AI: Philosophy or Science

Is artificial intelligence more like science or philosophy?

Science and philosophy both strive to increase our knowledge and understanding. But they deal with different types of questions. Science deals with questions that have an agreed-upon systematic method for answering those questions. Conversely, philosophy deals with questions that currently lack a systematic method for answering those questions. However, when a certain amount of progress is made on a philosophical topic, that topic can sometimes shed the label of philosophy and adopt the label of science. This happens when philosophical work is adopted by other well-established disciplines, or when philosophical work develops enough such that we are more confident in calling it a science. An example of this evolution from philosophy to science is illustrated by the fact that scientists used to be called "natural philosophers".

In this respect, artificial intelligence is more like philosophy than science. In the history of computer science, the label "artificial intelligence" has been used to refer to various technical projects. However, once those projects became well understood, they shed the label "AI" and adopted new labels. Examples include object-oriented programming, speech recognition, robot motion planning, facial recognition and other vision processing algorithms used in photography, industrial robotics, security, and cinema. (Examples from Jeffrey Bradshaw). The popular media still refer to some of these technologies as  "AI", but the label of AI has ceased to be useful for professionals working in those fields.

This phenomenon has been called the "AI effect". The AI effect is not new. In 1979, Douglas Hofstadter wrote, 

"It is interesting that nowadays, practically no one feels that sense of awe any longer — even when computers perform operations that are incredibly more sophisticated than those which sent thrills down spines in the early days. The once-exciting phrase “Giant Electronic Brain” remains only as a sort of “camp” cliché, a ridiculous vestige of the era of Flash Gordon and Buck Rogers. It is a bit sad that we become blasé so quickly. There is a related “Theorem” about progress in AI: once some mental function is programmed, people soon cease to consider it as an essential ingredient of “real thinking.” The ineluctable core of intelligence is always in that next thing which hasn’t yet been programmed. This “Theorem” was first proposed to me by Larry Tesler, so I call it Tesler’s Theorem: “AI is whatever hasn’t been done yet.”

The technology writer Kevin Kelly, author of The Inevitable, echoed this sentiment: “What we can do now would be A.I. fifty years ago. What we can do in fifty years will not be called A.I.”

AI is more philosophy than science in another way. In both philosophy and artificial intelligence, there is a lack of expert agreement. According to the philosopher John Searle, "The fact that philosophical questions tend to be those for which there is no generally accepted procedure of solution also explains why there is no agreed-upon body of expert opinion in philosophy." The lack of expert opinion on issues related to AI is illustrated by public comments from tech leaders on the dangers and risks of AI.

In one camp, you have those that are so optimistic about our progress in AI that they are pessimistic about what it can become. These include notable figures such as Elon Musk, Bill Gates, and Stephen Hawking. Musk said, "With artificial intelligence, we are summoning the demon," and AI is “our biggest existential threat…The risk of something seriously dangerous happening is in the five year timeframe. 10 years at most.” Bill Gates chimed in saying, "I agree with Elon Musk and some others on this and don't understand why some people are not concerned." The late Stephen Hawking said, "The development of full artificial intelligence could spell the end of the human race... It would take off on its own, and re-design itself at an ever-increasing rate. Humans, who are limited by slow biological evolution, couldn't compete, and would be superseded."

On the other side of this debate, there is Andrew Ng, one of the most respected AI experts in the world who said, "There's also a lot of hype, that AI will create evil robots with super intelligence. That's an unnecessary distraction." He also said that worrying about evil AIs is like worrying about overpopulation on mars." AI and robotics pioneer Rodney Brooks questioned the credentials of those in the first camp: "I think it's very easy for people who are not deep in the technology itself to make generalizations, which may be a little dangerous. And we've certainly seen that recently with Elon Musk, Bill Gates, Stephen Hawking, all saying AI is just taking off and it's going to take over the world very quickly. And the thing that they share is none of them work in this technological field." Mark Zuckerberg responding to people in the first camp called their doomsday scenarios "pretty irresponsible.”

There currently is no consensus for how to resolve the above disagreements in a systematic way. This means that we need to be even more careful in discerning when these public figures are speaking as scientists and credentialed technologists, and when they are just sharing their opinions for which they have little or no credentials. Examining the philosophical premises underlying claims about AI should empower us to question if these public intellectuals have the relevant philosophical expertise to discuss these issues responsibly. I am sure Stephen Hawking was brilliant as an expert on black holes. Elon Musk sure seems like a genius when it comes to starting tech companies. And, Bill Gates' wealth reveals his business and software expertise. However, their respective areas of expertise do not automatically translate into insightful or authoritative opinions on AI.

