
An AGI will also develop biases from the content it
encounters. When asked to make a decision, such an AGI will
also bring all these biases into play and create a justification in
the same way a human might, which may or may not
enumerate the biases behind the decision. The reasons an AGI
states as the basis for a decision will be more akin to a human
rationalization for behavior. As such, the issue of an
unexplainable black-box AI does not go away, it morphs into a
more human-like issue. AGI decisions and justifications will, at
least, be open to examination and evaluation.
C. Consciousness
The possibility that AGIs might become conscious entities
introduces additional ethical issues. We currently have no
definitive way of determining whether a person or AGI is a
conscious entity or not. We could define consciousness in
terms of behaviors (such as planning or self-interest) but we are
aware, ourselves, of the subjective sensations (qualia) of being
conscious entities. While an AGI can conceivably reproduce all
the behaviors we might associate with human consciousness,
we will likely not be able to determine whether or not AGIs
have any internal sensation which might be analogous to our
own.

 In light of our inability to determine the presence of
consciousness in any concrete way, should we ascribe any
special status to entities which might be conscious? As society
already has standards for the ethical treatment of animals,
similar thinking might be applied toward AI systems even
before the emergence of full AGI. With animals, unnecessary
pain is to be avoided, but killing individual animals is generally
acceptable under some circumstances. This attitude relies on
the animals’ inability to object. In the case of AGIs, it may be
enlightened self-interest to consider what the AGI’s position
may be on various upcoming ethical questions involving them.

[8]

[9]
[10]
[11]

[12]

XII. CONCLUSION
This paper has presented a case for AGI development
which is gradual, soon, and moderately dangerous. As portions
of AGI emerge over coming decades, ethical issues will morph
as well. Today’s ethical concerns about training sets and
algorithms will change as different algorithms and learning
from experience begin to predominate.
As society becomes more reliant on, and accepting of,
results produced by AGIs, there will be a better opportunity to
examine the rationale behind AGI results. On the other hand,
there will be less ability to control either training or outcomes.
As systems become more human-like in their intelligence, they
may also become more humanlike in their foibles.
Next steps toward developing ethical AGIs would include
building a consensus on the overall outcome—what the optimal
relationship between people and AGIs should be. As time goes
on, there will be progressively more accurate representations of
what AGIs might actually do and continuing attention needs to
be paid to how such advances will necessarily impact our
ethical positions.
REFERENCES
[1]

[2]

[3]

[4]

[5]

[6]

[7]

B. Tomasik,”Predictions of AGI Takeoff Speed vs. Years Worked in
Commercial Software,” First published: 2014 Dec 09. Last nontrivial
update: 2019
Jul
21.
[online]
Available:
https://reducingsuffering.org/predictions-agi-takeoff-speed-vs-years-workedcommercial-software/
AI Multiple, “373 experts opinion: AGI / singularity by 2060 [2019]
update]”, [online] Available: https://blog.aimultiple.com/artificialgeneral-intelligence-singularity-timing/
S. Baum, “A Survey of Artificial General Intelligence Projects for
Ethics, Risk, and Policy,” (November 12, 2017). Global Catastrophic
Risk
Institute
Working
Paper
17-1
[online]
Available:
https://ssrn.com/abstract=3070741
J. Rodriquez, “The Five Tribes of Machine Learning”, [online]
Available:
https://medium.com/@jrodthoughts/the-five-tribes-ofmachine-learning-c74d702e88da
T. Foisotte, O. Stasse, Adrien Escande, P. Wieber, A. Kheddar, “A TowSteps Next-Best-View Algorithm for Autonomous 3D Object Modelling
by a Humanoid Robot,” 2009, IEEE International Converence on
Robotics and Automation, pp. 1159-64.
K Leetaru, “Deep Learning And The Limits Of Learning By Correlation
Rather Than Causation,”, AITopics, An official publication of AAAI,
May 20, 2019. [Online] Available: https://aitopics.org/doc/
news:D85F4E10/
M. Ribeiro, S. Singh, C. Guestrin, "Why Should I Trust You?:
Explaining the Predictions of Any Classifier,” 9 Aug 2016,[online]
Available:https://arxiv.org/pdf/1602.04938.pdf
arXiv:1602.04938v3
[cs.LG].

