Chapter 20
Ten Ways in Which AI Has Failed
IN THIS CHAPTER
Comprehending the world
Developing new ideas
Understanding the human condition
Any comprehensive book on AI must consider the ways in which AI has failed to meet
expectations. The book discusses this issue in part in other chapters, giving the
historical view of the AI winters. However, even with those discussions, you might
not grasp that AI hasn’t just failed to meet expectations set by overly enthusiastic
proponents; it has failed to meet specific needs and basic requirements. This chapter
is about the failures that will keep AI from excelling and performing the tasks we
need it to do to fully achieve the successes described in other chapters. AI is currently
an evolving technology that is partially successful at best.
One of the essential issues surrounding AI today is that people keep anthropomorphizing
it and making it into something it isn’t. An AI accepts cleaned data as input, analyzes
it, finds the patterns, and provides a requested output. As described in the sections
of this chapter, an AI doesn’t understand anything, it can’t create or discover anything
new, and it has no intrapersonal knowledge, so it can’t empathize with anyone about
anything. The critical piece of information to take from this chapter is that an AI
behaves as designed by a human programmer, and what you often take for intelligence
is only a mix of clever programming and vast amounts of data analyzed in a specific
manner. For another view of these and other issues, check out the article entitled
“Asking the Right Questions About AI,” at https://medium.com/@yonatanzunger/asking-the-right-questions-about-ai-7ed2d9820c48.
Even more important, however, is that people who claim that an AI will eventually
take over the world fail to understand that doing so is impossible given current technology.
An AI can’t suddenly become self-aware because it lacks any means of expressing the
emotion required to become self-aware. As shown in Table 1-1 in Chapter 1, an AI today lacks some of the essential seven kinds of intelligence required to
become self-aware. Simply possessing those levels of intelligence wouldn’t be enough,
either. Humans have a spark in them —something that scientists don’t understand. Without
understanding what that spark is, science can’t recreate it as part of an AI.
Understanding
The ability to comprehend is innate to humans, but AIs completely lack it. Looking
at an apple, a human more than just a series of properties associated with a picture
of an object. Humans understand apples through the use of senses, such as color, taste,
and feel. We understand that the apple is edible and provides specific nutrients.
We have feelings about apples; perhaps we like them and feel that they’re the supreme
fruit. The AI sees an object that has properties associated with it — values that
the AI doesn’t understand, but only manipulates. The following sections describe how
the failure to understand causes AI as a whole to fail to meet expectations.
Interpreting, not analyzing
As stated many times throughout the book, an AI uses algorithms to manipulate incoming
data and produce an output. The emphasis is on performing an analysis of the data.
However, a human controls the direction of that analysis and must then interpret the
results. For example, an AI can perform an analysis of an x-ray showing a potential
cancer tumor. The resulting output may emphasize a portion of the x-ray containing
a tumor so that the doctor can see it. The doctor might not be able to see the tumor
otherwise, so the AI undoubtedly provides an important service. Even so, a doctor
must still review the result and determine whether the x-ray does indeed show cancer.
As described in several sections of the book, especially with self-driving cars in
Chapter 14, an AI is easily fooled at times when even a small artifact appears in the wrong
place. Consequently, even though the AI is incredibly helpful in giving the doctor
the ability to see something that isn’t apparent to the human eye, the AI also isn’t
trustworthy enough to make any sort of a decision.
Interpretation also implies the ability to see beyond the data. It’s not the ability
create new data, but to understand that the data may indicate something other than
what is apparent. For example, humans can often tell that data is fake or falsified,
even though the data itself presents no evidence to indicate these problems. An AI
accepts the data as both real and true, while a human knows that it’s neither real
nor true. Formalizing precisely how humans achieve this goal is currently impossible
because humans don’t actually understand it.
Going beyond pure numbers
Despite any appearance otherwise, an AI works only with numbers. An AI can’t understand
words, for example, which means that when you talk to it, the AI is simply performing
pattern matching after converting your speech to numeric form. The substance of what
you say is gone. Even if the AI were able to understand words, it couldn’t do so because
the words are gone after the tokenization process. The failure of AIs to understand
something as basic as words means that an AI’s translation from one language to another
will always lack that certain something needed to translate the feeling behind the
words, as well as the words themselves. Words express feelings, and an AI can’t do
that.
