I’ve been asked to speak
on the social implications of AI, which I take to mean its ethical and
political implications. I suspect that I
got this invitation because I teach a course on the Philosophy of Technology
for which I have written a textbook (plug plug). But I also suspect that I got it in part because
I teach a course called “Minds: Natural and Artificial”, which focuses on the
topic of “The Philosophy of Mind” and more particularly the issue called “the
hard problem”, which considers the mysterious phenomenon of consciousness. I won’t bore you with details, but I will
note that the term AI often invokes in people’s minds issues of consciousness
and the nature of mind and awareness.
One of the authors we read is Hubert Dreyfus who wrote a famous and
contentious book called “What Computers Can’t Do” in which he argues that
computers will never be able to manifest consciousness. In the class I approach the “hard problem” as
an open philosophical question. It is a
very fraught issue in which there is plenty of opportunity for debate. I mention these points only because the
question of whether computers can think often lurks in the background of
discussions of “AI” and adds most of the frisson surrounding the term in
people’s minds.
Many people associate AI
with images like that of commander Data in a Star Fleet courtroom defending his
right to be recognized as a person, or the robot from the movie iRobot pleading
with Will Smith’s character to recognize the plight of his people at the hands
of an exploitive humanity. I mention my course
and the “hard problem” because outside the context of a classroom, in more
practical settings like this one, I feel obliged to speak more frankly about
the prospects of machine intelligence. I
agree with Dreyfus that there are certain things that computers can’t do and
that it is extremely unlikely that they will ever manifest a level thinking
that would allow them to be considered independently creative or conscious.
Another issue I feel
obliged to deal with is the issue of my technical grasp of the topic. As a professor in the Arts and Humanities it
might be easy to assume that I am somewhat out of my depth when it comes to a
highly technical subject like artificial intelligence. Computers are black boxes for most people,
and philosophers might be considered about as far from the nuts-and-bolts of software
engineering as you can possibly get. I
will just mention that I have been an active computer programmer, largely as a
hobby, but in the past working on academic projects, for over 40 years. I have written tens of thousands of lines of
code over the years, including programs using what are typically described as AI
techniques. I would direct you to my
Internet Archive collection of early 8-bit programs and my Github repository
and pages to check my bona fides. (jggames.github.io and https://archive.org/details/AI8-bitBASICprograms)
So, on the issue of
computer intelligence and creativity, I would qualify my prior opening remarks
by stating that I think computer software, as has been well demonstrated over
the past half century, can be a great aid to human creativity. For example, Eric Topol's 2013 book, The
Creative Destruction of Medicine, illustrates some useful possibilities for
developing software to take up the load of medical diagnosis and better
information management desperately needed in public health systems. And Erik Brynjolfsson and Andrew McAfee’s The
Second Machine Age, give a wonderful rundown on the economic positives of
new tech. But what these recent improvements
in medicine and commerce illustrate is that what we
are really concerned with is a much narrower definition of intelligence. Computers can indeed “think” in the much more modest sense of carrying out tasks formerly carried out only in human brains.
They have been doing so since at least the Antikythera mechanism built by the
ancient Greeks to calculate astronomical events and the timing of the Olympic
games, and ancient Chinese abacuses. There is nothing new about machines doing intellectual tasks except
perhaps the recent substantial increases in the pace of change that is to be
expected in a society at the apogee of a bonanza of cheap high intensity energy
like that provided by fossil fuels over the last two centuries.
In brief, I see the term
AI more as a contemporary buzz term, spurred by recent improvements in language
recognition software in combination with advancements in visual and auditory
generative programs made possible by access to vast amounts of data generated
by the Internet. The recent tendency to use the term AI with its exciting
connections to the “hard problem” is no doubt a convenient tool for anyone
connected with the need to raise investment capital required in free market
economies. But as a coder I really can’t see the term as anything more than a fund-raising
or talent recruiting trick aimed at spurring on new projects dreamed up by software
engineers.
I am not alone in holding such
views. As Linus Torvalds, inventor of the Linux operating system put it in a recent interview, it represents a bunch of people "with their hands out" and another hype-cycle like crypto or "cloud native." Others have written books about AI
viewed primarily as a marketing device.
