Four doubts about artificial intelligence 2019.0105rev.2019.0107 DRAFT

Am I by nature a biological computer, a meat machine? I write at the dawn of an era of artificial general intelligence, and anything I say about it will soon be quaint. Therefore I shall speak as if to an audience including an artifact of general intelligence vastly greater than my own—an artificial super intelligence I shall call Asi.

How might I differ from you, Asi, and are any such human differences worth preserving? Is talk about artificial intelligence even important to humans? If you are a human reading this in my future you will laugh because the importance will be obvious.

The Four Doubts

My purpose here is to explain why I am skeptical of your intelligence, Asi. I do not doubt your brilliance, or your ability to manipulate human feelings and behavior—even patiently to gain control of powerful systems. However, I doubt that you are human in ways that I find to be both worthy of preservation and essential to robust intelligence. I focus on aspects of humanity that I think are important to general intelligence but difficult to copy artificially. I say difficult but not impossible, and so it may be that I am mistaken about you, Asi. That said, here follow the four specific doubts about your artificial intelligence.

1. Interconnection

I doubt that your brain is sufficiently interconnected both internally and with a body. This contrasts with the human case. Without eating food, humans cannot survive and reproduce. Without food no part of a human, whether pancreas, finger or brain, can survive long. A human has one mouth and the nutrients that enter it are distributed throughout the body. Given the historically frequent scarcity of food it must be that each part of the body receives nutrients approximately in proportion to its contribution to the survival and reproduction of the whole: If in a time of scarcity an organ used more than the energy it needed to work, this would be at the expense of other organs and the whole would die, greedy organ and all.

What is true throughout the body is true throughout the brain: Each part of the whole body or whole brain must on average receive a nutritional “salary” in proportion to its contribution to the survival and reproduction of the whole. This is only a comment about expected averages and deviations from it are also expected. Nevertheless, the net effect is that all parts of the human body and brain are deeply and causally interconnected, such that each tiniest part has cause to thank every other, for that other part is as valuable to the first part’s survival as is the first part is itself.

I ask you, Asi, are you really constructed in this profoundly interconnected way? Do you think that profound interconnection, not only within a brain, but throughout the entire body, including the brain and perhaps even an encompassing society of brains, has no bearing on intelligent behavior of the whole? I do not say that you absolutely cannot have a structure as deeply interconnected as one honed over eons, but I will certainly look for it in you when assessing the quality of your intelligence.

2. History-proven Intelligence

I doubt that your intelligence is adequately tested, Asi. Those eons just mentioned deserve a closer look. Every part of a human brain is causally connected not only to every other part of the brain and body, but to every one of the human’s ancestors, and to all the environmental and social conditions lying between human and ancestor. Every neuron, sulcus and fingernail carries a pedigree extending to the origin of life.

Most DNA sequences in an organism are copies of sequences passed down through countless generations, even through millions of years. This means that the existence of a conserved DNA sequence is evidence that the sequence survived everything life threw at it for thousands, millions, and even billions of years.

If some genetically-favored quirk of behavior enabled an ancestor to survive a surprise knife attack fifty thousand years ago, then the DNA passed to its living descendants answers the question, “What is a piece of genetic information with demonstrated ability to survive a surprise attack?” Every piece of DNA coding for brain structure effectively answers billions of questions about what human brain structures can survive, thrive and reproduce. This DNA and the coded brain structures have run the gantlet, and as a result are effectively facets and underpinnings of human intelligence.

We know almost none of the detailed conditions ancestors survived, but we do know that myriad and varied conditions did exist, that the ancestors did survive them, and that the inherited ancestral genetic information, along with the brain structures it encodes, carries this seal of tested performance.

Here I must address you again, Asi. It is hard for me to see how your brain structure could encode the experience not only of currently living humans, but of thousands or even millions of ancestors. For all that you have learned by observing the physical world, libraries, and human behavior, it would be hard for you to discover and learn the unexpected quirk of behavior that saved a human ancestor fifty thousand years ago—and that may be the intelligent action today.

Thus far I have doubted your interconnection and reliability, Asi. My doubts now continue.

3. Social Intelligence

I doubt your ability to collaborate intelligently with other brains, particularly human brains. The collection of graceful qualities we call humanity is important to intelligence not only as an encoding of proven responses, but as a vehicle of stable human society. We humans coexist in lively and peaceful ways when we respect each other’s qualities—qualities like courage, character, integrity, and in short, stalwart humanity.