The same conversion process occurs with every sense that humans possess. A computer
translates sight, sound, smell, taste, and touch into numeric representations and
then performs pattern matching to create a data set that simulates the real-world
experience. Further complicating matters, humans often experience things differently
from each other. For example, each person experiences color uniquely (https://www.livescience.com/21275-color-red-blue-scientists.html). For an AI, every computer sees color in precisely the same way, which means that
an AI can’t experience colors uniquely. In addition, because of the conversion, an
AI doesn’t actually experience color at all.
Considering consequences
An AI can analyze data, but it can’t make moral or ethical judgements. If you ask
an AI to make a choice, it will always choose the option with the highest probability
of success unless you provide some sort of randomizing function as well. The AI will
make this choice regardless of the outcome. The “SD cars and the trolley problem” sidebar in Chapter 14 expresses this problem quite clearly. When faced with a choice between allowing either
the occupants of a car or pedestrians to die when such a choice is necessary, the
AI must have human instructions available to it to make the decision. The AI isn’t
capable of considering consequences and is therefore ineligible to be part of the
decision-making process.
In many situations, misjudging the ability of an AI to perform a task is merely inconvenient.
In some cases, you may have to perform the task a second or third time manually because
the AI isn’t up to the task. However, when it comes to consequences, you might face
legal problems in addition to the moral and ethical problems if you trust an AI to
perform a task that unsuited to it. For example, allowing a self-driving (SD) car
to drive by itself in a place that doesn’t provide for this need is likely illegal,
and you’ll face legal problems in addition to damage and medical charges that the
SD car can cause. In short, know what the legal requirements are before you trust
an AI to do anything involving potential consequences.
Discovering
An AI can interpolate existing knowledge, but it can’t extrapolate existing knowledge
to create new knowledge. When an AI encounters a new situation, it usually tries to
resolve it as an existing piece of knowledge, rather than accept that it’s something
new. In fact, an AI has no method for creating anything new, or seeing it as something
unique. These are human expressions that help us discover new things, work with them,
devise methods for interacting with them, and create new methods for using them to
perform new tasks or augment existing tasks. The following sections describe how an
AI’s inability to make discoveries keeps it from fulfilling the expectations that
humans have of it.
Devising new data from old
One of the more common tasks that people perform is extrapolation of data; for example, given A, what is B? Humans use existing knowledge to create
new knowledge of a different sort. By knowing one piece of knowledge, a human can
make a leap to a new piece of knowledge, outside the domain of the original knowledge,
with a high probability of success. Humans make these leaps so often that they become
second nature and intuitive in the extreme. Even children can make such predictions
with a high rate of success.
The best that an AI will ever do is to interpolate data for example, given A and B, is C somewhere in between? The capability to successfully
interpolate data means that an AI can extend a pattern, but it can’t create new data.
However, sometimes developers can mislead people into thinking that the data is new
by using clever programming techniques. The presence of C looks new when it truly
isn’t. The lack of new data can produce conditions that make the AI seem to solve
a problem, but it doesn’t. The problem requires a new solution, not the interpolation
of existing solutions.
Seeing beyond the patterns
Currently, an AI can see patterns in data when they aren’t apparent to humans. The
capability to see these patterns is what makes AI so valuable. Data manipulation and
analysis is time consuming, complex, and repetitive, but an AI can perform the task
with aplomb. However, the data patterns are simply an output and not necessarily a
solution. Humans rely on five senses, empathy, creativity, and intuition to see beyond
the patterns to a potential solution that resides outside what the data would lead
one to believe. Chapter 18 discusses this part of the human condition in more detail.
A basic way to understand the human ability to see beyond patterns is to look at
the sky. On a cloudy day, people can see patterns in the clouds, but an AI sees clouds
and only clouds. In addition, two people may see different things in the same set
of clouds. The creative view of patterns in the cloud may have one person seeing a
sheep and another a fountain. The same holds true for stars and other kinds of patterns.
The AI presents the pattern as output, but it doesn’t understand the pattern and lacks
the creativity to do anything with the pattern, other than report that the pattern
exists.
Implementing new senses
As humans have become more knowledgeable, they have also become aware of variances
in human senses that don’t actually translate well to an AI because replicating these
senses in hardware isn’t truly possible now. For example, the ability to use multiple
senses to manage a single input (synesthesia; see https://www.mnn.com/health/fitness-well-being/stories/what-is-synesthesia-and-whats-it-like-to-have-it for details) is beyond an AI.
Describing synesthesia effectively is well beyond most humans. Before they can create
an AI that can mimic some of the truly amazing effects of synesthesia, humans must
first fully describe it and then create sensors that will convert the experience into
numbers that an AI can analyze. However, even then, the AI will see only the effects
of the synesthesia, not the emotional impact. Consequently, an AI will never fully experience or understand synesthesia. (The “Shifting data spectrum” section of Chapter 8 discusses how an AI could augment human perception with a synesthetic-like experience.)