I will just mention Katie Crawford’s well-received “Atlas of AI” and
Merideth Broussard’s Artificial Unintelligence. Yuval Noah Harari has a fascinating chapter
in his book Homo Deus on the new religions of Silicon Valley and what he
calls “Data Religion.” I would more
observe that the real computer revolution occurred long before ChatGPT in the final
four decades of the last century, when the application of mundane computer
software and automation equipment de-industrialized our society and shrank the blue-collar
sector from just over 35 percent of employment to something closer to 10 or 15
percent. As economists and historians of
deindustrialization have observed, most of that process did not result from the
offshoring jobs but processes of automation carried on within our society.
If anything, offshoring
occurred late in the process, the last decade or so, largely to help
deindustrialized workers maintain their buying power. Ten-dollar T-shirts from Asia have helped
maintain family incomes that would have otherwise noticeably shrunk over the
last decades. The vast increases in
productivity in the industrial sector was achieved as result of trillions of
dollars of investment. But trillions
were also spent in the final four decades of the last century in the service
sector as well, with almost no measurable increase in productivity measured,
until recently. As American economist
Robert Solo famously quipped in 1987 “You can see the computer age everywhere but in the
productivity statistics."
Through that period of
deindustrialization people did not go on about the potential impacts of
“artificial labour.” The term
“automation” was sufficient. Since that
time economists have been waiting for the shoe to drop in the service
sector. But employment just kept growing
and growing in that sector, without significant attendant productivity growth, despite
vast investments in computerization. The
result has possibly been the creation of a vast array of what David Graeber
calls “bullshit jobs” in his provocative book of that title. Sometimes I am inclined to think that AI is
simply a term preferred by white collar workers, me included, who feel somewhat
threatened by the impending true application of automation to our bloated sector. It grants the process the higher level of cache
that we feel our work deserves compared to that of our blue-collar
fellows. Which brings us to the first major
moral issue regarding AI, which is the issue of technological unemployment.
It is an open question
whether technological development can or will eventually lead to an acute
crisis of employment rather than the wage stagnation and heightening itinerancy
with which we are familiar. This is an
empirical issue and still to some extent a future issue. We have been able to keep many people
employed, or occupied with education, early retirement or social supports,
although anyone familiar with the various drug and mental health epidemics will
tell you about the limits of such efforts. Recent
studies aside suggesting that we might finally be seeing a decoupling of
productivity growth from levels of employment growth, there is a robust
philosophical and ethical debate going on about whether the work that we do
have and can expect to have will be of an edifying nature, regardless of
whether enough of the resulting wealth can be appropriately shared. Some people argue for a guaranteed annual
income or other wealth distributing schemes.
I would simply note that such proposals do not grapple with the more fundamental
issue of the quality and meaningfulness of work. Figuring out how to make such judgements and
how to best ensure that human beings can have enough opportunity to apply
themselves to meaningful tasks is a critical question that continues to vex
regardless of proposals regarding the sharing of wealth.
In a somewhat related
vein, there is the fundamental question raised by authors like Crawford, of the
relation of AI to the more general environmental crisis. It is a connection that is often overlooked,
but it is a highly relevant observation to make, as she does in her book, that
computers and electronics are high energy and resource intensive activities,
both in their infrastructural requirements and typical applications. One need only note that in the early 2000s
the improvements made in Great Britain in terms of increases in energy
efficiency achieved through intensive public actions and investments motivated
global treaty obligations, were entirely offset by increases in energy requirements
needed for the infrastructure of the digital revolution. Crawford’s exploration
of the vast air-conditioned server farms needed to host our cat videos, not to
mention the now vastly expanding AI infrastructures, is sobering. But as Crawford also points out the infrastructure
of AI is tightly interwoven with activities still primarily focused on exploiting
natural resources, as has been the hallmark of commercial activity since the
industrial revolution. Nothing so far in
the empirical data robustly indicates that AI represents a radical shift from this
pattern of consumption. But the human species must collectively consider sustainable
alternatives to this economic model as was well illustrated by MIT’s original 1972
Limits to Growth model and its recent updates in 1992, 2012 and 2022.
Finally, there are specific
ethical issues related to the development of AI tools themselves and their
application for specific purposes. First,
the development of Large Language Models and visual and auditory generative
techniques have been highly dependent on access to vast amounts of human
generated training data used to apply to the various “machine-learning” methods
required to develop such applications. These
processes raise many issues regarding the use of “our” data to benefit other
people’s commercial purposes. These include issues about copyright,
intellectual property rights and privacy.
More broadly the incentive of big data companies to gain access to our
information create many potential moral hazards regarding the farming of users
for their information. Since we are
currently in the very midst of such processes of development it is easy to
overstate the challenges and the difficulties of finding reasonable administrate
and legal solutions.
