What is Artificial Intelligence? An examination of a Google
search will list thousands of definitions, many convoluted and circular, namely
defining intelligence as intelligence. As we have noted elsewhere, the problem
of not having a clean and clear definition makes it impossible to create laws,
yet this never seems to stop Governments, resulting of course in endless
litigation and confusion. Our intent herein is not to define AI per se, since
we believe that at best it is a work in progress and at worst the wrong words
to begin with, but to present some paradigms and elements which may prove
useful.
In a simplistic sense, AI takes some input that is to be examined and provides an output to the putative question provided in the
input. It does so by relying on a massive amount of exogeneous information that has been processed by an element called a
neural network (NN) for example. The NN has been designed and trained so that
any input aligned with the class of trained data can or should produce an
answer. Some answers can be presented simply as yes or no, and others more
complex and in a text form using a natural language processing system as an
adjunct.
Another simple example is shown below. Here we take a
pathology slide, not even identifying it by organ, and we seek to identify by
organ and malignant status. The input
is an image and the output
is a classification of N possible organs and M
possible states. The system has been “trained” with potentially millions of
identified images.
However what AI has in common is a form of “learning” from
prior data sets and then developing algorithms on handling new data demands to
provide answers or actions. What we see is that AI is a concatenation of
inputs, data sets, learning algorithms and output
mechanisms. In the simplest sense, on can ask
a question and receive an answer, if the data set contains the data adequate
for learning.
We examine here the potential extensions of this set of
constructs. AI can go from the simplest input/output paradigm to a fully autonomous entity that initiates interactions, gathers
information, constructs mechanisms, and
provides actions while continuously monitoring its own performance, seeking
increased optimization.
The putative “danger” of an AI system lies in the realm of the autonomous AI entity
(AAIE) embodiments. Namely, an AI entity totally independent of any human
interaction. Namely, it begs the question; can an AI system become totally
independent of any human agency? If so, then what limits can be placed upon its
actions? What can be done to enforce such limits?
We have a clear example of unenforced limits in a small
sense with COVID-19. A virus released into the society and its propagation
facilitated by an unprepared set of Governments resulting in the death of
millions and a near collapse of economies. Autonomous AI systems are many orders of magnitude more
deadly to humanity as a whole.
Our objective herein is to examine AI systems and
specifically to consider canonical models demonstrating the putative
progression to a fully autonomous AI entity, one capable of independent
actions both computationally and physically. The latter model we call the
Autonomous AI Entity, AAIE. This is an entity that operates independent of
human interaction and makes judgements on its own. Further it has the capability
of using and assembling instruments as externalities to effect its intentions.
We often hear about the fears of AI devoid of any
specificities. In order to understand what the risks may be one must understand
what evolution can occur and what areas should be limited if
any. In many ways it is akin to bio research on new organisms. We know that
COVID is a classic example of bio-research gone wild.
Basically, the fundamental structure of AI as currently
understood is some entity which relies on already available information that is
used by some processing elements to perform actions. Now, in contrast to what
we have argued here, there is that this exogeneous Information set, provided by
humans, may become self-organizing in an autonomous mode entity. Namely as we
approach an autonomous mode this set of information may be generated by the
entity itself, and no longer reflecting any reliance on a human.
The neural net paradigm has been evolving for almost the
past fifty years. Simply stated the neural net paradigm assumes a computer
entity, that takes a massive amount of exogeneous information to train a
network, so that when some input entity is presented, it can produce an output
entity that correctly reflects the body of information available to the
computer entity. To accomplish this one needs significant amounts of
information, memory and processing. Thus, conceptually one had the structure
constructs yet it required the development and availability of memory and
processing power to take the steps we see today. Thus, NN are not new but only
constrained by technology.
In addition, the nature of inputs and outputs is also an
evolving area. For the output we may want some natural language processor and
say for the input the ability to gather and process images. In fact, the input
must eventually gather all types of entities; video, image, taste, smell,
touch, voice, etc. In fact, multimedia inputs and outputs will be essential.