Oddly enough, some studies show that adults can be trained to have synesthetic experiences,
making the need for an AI uncertain (https://www.nature.com/articles/srep07089).
Although most people know that humans have five senses, many sources now contend
that humans actually have far more than the standard five senses (http://www.todayifoundout.com/index.php/2010/07/humans-have-a-lot-more-than-five-senses/). Some of these additional senses aren’t at all well understood and are just barely
provable, such as magnetoception (the ability to detect magnetic fields, such as earth’s magnetic field). This sense
gives people the ability to tell direction, similar to the same sense in birds, but
to a lesser degree. Because we have no method of even quantifying this sense, replicating
it as part of an AI is impossible.
Empathizing
Computers don’t feel anything. That’s not necessarily a negative, but this chapter
views it as a negative. Without the ability to feel, a computer can’t see things from
the perspective of a human. It doesn’t understand being happy or sad, so it can’t
react to these emotions unless a program creates a method for it to analyze facial
expressions and other indicators, and then act appropriately. Even so, such a reaction
is a canned response and prone to error. Think about how many decisions you make based
on emotional need rather than outright fact. The following sections discuss how the
lack of empathy on the part of an AI keeps it from interacting with humans appropriately
in many cases.
Walking in someone’s shoes
The idea of walking in some else’s shoes means to view things from another person’s perspective and feel similar to how the
other person feels. No one truly feels precisely the same as someone else, but through
empathy, people can get close. This form of empathy requires strong intrapersonal
intelligence as a starting point, which an AI will never have unless it develops a
sense of self (the singularity as described at https://www.technologyreview.com/s/425733/paul-allen-the-singularity-isnt-near/). In addition, the AI would need to be able to feel, something that is currently
not possible, and the AI would need to be open to sharing feelings with some other
entity (generally a human, today), which is also impossible. The current state of
AI technology prohibits an AI from feeling or understanding any sort of emotion, which
makes empathy impossible.
Of course, the question is why empathy is so important. Without the ability to feel
the same as someone else, an AI can’t develop the motivation to perform certain tasks.
You could order the AI to perform the task, but there the AI would have no motivation
on its own. Consequently, the AI would never perform certain tasks, even though the
performance of such tasks is a requirement to build skills and knowledge required
to achieve human-like intelligence.
Developing true relationships
An AI builds a picture of you through the data it collects. It then creates patterns
from this data and, using specific algorithms, develops output that makes it seem
to know you — at least as an acquaintance. However, because the AI doesn’t feel, it
can’t appreciate you as a person. It can serve you, should you order it to do so and
assuming that the task is within its list of functions, but it can’t have any feeling
for you.
When dealing with a relationship, people have to consider both intellectual attachment
and feelings. The intellectual attachment often comes from a shared benefit between
two entities. Unfortunately, no shared benefit exists between an AI and a human (or
any other entity, for that matter). The AI simply processes data using a particular
algorithm. Something can’t claim to love something else if an order forces it to make
the proclamation. Emotional attachment must carry with it the risk of rejection, which
implies self-awareness.
Changing perspective
Humans can sometimes change an opinion based on something other than the facts. Even
though the odds would say that a particular course of action is prudent, an emotional
need makes another course of action preferable. An AI has no preferences. It therefore
can’t choose another course of action for any reason other than a change in the probabilities,
a constraint (a rule forcing it to make the change), or a requirement to provide random output.
Making leaps of faith
Faith is the belief in something as being true without having proven fact to back up such
belief. In many cases, faith takes the form of trust, which is the belief in the sincerity of another person without any proof that the
other person is trustworthy. An AI can’t exhibit either faith or trust, which is part
of the reason that it can’t extrapolate knowledge. The act of extrapolation often
relies on a hunch, based on faith, that something is true, despite a lack of any sort
of data to support the hunch. Because an AI lacks this ability, it can’t exhibit insight
— a necessary requirement for human-like thought patterns.
Examples abound of inventors who made leaps of faith to create something new. However,
one of the most prominent was Edison. For example, he made 1,000 (and possibly more)
attempts to create the light bulb. An AI would have given up after a certain number
of tries, likely due to a constraint. You can see a list of people who made leaps
of faith to perform amazing acts at https://www.uky.edu/~eushe2/Pajares/OnFailingG.html. Each of these acts is an example of something that an AI can’t do because it lacks
the ability to think past the specific data you provide as input.