By Dee Smith
A leading researcher in AI and cognitive science recently told me that we are nowhere close to developing any AI system that has the real-world cognitive and reasoning abilities of a rat. This leads to the question of whether there is something fundamental in consciousness and “alive-ness” that is not computational and that computers thus may never grasp.
In fact, there is. And this points up a significant element to consider in assessing AIs. For example, an important part of mathematics is called “non-computable functions.” They are called that because they cannot be solved by computation: by using any possible algorithmic instructions. These include Markovian functions, all of quantum dynamics, Cantor’s hierarchy of infinities, and even, astonishingly, proofs of basic functions of arithmetic! No system that “thinks” on the basis of algorithms can ever comprehend or solve these non-computable functions (although it could possibly simulate that it did). So, an important part of mathematics, and of the reality it describes, will always be unseen by—and missing from—algorithm-based AI, which is all of AI and all of computers as currently conceivable. If you mention this objection to key researchers at a major Silicon Valley AI company, their answer is “well, those things are not really that important.” I’m not kidding: I have done so, and the response has the feel of “pay no attention to the man behind the curtain.”
Added to all of this is the unreliability of computers in general and of AIs in particular. As anyone who uses any computer-based system knows, computers break down in many different ways all the time. They can only be kept running by having humans who are willing and able to fix them being available on a continuous, 24/7 basis. Homeostatic living beings do not of course require such intervention, not the simplest ones, such as viruses, nor the known most complex, like humans.
And of course, cyber systems are very vulnerable to malicious attacks as well, sometimes with results that become even harder to address. In fact, it has been demonstrated that sophisticated AI systems “attract” attacks from various malicious or just mischievous parties. Then there are the massive hallucinations that AIs exhibit, such as citing articles or legal cases that do not exist. Some of these are in fact worse in more advanced systems. And these are the machines that their proponents will task with solving all the problems of the world?
Electronic computing systems on which AIs operate are also subject to breakdowns caused by subatomic particles from space. Cosmic rays hit the Earth’s atmosphere, and the collisions create showers of high-energy particles like neutrons and muons. When such ionizing energetic particles hit the electronics of computers, they can cause “single-event upsets.” For example, a “bit flip” happens when a memory cell is hit by a particle and changed from a “0” to a “1.” Such events have caused unexpected errors and crashes in large systems like supercomputers. Their frequency has forced computer scientists to deploy so-called “error correcting code” to try to find and fix these events. But when a single event error happens in a critical system (for example, an autonomous AI system) at a critical moment, the outcomes can be catastrophic. Biological life has evolved over more than 3 billion years on Earth to be more robust and not so immediately vulnerable to cosmic rays and other natural radiation, which bombards us all the time. Electronic computing systems, which have a history of at most 80 years, do not have the same resilience.
If you told someone 25 years ago that they would have a device roughly the size of a deck of cards in their pocket that could tell them how to drive from a specific office in Houston, Texas, to a specific hotel in Anchorage, Alaska—step by step, turn by turn—most people would have commented that such a device would have to be astonishingly intelligent. Today, of course, we don’t think of our phones, and their Google Maps or Apple Maps, that way at all. Similarly, some of our feelings (and they are feelings) about AI may simply have to do with its novelty at this point.
There is a great deal of sloppy thinking, driven by both fear and exuberance, going on in this domain. The question is not whether AIs are conscious and/or dangerous because they think like humans. I would argue that to the extent they may think, they think like AIs, not like humans.
Many animals are now recognized as conscious, but they do not think like humans. Alligators do not think like humans, but they are dangerous, and a swamp full of alligators is even more dangerous.
Perhaps the most fundamental issue at play in all this is that human beings want a savior. If this is not to be a religious savior, the fashion today is for it to be a technological one. AI is the perfect candidate for this, as it seems magical: you cannot see how it does what it does, nor what is going on inside it, and it can certainly look aware/volitional/conscious/alive . . . or something similar but unaccountable.
It is curious that people choose to believe that any Large Language Model AI—trained as they are on a significant portion of all human output as recorded in text and images on the Internet—could possibly be a candidate for a savior that would “fix” the world. In no small part, this is because there is at least as much bad as good in human output, and further, because this assumes that the AIs can learn through some moral evolution, and/or by human instruction and intervention, to act only benevolently.