We use the neural net construct as a place holder. One
suspects there may be significant evolutions in these elements. One need look
no further that what we have seen in the past 40 years. The driver for the
evolutions will be processing complexity as well as computing complexity. One
also suspects that there will be significant evolutions in memory for the
learning data.
Also, paradigms on human neural processing may open avenues
for new architectures. This is a challenging area of research. The biggest risk
we face is the gimmick constructs that are currently driving the mad rush.
The risk of autonomy was perceived in broader terms by Wiener in his various writings. The development of
AEs is the development of entities that can displace if not annihilate man. We
see that AEs can restructure their own environment and that control of AEs may
very well be out of the hands of their developers. In fact, the developer may
not even be aware of when such an autonomous act occurs.
On has always considered the insights of Shannon and his Information Theory and the broader
constructs of Wiener and Cybernetics. One suspects we are leaving
the world of Shannon and entering that of Wiener.
If AEs are to be considered
intelligent than how would we compare that to human intelligence. Would an AE
consider humans just an equivalent primordial slime, an equal, a superior, or
just some nuisance inferior species? Can we measure this or is it even
measurable.
The areas of greatest risk are legion in AI. They range from simple
misinformation, to psychological profiling, then influencing and controlling
large groups, and finally as full autonomy is obtained, the ability to manipulate their
environment.
Without some moral code or ethical framework, AEs can act in
whatever manner they so choose, often taking leads from the data input that may
have or create themselves.
There have been multiple lists of AI risks. The problem is that all
that have been generally available lack any framework for such listing. They
generally make statements regarding privacy, transparency, misinformation,
legal and regulatory etc. These are for the most part content free sops. One needs,
actually demands, the canonical evolution we have presented herein to understand what
the long-term risks may be. Having a construct to work with then policies may
evolve.
5
Stability of
Autonomous Entities
Autonomous entities, AE, can result in unstable constructs.
The inherent feedback may result in the AE in cycling in erratic ways that are
fundamentally unstable. This again is a concern that Wiener expressed. Stability of an AE may be impossible. They may be driven
by the construct of, “on the one hand but on the otherhand”. This is a
construct without a moral fabric, without an underlying code of conduct.
Isaac Asimov in his robot novels present the three rules of
robotics. However, AI is much more
than robotics. Robots, in the Asimovian world, were small self-contained
anthropomorphic entities.
In our construct the AI Autonomous entity is an ever-expanding entity capable
are unlimited capabilities. Moreover, these autonomous entities can evolve and
expand independent of human interaction or control. Thus, the key question is;
what can be done to protect humanity if not all of earthly entities from an
overpowering and uncontrollable autonomous entity?
First one must admit the putative capacity of existence for
such an entity. Second one must recognize that the creation of these entities
cannot be prevented since an adversary may very well do so as a means of a
threat or control. Third, creation of such entities may very well be in the
hands of technologists who lack and moral foundation and will just do so
because they can do it. Thus, it is nearly impossible for this entity to be a
priori controlled.
Therefore, at best one can a posteriori control such
entities. This requires advanced surveillance and trans-governmental control
mechanisms. Namely it can be possible to sense the existence and development of
such systems via various distributed network sensing mechanisms. When detected
there must be prohibitive actions in place and immediately executable in a
trans-border manner.
The Asimovian Robot is an anthropomorphic entity. In Asimov’s world the robot was a
stand-alone creature, one of many, with capabilities limited by its
singularity. Robots were just what they were and no more. An AI
Entity is a dynamically extensible entity capable of unlimited extension akin
to a slime mold, a never-ending extension of the plant. The AI Entity may morph
and add to itself what it internally sees a need for and take actions that are
solely of its own intent. Thus, there is a dramatic difference between a Robot
and an AI Entity. The challenge is that trying to apply the three laws of
robotics to an entity that controls its own morphing is impossible.
