<html>
<head>
<meta http-equiv="content-type" content="text/html; charset=UTF-8">
</head>
<body>
<div class="tls-article-intro-primary columns is-gapless
is-multiline is-centered">
<div class="tls-article-intro-primary__label-wrapper column is-12">
<div class="tls-article-label" data-featured="false"><span
class="tls-article-label__category"><a
href="https://www.the-tls.co.uk/categories/philosophy/contemporary-philosophy/">Contemporary
philosophy</a></span><span
class="tls-article-label__separator">|</span><span
class="tls-article-label__article-type">Book Review</span></div>
</div>
<div class="tls-article-intro-primary__headline-wrapper column
is-12-mobile is-10-tablet is-8-desktop is-8-widescreen">
<h1 class="tls-headline">Computers don’t give a damn</h1>
</div>
<div class="tls-article-intro-primary__standfirst-wrapper column
is-12-mobile is-10-tablet is-8-desktop is-8-widescreen">
<h2 class="tls-standfirst">The improbability of genuine thinking
machines</h2>
</div>
<div class="tls-article-intro-primary__byline-wrapper column
is-12">
<div class="tls-byline"><span class="tls-byline__by">By </span><span
class="tls-byline__name" data-disablelink="false"><a
href="https://www.the-tls.co.uk/articles/promise-of-artificial-intelligence-brian-cantwell-smith-book-review/#">tim
crane</a></span></div>
</div>
</div>
<div class="tls-single-article__lead-image">
<div class="tls-lead-image columns is-gapless is-multiline
is-centered">
<div class="tls-lead-image__wrapper column is-10-desktop
is-10-widescreen"><img class="tls-lead-image__image"
src="https://www.the-tls.co.uk/s3/tls/uploads/2020/05/AssetAccess81.jpg"
alt="“Archive Dreaming” installation at SALT Galata,
Istanbul, by Refik Anadol, 2017. It uses AI to sort and
combine 1.7 million historical documents" width="786"
height="442">
<div class="tls-media-information" data-withcredit="true"
data-wraptext="true"><span
class="tls-media-information__caption">“Archive Dreaming”
installation at SALT Galata, Istanbul, by Refik Anadol,
2017. It uses AI to sort and combine 1.7 million
historical documents</span><span
class="tls-media-information__credit">© Getty Images</span></div>
</div>
</div>
</div>
<div class="tls-single-article__wrapper columns is-centered
is-gapless is-multiline">
<div class="tls-single-article__sidebar column is-4-desktop
is-10-tablet">
<div class="tls-single-article__sidebar-content">
<div class="tls-article-stamp tls-read-this-issue">
<div class="tls-article-stamp__meta-wrapper"
data-withimage="true">
<div class="tls-article-stamp__label">May 15, 2020</div>
<div class="tls-article-stamp__link"><a tabindex="-1"
class="tls-link tls-link--variant-primary
tls-link--theme-default tls-link--size-small false "
href="https://www.the-tls.co.uk/issues/may-15-2020/"
data-size="small">Read this issue</a><span
class="tls-link--icon"><svg width="1em" height="1em"
viewBox="0 0 20 20"></svg></span></div>
</div>
</div>
</div>
</div>
</div>
<div class="tls-book-details columns is-gapless is-multiline
tls-sidebar-bookdetails">
<div class="tls-book-details__wrapper ">
<div class="tls-book-details__unit">
<h3 class="tls-component-heading">In this review</h3>
<div class="tls-list-item-book-review">
<div class="tls-list-item-book-review__book-meta-wrapper">
<div class="tls-list-item-book-review__book-title">THE
PROMISE OF ARTIFICIAL INTELLIGENCE</div>
<div class="tls-list-item-book-review__book-details">Reckoning
and judgment<br>
184pp. MIT Press. £20 (US $24.95).</div>
<div class="tls-list-item-book-review__author-details">Brian
Cantwell Smith</div>
</div>
<div class="tls-list-item-book-review__buy-button"><a
tabindex="-1" class="tls-link tls-link--variant-primary
tls-link--theme-default tls-link--size-small false "
href="https://shop.the-tls.co.uk/the-promise-of-artificial-intelligence-9780262043045.html?utm_source=tls&utm_medium=mainsite&utm_campaign=review"
data-size="small">Buy</a><span class="tls-link--icon"><svg
width="1em" height="1em" viewBox="0 0 20 20"></svg></span></div>
</div>
</div>
</div>
</div>
<div class="tls-article-body">
<p class="has-drop-cap dropcap@1">In 1965, Herbert Simon, one of
the founders of the new science of Artificial Intelligence (AI),
wrote that “machines will be capable, within twenty years, of
doing any work that a man can do”. He was wrong of course – but
maybe his mistake was only a matter of timing.</p>
<p>If Simon were to see what computing machines were capable of,
fifty-five years after he made this remark, surely even he would
be amazed. A single smartphone contains more computing power
than all the world’s computers in 1965 put together. And many of
the philosophical arguments against the possibility of AI from