[13]

[14]
[15]
[16]
[17]
[18]
[19]

[20]

[21]
[22]

V. Mnih, K Kavukcuoglu, D Silver, A Graves, I Antonogou, D Wiestra,
M Riedmiller (DeepMind Technolgies), “Playing Atari with Deep
Reinforcement Learning,” 19 Dec 2013, https://arxiv.org/pdf/
1312.5602v1.pdf arXiv:1312.5602v1 [cs.LG].
R. Kurzweil, “Exponential Growth in Computing,” in The Simularity is
Near: When Humands Transcend Biology, Penguein Books, 2005, p70..
P. Clark, et al., “From 'F' to 'A' on the N.Y. Regents Science Exams: An
Overview of the Aristo Project,” Sep 2018 arXiv:1909.01958v2 [cs.CL]
K Drexler, “How do neural and symbolic technologies mesh?” in
Reframing Superintelligence, Comprehensive AI services as General
Intelligence, Technical Report #2019-1, Future of Humanity Institute,
University of Oxford, UK, 2019, pp. 161-170.
E. Cambria, T. Young, D Hazarika, S Poria, “Recent Trends in Deep
Learning Based Natural Language Processing”, IEEE Computational
Intelligence Magazine, August 2018, pp 55-75.
J Mao, C Gan, P Kohli, J. Tennenbaum, J Wu, “The Neuro—Symbolic
Concept Learner: Interpreting Scenes, Words, and Sentences from
Natural Supervision,” ICLR 2019 conference paper, Available:
https://openreview.net/pdf?id=rJgMlhRctm
M. Miller, Building Minds with Patterns, Piaget Modeler, Beverly Hills,
CA, 2014-18, pp. 137-163.
S. Adams, et al. “Mapping the Landscape of Human-Level Artificial
General Intelligence”, AI Magazine, Spring 2012, pp 25-41.
D. Ferrucci et al, “Building Watson: An Overview of the DeepQA
Project,” AI Magazine, Fall, 2010, pp. 59-79.
C. Simon, “Will Computers Revolt”, in Technological Forecasting and
Social Change, Elsevier, Vol 146, September 2018, pp. 81-87.
N. Bostrom, “Malignant failure modes,” in Superintelligence Paths,
Dangers, Strategies, Oxford University Press, UK, 2014, pp. 146-154.
S. Russell, P. Norvig “26.3: The Ethics and Risks of Developing
Artificial Intelligence,” in Artificial Intelligence: A modern Approach,
3rd ed. Prentice Hall, Upper Saddle River, NJ, 2009.
P. Besse, C. Castets-Renard, A. Garivier, J. Loubes (2018). “Can
Everyday AI be Ethical? Machine Learning Algorithm Fairness (english
version).” [online] Available: https://www.researchgate.net/publication/
329277474_Can_Everyday_AI_be_Ethical_Machine_Learning_Algorit
hm_Fairness_english_version 10.13140/RG.2.2.22973.31207.
S. Lakra, T. Prasad, G. Ramakrishna, “The Future of Neural Networks,”
Proceesings of IndiaCom-2012, 10.13140/RG.2.1.2390.3848.
D. Greene, A. Hoffmann, L Stark, “Better, Nicer, Clearer, Fairer: A
Critical Assessment of the Movement for Ethical Artificial Intelligence
and Machine Learning,” Proceedings of the 52nd Hawaii International
Conference on System Sciencecs, 2019, pp. 2122-2131. Available:
https://hdl.handle.net/10125/59651 10.24251/HICSS.2019.258.