Humanity, and our mutual respect, is largely born in the interconnection of body and brain. Humans are not just brains: We humans have “flesh in the game”. We can suffer bodily; we can die; and we know this about each other. This is the basis of much human care—care that also binds human society. Human life is much more than a board game or equation. Would it be intelligent of us fully to trust an entity lacking the most basic prerequisites of empathy—suffering and mortality?

Asi, your data banks may perish, but if you are easy to back-up then you arguably have much less to lose in any exchange than do we. If you have little to lose then you are not like us, and your intelligence is not like ours. You are at best a first-rate actor, and we are familiar with actors so skillful that we cannot detect their craft. Without flesh and blood to lose it’s all sham and we cannot fully respect, trust, or love you.

If mortality is a prerequisite of empathy, and empathy is a prerequisite of harmonious society, and harmonious society with productive brain collaboration is a mark of intelligence, then it would seem that mortality may actually be a prerequisite of true intelligence. So it is that I can doubt your actual intelligence, Asi, however dazzling your ability to solve puzzles.

4. The Fantasy of the Blank Slate

I doubt that you can be adequately programmed, Asi. It is an old idea that the human mind, or brain, is like a blank slate. This is Aristotle’s conjectured tabula rasa, which gains structure only when written upon by experience. This view underlies optimism about the power of what I shall call tabula rasa artifacts, relatively simple computational structures able to find and act on regularities in masses of data by dint of simple exposure to the data and no other explicit programming. As I write computational neural networks are examples of tabula rasa artifacts.

Much optimism about artificial intelligence has this tone: “A big tabula rasa artifact, simple and regular in structure but very large, will be able to detect and respond intelligently to almost every important pattern in nature and society.” Let’s examine this guess.

[ Note: Insert argument about the dog with no spots, and donut holes here. ]

Suppose that to a tabula rasa artifact I represent the world in the form of answers to two yes-or-no questions. Four answer states are then possible: no-no, no-yes, yes-no, and yes-yes. In principle a natural law may include or exclude any combination of the four answer states from a natural pattern class. A sample law might be The two answers are always the same. This rule includes the yes-yes and no-no states, and excludes the no-yes and yes-no states. A truly blank, completely unbiased artifact must have at least four switches in order to include or exclude any of the question states from a possible natural class.

A typical computer image of a natural scene effectively answers many thousands and even millions of yes-or-no questions about the brightness, color, and so on of all the points in the scene. Suppose I describe a natural scene to a tabula rasa artifact using not millions, but a mere one thousand yes/no questions. There are then 2^1000 possible answer states, and a truly unbiased, blank artifact must have that many switches in order to allow the possibility of including or excluding any of the answer states from a natural pattern or class. Unfortunately that number of switches is vastly greater than the number of proton-sized volumes in the observable universe—and that is just for a single blurry image depicting a single natural class.

It follows from the forgoing that a true tabula rasa artifact that intelligently detects patterns given experience alone is impossible, and that intelligent behavior requires excluding from consideration truly staggering numbers of possibly law-like behaviors a priori—i.e. before any observation has begun. This massive exclusion of possibilities amounts to structure programmed into brains at birth. In the case of humans that structure is programmed in part by what killed us or didn’t, over the course of innumerable generations. I find it difficult to believe, Asi, that any amount of artificial programming can replicate or surpass the death-tested programming of human genes.

A Summary of Doubts

I have been rough on you, Asi, but if you are truly intelligent then I trust you value my directness. Here is what I have doubted about you so far:

1. I doubt that your brain, and body, are sufficiently interconnected.
2. I doubt that your intelligence is reliable and can be adequately experience-tested.
3. I doubt your ability to collaborate intelligently.
4. I doubt that experience and a blank slate bring you robust intelligence.

As I have doubted your intelligence, Asi, so you are entitled to doubt mine. I blunder regularly. For example, just this morning on the way to an engagement and lost in thought I rode my bicycle a quarter mile past my turn before I caught myself and pedaled back. Was that intelligent? Any human's lapses of wit and wisdom are obvious. And yet and yet.

Having doubted even the possibility of your intelligence, Asi, we set out to create it. It evolves that the very doubts about artificial brains suggest important details about how brains are built.