the 1960s and 70s fell flat on their face as the technology
advanced. The philosopher Hubert Dreyfus also claimed in 1965
that “no chess program can play even amateur chess” – true at
the time, but proved false soon after. When in 1997 the IBM
programme Deep Blue beat the chess champion Garry Kasparov, this
conclusively destroyed the idea that world-class chess was
something computers can’t do; Kasparov has commented recently
that “today you can buy a chess engine for your laptop that will
beat Deep Blue quite easily”. And the familiar claim that
computers could never really use their stored knowledge as well
as human beings was shaken in 2011 by IBM’s Watson programme,
which won the top $1 million prize on the American game show <i>Jeopardy</i>,
beating the best human competitors.</p>
<p>AI sceptics used to claim that computers would never be able to
recognize human faces or human speech, translate speech into
text, or convert handwriting to printed text. But today’s phones
can do all these things. Dreyfus had made gentle fun of the
grand claims of AI, quoting a fanciful newspaper report from
1968 about “a new idea in gifts … a genuine (if small) computer,
that costs around $20. Battery operated, it looks like a
portable typewriter. But it can be programmed like any big
computer to translate foreign languages, diagnose illnesses,
even provide a weather forecast”. What seemed then like wild
science fiction is now our everyday reality.</p>
<p>So Simon’s claim may have been proved false, but maybe he was
only a few decades out. The achievements of actual AI – that is,
the kind of technology that makes your smartphone work – are
incredible. These achievements have been made possible partly by
developments in hardware (in particular the increased speed and
miniaturization of microprocessors) and partly because of the
access to vast amounts of data on the internet – both factors
that neither Simon nor Dreyfus could have predicted. But it
means that enthusiastic predictions for AI are still popular.
Many believe that AI can produce not just the “smart” devices
that already dominate our lives, but genuine thinking machines.
No one says that such machines already exist, of course, but
many philosophers and scientists claim that they are on the
horizon.</p>
<p>To get there requires creating what researchers call
“Artificial General Intelligence” (AGI). As opposed to a
special-purpose capacity – like Deep Blue’s capacity to play
chess – AGI is the general capacity to apply intelligence to an
unlimited range of problems in the real world: something like
the kind of intelligence we have. The philosopher David Chalmers
has confidently claimed that “artificial general intelligence is
possible … There are a lot of mountains we need to climb before
we get to human-level AGI. That said, I think it’s going to be
possible eventually, say in the 40-to-100-year time frame”. The
philosophers John Basl and Eric Schwitzgebel are even more
optimistic, claiming it is “likely that we will soon have AI
approximately as cognitively sophisticated as mice or dogs”.</p>
<p>The intellectual enthusiasm for the possibility of AGI is
matched by the vast sums invested in trying to make it a
reality. In July 2019, Microsoft announced that it would invest
$1 billion in Sam Altman’s OpenAI, a for-profit company which
aims to use AI for the “benefit of mankind as a whole”. The
British company Deep Mind, led by the computer
scientist/neuroscientist Demis Hassabis, was bought by Google
for $500 million in 2014. Deep Mind’s best-known achievement to
date is the machine AlphaGo, which in 2016 beat Lee Sedol, world
champion of the ancient game of Go. Go is vastly more complex
than chess – it is sometimes said to be the most complex game
ever created – and standard AI chess-playing methods had never
been successfully applied to it. The computing methods used by
AlphaGo are often touted as one of keys to “solving
intelligence”, as Deep Mind’s own publicity puts it.</p>
<p>Brian Cantwell Smith’s new book is a provocative expression of
scepticism about these recent claims on behalf of AI, from a
distinguished practitioner in the field. His overall argument is
based on a distinction between what he calls “reckoning” and
“judgment”. Reckoning is understood here in its original
etymological sense: as calculation, like addition and
subtraction. Judgment, by contrast, is something more. It is
described by Smith as “an overarching, systemic capacity or
commitment, involving the whole commitment of the whole system
to the whole world”. Our thinking involves not just some kind of
simple on-off representation of things around us, but an entire
emotional and value-laden involvement with the world itself.