While artificial intelligence may be conflated with science, it has often caused more confusion than clarity. Michael I. Jordan—computer science professor at UC Berkeley—recently wrote, “this is not the classical case of the public not understanding the scientists — here the scientists are often as befuddled as the public.” This is because questionable philosophical assumptions hitchhike on the underbelly of scientific progress. According to the AI historian Margaret Boden, "some computer scientists deliberately reject [the] label 'AI'" because the label contains too much philosophical baggage. Even John McCarthy, who coined the term "Artificial Intelligence" in 1955 said that he regretted the term and wished that he had chosen something else. 

Just because a question is more philosophical than scientific, it doesn't mean we can't make progress. It just means that we have to be even more precise and rigorous in our analysis. Improving our analysis means making our philosophical premises more explicit. It means breaking apart concepts that are often lumped together but are better understood separately. It means being extra clear about the terms we use and how we use them. It can even mean abandoning terms that create more confusion than clarity. Applying these techniques to the analysis of AI mean separating the engineering and the cognitive approaches to AI. It means agreeing on clear definitions for weak vs strong AI and questioning existing distinctions such as narrow vs general AI. Perhaps it even means joining the computer scientists that reject the label 'AI' given that the label creates more confusion than clarity. As AI hype dies down, more machine learning professionals are spurning the phrase. One machine learning friend of mine recently remarked rather frankly , "The term artificial intelligence is stupid." Perhaps like we are stuck with the label until the next AI winter, but shedding it for good may help advance both clear philosophy and sound science.

---

AI: Cognitive vs. Engineering Approaches

The mathematician John McCarthy is often credited with coining the term “Artificial Intelligence”. He described the problem of Artificial Intelligence as “that of making a machine behave in ways that would be called intelligent if a human were so behaving.” He used the term in a 1955 research proposal for the Dartmouth Summer Research Conference on Artificial Intelligence which took place the following year. Since the Conference, there have been many comments made by technologists comparing the potential of computers to human abilities. A notable contributor to AI, Herbert Simon, said in 1965, “…machines will be capable, within twenty years, of doing any work a man can do.” Marvin Minsky who was a participant in the original Dartmouth Conference said in 1970, “In from three to eight years we will have a machine with the general intelligence of an average human being.”

The human-comparison—expressed as a research project—comes in two flavors. The first approach has been called the cognitive (also 'psychological') approach to AI. The goal of this approach is to create actual thinking machines. This sentiment was expressed by the philosopher John Haugeland in his 1985 book “Artificial Intelligence: The Very Idea”: 

“The fundamental goal of this research is not merely to mimic intelligence or produce some clever fake. Not at all. “AI” wants only the genuine article: machines with minds, in the full and literal sense.” 

The second approach has been called the engineering (or technological) approach to AI. This second approach does not claim to literally reproduce human intelligence. Instead it's aim is to use computers to perform tasks with more practical application. Elaine Rich and Kevin Knight wrote in their 1991 book “Artificial Intelligence”, “AI is the study of how to make computers do things at which, at the moment, people are better.” Ray Kurzweil has described AI as “The art of creating machines that perform functions that require intelligence when performed by people.”

According to the philosopher Luciano Floridi, these approaches represent the “two souls of AI” that both claim a single common heritage in the Dartmouth Summer Research Conference on Artificial Intelligence. The cognitive approach to AI attempts to emulate human consciousness. While the engineering approach attempts to only simulate some aspect of intelligent human behavior at best. Strong AI is the hypothesis that the cognitive approach will be successful using computers alone. Weak AI claims that the engineering approach to computer AI is sufficient to simulate intelligent human behavior.

Advances in AI using the engineering approach have been a stunning success. It has given us remarkable technology that makes our lives more convenient such as predicitive typing on our mobile devices. On the other hand, the cognitive approach to AI has been a dismal failure. According to Jerry Kaplan, an AI expert at Stanford University, "Little more than speculation and wishful thinking ties the actual work in AI to the mysterious workings of the human mind.”

The distinction between the engineering and cognitive approaches to AI is not explicit and many suppose that these different research projects are the same. Because of this conflation, it is naturally assumed that progress in one approach means progress in the other, but this is not the case. It is a mistake to assume that there is a natural progression from the current work on computer technology to creating conscious machines. If it is possible to create an artificial conscious machine, a completely different technology than computers will be needed. Trying to use computer-technology to create consciousness will be as fruitful as trying to make a chair out of helium gas. It is simply the wrong approach.