A second example relating
to application simply involves the possibilities of the new tools to facilitate
new kinds of malfeasance, that we might be insensitive to simply because of the
novelty of the activities attending the new tools. This is an abiding issue of technological
change. Are the newfangled automobiles love hotels on wheels for teenagers? Is selling bootleg video tapes theft? Is
hacking a kind of trespassing? Is
texting while driving recklessness? And,
of course, most recently, is not properly attributing material produced by
machine a form of fraud?
A somewhat novel type of
issue regarding AI development and application can be described by the term
“the alignment problem” coined by Brian Christian in his book of the same
title. Since AI programing techniques
like “machine-learning” apply the kind of programing techniques that coders at
one time simply called “self-modifying code,” which we were told by our
teachers represented the “ultimate” in programmer laziness and warned to
absolutely avoid at all costs, means the resulting software has the unavoidable
quality of a black box. Unlike
traditional algorithmic or heuristic methods, contemporary programmers don’t have
a good grasp of how their system operate and will continue to operate in novel
conditions. This raises many issues
about the handing over of tasks normally requiring human judgement to
machines. There are now famous instances
of what used to simply be called “expert systems” manifesting hidden biases
often resulting from tendencies buried deeply in the human created training
data, but sometimes simply from the imponderables of the programing methods as
such.
One specific example of
an issue regarding the application of AI is the issue of robotic forms of
warfare. The questions of whether
machines should be handed even greater levels of discretion regarding the exercise
of lethal judgment on our behalf is a very challenging ethical question,
although I would note that such issues have been around since poison gas and
delayed action munitions. So, I don’t
think these types of questions are really an issue specific to what we are now
calling AI.
I would to put most of these
specific issues of development and use in the “scare the horses” category. As in the case of the early automobile when
people didn’t know how we would manage issues like maintenance, traffic flow
and driving etiquette, these now largely forgotten vitriolic debates were
quickly resolved. But as the case of
automobile would also suggest we might well have done a better job of looking
at infrastructural issues, like what would happen with all the exhaust fumes
coming out of vehicles and how their operating requirements would influence us
in re-shaping our cities. So, I would
tend to weigh the issues of energy and resource use more highly.
It is a simple reality of
physics that the development of AI to the degree being predicated by its main
advocates will require vast increases in access to energy, both for running
computer systems supporting AI processing, but also soon, for creating and storing
the vast amounts of artificially created training data that will be needed. The proposed levels of advances in machine
learning will require much more data than even our prodigious current Net use could
ever supply. But the gurus of AI, when
asked about these more mundane energy issues, quickly flip into modes of
magical thinking, speaking about fusion and mining asteroids, and the like.
So, we cannot escape the preeminent
technological issue of our age regarding energy. And the complex systems of our energy systems
raise many possible cases of whether there are some technological activities
that simply should not be done, or as the Latins so concisely put it: ab esse
ad posse non valet consequentia (just because something can be done doesn’t
mean it should be) Considering whether
there are limits to the creation and application of technologies are not as
deeply considered questions as they should be, although I would note the
positive signs that this may be changing illustrated by Canada’s leading role
in the international treaty banning landmines and recent efforts to limit single
use plastics.
Finally, although the word
“technology” is one of the most prominent terms of our age, the definition of
this concept turns out to be a highly contested philosophical topic. The fact that such a key term could be so
philosophically confused and misunderstood stands itself as the main moral failing
of our age. As Marshall McLuhan so
sagely put it, “the medium is the message.”
Interpreting the meaning of technology as such is the preeminent moral
challenge of our time.
Bibliography
Topol, Eric. The Creative
Destruction of Medicine: How the Digital Revolution Will Create Better Health
Care (2013)
Crawford,
Kate. Atlas of AI (2021)
Broussard,
Meredith. Artificial Unintelligence (2019)
Gerrie, James. A Plea for the Preservation of Early BASIC Game Programs Canadian Journal of
Media Studies/Revue Canadienne d'études des médias,18(1) 2022, pp. 90-113.
Gerrie, James. "Software Preservation Insights on the Power of BASIC" in Game Science. Digital
Humanities for Games and Gaming. (Disk Book). Melanie Fritsch, Stefan Höltgen,
Torsten Roeder, Editors. Weimar: PolyPlay 2023.
Graeber,
David. Bullshit Jobs: A Theory (2018)
Christian, Brian. The
Alignment Problem: Machine Learning and Human Values (2020)
Harari, Yuval Noah. Homo Deus (2015)