Computers have none of this. As the philosopher John Haugeland
(a major influence on Smith) used to say, “computers don’t give
a damn”. Giving a damn is a precondition of “judgment” in
Smith’s sense, and anything that amounted to a real AGI would
need to exercise judgment, and not simply calculate.</p>
<p><i>The Promise of Artificial Intelligence </i>gives a brief
and intelligible survey of two main stages in the history of AI.
The first stage, starting in the 1960s, was what Haugeland
christened “Good Old-Fashioned AI” (GOFAI) which solved
computing problems by using explicit representations of general
principles and applying them to particular situations. (Think of
doing a mathematical proof or presenting an argument in logic.)
“Second wave” AI, which started to emerge in the 1980s, began
from the opposite end, so to speak: deriving general conclusions
from huge amounts of simple data as input. This kind of
approach, variously called machine learning or deep learning,
has had considerable success at things that GOFAI was very bad
at, like pattern recognition, or updating knowledge on the basis
of input.</p>
<p>First wave AI, it was often said, misconceived the nature of
thinking: very little thinking resembles calculating or proving
theorems. But Smith goes further, and argues that “the deeper
problem is that it misconceived the world”. GOFAI assumed that
“the world comes chopped up into discrete objects”, and because
of this it analysed reasoning into its components by using
formal logic (the basic ideas of which underlie modern
computing). Smith argues that first wave theorizing assumed that
the world must be structured in the way that logic structures
language: objects correspond to names, properties correspond to
predicates or general terms. Things fit together in the world as
symbols fit together in a logical language. Smith claims that
this is one main reason why the GOFAI project failed: it failed
to take account of the “fabulously rich and messy world we
inhabit”, which does not come in a “pre-given” form, divided up
into objects.</p>
<p>Second wave AI, according to Smith, does not make this mistake.
It does not assume a “pre-given” ontology or structure for the
world, and for that reason, he argues, it has made progress in
the areas where GOFAI failed: in particular, with tasks like
face recognition, text processing and (most famously) the game
of Go. The distinctive feature of deep learning machines is
their ability to detect patterns in large (sometimes huge)
amounts of data. The machines “learn” by being given an
indication by the programmer of which patterns are the important
ones, and after a while they can produce results (for example,
moves in a game) that surprise even the programmers. This is in
contrast to first wave AI programmes which attempted to
anticipate in advance how input from the real world should be
responded to in every conceivable situation. Those early AI
machines that worked only did so in very constrained made-up
environments, sometimes called “microworlds”.</p>
<p>Nonetheless, Smith thinks that we should not be too optimistic
about the ability of second wave AI to create AGI. Machine
learning may not start with general rules which make ontological
assumptions, but it does start with data that is already
processed by humans (eg things that we classify as faces, or as
road traffic at an intersection and so on). Much machine
learning, as Smith says, is “dedicated to sorting inputs into
categories of manifest human origin and utility”. So even if
they are more sensitive to the messy world, second wave AI
machines are still tied up with the programmers’ own
classifications of reality – indeed, it is hard to see how they
could be otherwise designed.</p>
<p>Smith is surely right that AI’s recent successes give us little
or no reason to believe in the real possibility of genuine
thinking machines. His distinction between reckoning and
judgment is an important attempt to identify what it is that is
missing in AI models. In many ways (despite his protest to the
contrary) it echoes the criticisms of Dreyfus and others, that
AI will not succeed in creating genuine thinking unless it can
in some way capture “common sense”. And just as common sense
(part of Smith’s “judgment”) cannot be captured in terms of the
“rules and representations” of GOFAI, nor can it be captured by
massively parallel computing drawing patterns from data.</p>
<p>To make this point about judgment, Smith does not actually need
the more ambitious ontological claims, that the world does not
have natural divisions or boundaries, that all classification is
simply a result of human interest, and so on. Maybe these claims
are true, maybe not – for many centuries philosophy has wrestled
with them, and they are worth debating. But we should not need
to debate them in order to identify the fundamental
implausibility of the idea that AGI is on the horizon.</p>
<p>This implausibility derives from something intrinsic to the
success of AI itself. For despite the sophistication of machine
learning, the fact remains that like chess, Go is still a game.