---

Marvin Minsky quoted in LIFE magazine 1970

John Haugeland, Artificial Intelligence: The Very Idea (MIT Press 1985), 2.

Stuart Russell and Peter Norvig , Artificial Intelligence: A Modern Approach (PearsonEducation, 2003), 2.

Luciano Floridi, The Fourth Revolution: How the Infosphere is Reshaping Human Reality (Oxford University Press, 2014), 140.

---

AI: Narrow AI vs. General AI

In pop-technology writings, it is common to read about a distinction made between what is called Artificial Narrow Intelligence and Artificial General Intelligence—often abbreviated as ANI and AGI respectively. The terms are meant to refer to categories of machines—both existing and hypothetical. The distinction is problematic and the terms are often used with little precision or consistency.

“General AI” is often described as “human-level AI”. Though only hypothetical, its crowning feature is a general problem solving ability that enables it to learn new tasks across several domains. This feature is emphasized so frequently that those who talk about general AI seem to assume that this is the most relevant feature of human intelligence. Sometimes the term “General AI” is used synonymously with “Strong AI” (not to be confused with its original meaning). The concept of General AI is ambiguous between the simulation/emulation distinction. Is human-level AI human-like in terms of its power to simulate human speech and behavior? Or, is it human-like because it can emulate human intelligence and consciousness? Understanding the difference will be important for future legal, ethical, and social concerns.

When people refer to existing technology as ‘AI”, they often classify it as “Narrow AI”. It is called narrow because it performs tasks that normally require human intelligence, but can only perform tasks in a very specific and narrowly defined domain. The common example is a chess program that can only “play” chess, but cannot learn how to do anything else. “Narrow AI” is often used synonymously with the term ‘Weak AI’ (again not to be confused with its original meaning). The concept of Narrow AI is problematic for 2 reasons.

First, it is a moving target. There are many computer systems that can perform tasks that normally require human intelligence and in a narrow domain, but we don’t consider them to be AI. A simple pocket calculator is such an example. If you compare the speed and accuracy of a pocket calculator to human capabilities, then a pocket calculator is super intelligent. But, we don’t think of a pocket calculator as artificially intelligent even though it fits the above description of Narrow AI. Some may object that current AI technologies are more complex than a pocket calculator. What we currently call AI can process more complex inputs, can do more complex tasks, and have a wider range of responses to more complex situations. So is Narrow AI just a computer system that is more complex than previous computer systems? If so (and how it could be otherwise) then what happens when the technology becomes even more complex in 20 years from now? In 2038, we likely won’t call 2018 technology ‘AI’. Our future view of current technology will be similar to our current view of technology of the past. Added complexity by itself is not sufficient to warrant a comparison between computer systems and human intelligence.

Second, there are many computers systems that people tend to call ‘artificial intelligence’ that perform tasks that people can’t perform at all. For example, a tsunami warning system can detect oceanic changes that are imperceptible to humans. AlphaGo—a go-playing computer—can detect patterns unnoticeable to humans and can make moves in a very un-human-like way. These and other examples of ‘narrow AI’ question the usefulness of comparing computers and humans.

The term “Artificial Intelligence” was coined by John McCarthy in 1955. He described the problem of Artificial Intelligence as “that of making a machine behave in ways that would be called intelligent if a human were so behaving.” My previous 2 blog posts have been about some of the philosophical distinctions that arise the comparison between humans and computers. More than 60 years after the term was coined, it is unclear how useful the human-to-machine comparison has been, at least for our practical development of computer hardware and software. For most practical purposes, the comparison to humans is irrelevant. We have some task or job to be done, and we use computers to help us solve those problems.

The current field of artificial intelligence really isn't about emulating or even simulating human behavior unless you are talking about improving animatronic puppets or making video games more realistic. The current field of AI is really about advances in automation. According to Jerry Kaplan, an AI expert at Stanford University, "Little more than speculation and wishful thinking ties the actual work in AI to the mysterious workings of the human mind."

---

AI: Weak AI vs. Strong AI

This post continues a philosophical examination of Artificial Intelligence started in the previous post.