We have noted herein that the early developments of AI
revolve around increased processing and interaction complexity. However, there
comes a point when externalities become the dominant factor, namely the ability
of the AI entity to interact with its external environment, first with the help
of a human, then with existing external entities and then with the ability to
create and use its own externalities. This progression then leads to the AAIE which if not properly delimited can result in
harms.
Canonical Forms have multiple uses. First, they provide
structure. Second, they allow for defining issues and elements. Third they are
essential if any regulatory structure is imposed. We have seen this in
Telecommunications Law where elements and architecture is critical to
regulation. However, as in Telecom and other areas, technology evolves and
these Canonical Forms may do so likewise. Thus, they are an essential starting
point and subject to modification and evolution.
As we have observed previously, the conversion of various
sensory data to system processable data is a critical step. The human and other
animal sensory system have evolved over a billion years to maximize the
survival of the specific species. The specific systems available to AI are
still primitive and may suffer significant deficiencies.
However, in a AAIE system, self-evolution may occur at an order of multi magnitudes
faster that the evolution we have in our species. What direction that evolution
takes is totally uncertain. The effects of that evolution will also determine
what an AAIE does as it perceives its environment.
A group at MIT has recently made a regulatory proposal for
AI. They recognize, albeit
rather in a limited manner, that one must define something to regulate it. They
thus note:
It is important (but difficult) to define what AI is, but
often necessary in order to identify which systems would be subject to
regulatory and liability regimes. The most effective approach may be defining AI
systems based on what the technology does, such as “any technology for making
decisions or recommendations, or for generating content (including text,
images, video or audio).” This may create fewer problems than basing a
definition on the characteristics of the technology, such as “human-like,” or
on technical aspects such as “large language model” or “foundation model” –
terms that are hard to define, or will likely change over time or become
obsolete. Furthermore, approaches based on definitions of what the technology
does are more likely to align with the approach of extending existing laws and
rules to activities that include AI.
Needless to say, the definition is so broad that it could
include a coffee maker or any home appliance. As we have argued herein, AI
inherently contains an element whereby massive data if collected and processed
by some means that permits a relationship between an input and output to be
posited. Also, and a key factor, is that the relationship between input and
posited output is hypothesized by some abstraction of data sets chosen by the
designer and potentially modified by the system.
The MIT group then states:
Auditing regimes should be developed as part and parcel
of the approach described above. To be effective, auditing needs to be based on
principles that specify such aspects as the objectives of the auditing (i.e., what
an audit is designed to learn about an AI system, for example, whether its
results are biased in some manner, whether it generates misinformation, and/or
whether it is open to use in unintended ways), and what information is to be
used to achieve those objectives (i.e., what kinds of data will be used in an
audit)
This is the rule of the select telling the masses what to
believe! It seems academics just can’t get away from this control mechanism.
They further note:
For oversight regarding AI that lies beyond the scope of
currently regulated application domains, and that cannot be addressed through
audit mechanisms and a system similar to that used for financial audits, the
federal government may need to establish a new agency that would regulate such
aspects of AI. The scope of any such regulatory agency should be as narrow
as possible, given the broad applicability of AI, and the challenges of
creating a single agency with broad scope. The agency could hire highly
qualified technical staff who could also provide advice to existing regulatory
agencies that are handling AI matters (pursuant to the bullets above).
(Such a task might alternatively be assigned to an existing agency, but any
existing agency selected should already have a regulatory mission and the
prestige to attract the needed personnel, and it would have to be free of
political and other controversies from existing missions that could complicate
its oversight of AI.) A self-regulatory organization (like the Financial
Industry Regulatory Authority, FINRA, in the financial world) might undertake
much of the detailed work under federal oversight by developing standards and
overseeing their implementation.
Again, another Federal entity, and as academics do, they
assume a base of qualified staff, an oxymoron for any Government entity. As we
have noted previously, if you can’t define it, you can’t regulate it. Also, as
is all too well known, all regulations have “dark sides”.
Posts