It has rules and a clear outcome which is the target for
players. Deep learning machines are still being used to achieve
a well-defined goal – winning the game – the meaning of which
can be articulated in advance of turning on the machine. The
same is true of speech and face recognition software. There is a
clear goal or target – recognizing the words and faces – and
successes and failures in meeting this goal are the input which
helps train the machine. (As Smith says, “recognition” here
means: correctly mapping an image onto a label: nothing more
than that.)</p>
<p>But what might be the goal of “general intelligence”? How can
we characterize in abstract terms the problems that general
intelligence is trying to solve? I think it’s fair to say that
no one – in AI, or philosophy, or psychology – has any idea how
to answer this question. Arguably, this is not because it is an
exceptionally difficult empirical question, but rather that it
is not obviously a sensible one. I suppose someone might say, in
the spirit of Herbert Simon (whose famous AI programme was
called the “General Problem Solver”), that general intelligence
is the general ability to solve cognitive problems. This might
seem fine until we ask ourselves how we should identify, in
general terms, the cognitive problems that we use our
intelligence to solve. How can we say, in general terms, what
these problems are?</p>
<p>Consider for example, the challenges faced in trying to create
a genuine conversation with a computer. Voice assistants like
Siri and Alexa do amazingly well in “recognizing” speech and
synthesizing speech in response. But you very quickly get to the
bottom of their resources and reach a “canned” response (“here
are some webpages relating to your inquiry”). One reason for
this, surely, is that conversation is not an activity that has
one easily expressible goal, and so the task for the Siri/Alexa
programme cannot be specified in advance. If the goal of
conversation were to find information about a subject matter,
then directing you to a website with relevant information could
be one reliable way of achieving that goal. But of course this
is not the sole thing to which we direct our intelligence when
talking with others.</p>
<p>What, then, is the overall goal of conversation? There isn’t
one. We talk to pass the time, to express our emotions, feelings
and desires, to find out more about others, to have fun, to be
polite, to educate others, to make money … and so on. But if
there is no single goal of conversation, then it is even less
likely that there is one goal of “general intelligence”. So no
wonder AI researchers struggle to even define the “task domain”
for AGI.</p>
<p>As Smith’s book shows, the claims for the possibility of AGI
ignore the huge differences between the relatively well-defined
areas where AI has succeeded, and the barely defined domain of
“general intelligence”. This is, on its own, enough of a reason
for scepticism about extrapolating beyond the spectacular
successes of actual AI to the real possibility of AGI. Smith’s
arguments about the ontological assumptions of AI, whatever
their merits, are not necessary to make this point.</p>
<p>Yet I suspect that many still have this lingering sense that
AGI must be possible, and that the difference between real human
thinking and what computers do is just a matter of complexity.