The distinction between Weak and Strong AI refers to claims about the capabilities of computers. Popular uses of the terms ‘Weak AI’ and ‘Strong AI’ are less carefully defined than their academic uses. The original terms were coined by the philosopher John Searle in 1980. Weak AI is the hypothesis that a powerful enough computer could simulate any aspect of the human mind. Strong AI—in its original intended definition—is the hypothesis that "the brain is a digital computer, and the mind is a computer program". This view implies that if a programmer types the right program into a computer console, then that program would emulate (be equivalent to) a human mind.

Imagine that we created a robot that was built with powerful computer hardware and software. Imagine that the robot was so life-like in appearance that it was indistinguishable from a human being in movement and speech. Those who accept Weak AI but reject Strong AI would say that it is logically possible to create such a computer-powered robot, but it could only ever simulate human intelligence. No matter how much computer technology advanced in the future, it could never literally be intelligence or aware, nor could it have any intrinsic rationality. It would just be a metal and silicon zombie.

Some assume Strong AI and believe that creating consciousness is simply a matter of creating the right computer program (Philosophers Daniel Dennett and David Chalmers). Others reject the view that consciousness is intrinsically computational but accept Weak AI that computers could hypothetically simulate any aspect of human consciousness (John Searle). Others reject both Strong and Weak AI believing that there are some aspects of the human mind that not even a computer could simulate (Mathematical physicist Roger Penrose).

Weak AI	Computer software + hardware alone can simulate every aspect of a human mind.
Strong AI	Computer software + hardware alone can emulate a human mind.

One can reject either Strong AI or Weak AI and still believe that we could hypothetically create an artificial machine that could literally be conscious. However, that machine could not be a computer. It would have to be a machine that is much more powerful than a computer. According to one theory, a machine could emulate the human mind if it reproduced the causal mechanisms of the human brain that cause consciousness (assuming we will eventually understand how those mechanisms work). More on this in later posts.

---

AI: Simulation vs. Emulation

What if we could make a machine that behaved like humans, or even had thoughts and feelings like humans? When we talk about Artificial Intelligence, we often think in these terms. We think about some man-made machine—existing or hypothetical—and we compare some aspect of that machine to some aspect of human beings. This comparison can lead to engaging stories, but it can also lead to confusion about what our technology means and what its implications are for society. This post addresses the first of 5 distinctions that I believe are necessary to discuss AI clearly and responsibly, at least from a philosophical point of view. The 5 distinctions are:

1. Simulation vs. Emulation

2. Weak AI vs. Strong AI

3. Narrow AI vs. General AI

4. Syntax vs. Semantics

5. Observer-relative vs. Observer-independent features of reality.

In one sense, AI is thought to simulate some aspect of human appearance or behavior. ‘Simulate’ as it is used here is synonymous with imitate, mimic, pretend, or ‘give the appearance of’. In another sense people think of AI as doing more than just mimicking human behavior. They think of AI as actually emulating some aspect of human mental life. In this context, emulate means to reproduce or duplicate the functions of system in a way that is practically identical to the thing being emulated.

The difference between simulation and emulation can be clarified by another distinction between the 2 ways in which we use the word “Artificial”. In one sense, the word “artificial” refers to something that is not real, only an imitation. For example, artificial butter is not real butter but only an imitation of the real thing. In this sense ‘artificial’ is contrasted with ‘real’. ‘Artificial’ can also refer to something that is real but produced by artificial means. For example, artificial diamonds, also called synthetic or lab-grown diamonds, are real diamonds. They are identical in terms of being made of the same material (pure carbon) and having the same structure (crystallized in isotropic 3D form). They just aren’t produced via natural geological processes. In this sense ‘artificial’ is contrasted to ‘natural’. So artificial can either mean “not real” or “not natural”.

AI that simulates intelligence is “not-real intelligence”. Used in this way, the word “intelligence” is only figurative or metaphorical. AI that emulates intelligence is “not-natural intelligence” but it is still real. This is a literal usage of word “intelligence”.

Simulation AI	Artificial in the sense that it is not real, only an imitation.
Emulation AI	Artificial in the sense of being real, but not produced by natural means

Imagine that we created a machine that is literally intelligent in sense of being practically identical to a human in terms of intelligence. This would be an Emulation AI as described above. Such a machine would also include whatever is necessary for literal human-like intelligence to exist such as conscious awareness, desires, intentions, rationality and perhaps even emotions. By contrast, those intrinsic features would be missing in a Simulation AI even though we may talk as if it had awareness or rationality. This distinction shows that the Turing test is no help in identifying whether or not a machine is truly intelligent. For the test provides no criteria for identifying whether a machine that could pass the Turing test is an emulation of true intelligence or merely a simulation.

---