What lies behind this conviction? One widespread idea is that
since the human brain is just a complex biological (and
therefore material) machine, it must be possible in principle to
artificially reproduce what the human brain does (thinking,
perceiving, feeling, imagining, being conscious, etc) by
building something that functions in exactly the same way as the
brain. And whatever we thereby build will be an artificial
version of our mental processes: an AGI.</p>
<p>This argument is based on two ideas: first, that thinking and
other mental processes go on in the brain; second, the brain is
a machine or mechanism. So if we can uncover the principles that
make this mechanism work, and we have adequate technology – the
argument goes – then we should be able to build a machine that
implements these principles, without leaving anything out. One
of the pioneers of deep learning, Yoshua Benigo, puts it this
way: “I don’t know how much time it’s going to take, but the
human brain is a machine. It’s a very complex one and we don’t
fully understand it, but there’s no reason to believe that we
won’t be able to figure out those principles”.</p>
<p>Of course building an artificial copy of a real brain is
nowhere close to today’s scientific reality. But if we believe
that we are at bottom material beings – if we take away all our
matter then there is nothing left of us – then such replication
seems possible in principle, even if it is never actually
realized. Suppose, then, a brilliant scientist of the future
could replicate in an artificial construction everything a
person (and their brain) does. Obviously this replica would also
be able to think, since thinking is one of the things a person
does. It is undeniable that making an artificial replica of a
person and all their features would be making an “artificial
intelligence” in an an obvious sense: simply because
intelligence is one of the features of people.</p>
<p>The question is, what does this have to do with AI? If the way
to create a real artificial thinker is to find out first how the
brain works, then you would expect AI researchers to try and
figure out how the brain or the mind actually works – that is,
to become neuroscientists or psychologists. But this is not how
AI researchers operate. Just as the invention of the aeroplane
did not require building something that flies in exactly the way
a bird flies, so the inventors of AI did not feel bound to copy
the actual workings of human brains. And despite the fact that
deep learning computers use what are called “neural networks”,
the similarity to the brain here is at a very abstract level.
Indeed, many pioneers of deep learning occupy themselves with
very abstract questions about general intelligence – Benigo says
his goal is “to understand the general principles of
intelligence” – rather than with the messy business of the
actual working of the human mind or brain.</p>
<p>This lack of focus on the way human minds (or brains) actually
function goes back to the beginnings of AI, and it was clearly
one of AI’s strengths. By ignoring the complexity of actual
human thinking, and the messy “wetware” of the actual human
brain, AI could get machines to solve difficult problems without
having to bother with how we would solve them. Watson, the IBM
website tells us, “is not bound by volume or memory. Watson can
read millions of unstructured documents in seconds”. That’s not
something we can do. The learning involved in training AlphaGo
involved millions of practice games of Go – as the cognitive
scientist and deep learning sceptic Gary Marcus has pointed out,
this is “far more than a human would require to become world
class at Go”. This means that whatever it explains about their
success, it is not the similarity of these machines to human
thinkers. So IBM should not claim that Watson is “built to
mirror the same learning process that we have, a common
cognitive framework that humans use to inform their decisions”.</p>
<p>Actual AI does not rely substantially on any detailed
psychological or neuroscientific research. So the fact that the
brain is a material mechanism which could, in principle, be
replicated artificially gives no support to the idea that AI as
it actually is could build an AGI. In fact, given the way AI has
actually proceeded, in splendid isolation from neuroscience, it
is likely that any attempt to replicate the brain would require
ideas very different from those traditionally used by AI. To say
this is not to disparage AI and its achievements; it is just to
emphasize the obvious fact that it is not, and has never been, a
theory of human thinking.</p>
<p class="square">Philosophical and scientific discussions of AI
have tended towards one of two extremes: either that genuine
artificial thinking machines are just on the horizon, or they
are absolutely impossible in principle. Neither approach is
quite right. On the one hand, as Smith, Marcus and others have
explained, we should be sceptical that recent advances in AI
give any support to the real possibility of AGI. But on the
other, it is hard to deny the abstract philosophical claim that
if you could replicate a human brain in such a way that would
produce something that did everything the brain did, then that
thing would be a “thinking machine” in one clear sense of that
phrase. However, this mere possibility does not mean that
today’s AI is anywhere near creating genuine thinking machines.
On the contrary: when you examine the possibility more closely,
it shows why AI is unlikely ever to do this.</p>
<p><em><strong>Tim Crane</strong> is Professor of Philosophy at
the Central European University and Philosophy editor of the </em>TLS<em>.
The third edition of his book </em>The Mechanical Mind<em>
was published in 2015</em></p>
</div>
</body>
</html>