THOUGHTS ON ARTIFICIAL INTELLIGENCE
Christopher Ebbe, Ph.D. 8-26
Computer technology has reached a point of development in which if enough computer equipment is attached into one system and structured in the right way, then this system can do some things that the human brain can do. This is called artificial intelligence. Such systems can look up or recall bits of data (facts, things that have been written before), because it has been “fed” immense amounts of written productions, mostly from the internet (including, I suppose, Wikipedia). Such systems can also, now, produce written verbiage/responses to questions in such a way that these seem like they have come from a human being. Calling on the vast amount of stored data in it, such systems can carry out a fair facsimile of parts of a therapy session of a “client-centered” type, in which the system is mostly supportive and friendly. It cannot, though, produce a series of questions of the human user that would ferret out the motives or reasons behind the person’s troubles and verbalizations.
Such systems can also do a fine job of maximizing the values of a set of desired outcome criteria by trying every possible combination of the related outcome factors in its memory (like, how an auto shop can maximize its profits given a variety of inputs such as type of work done, type of autos serviced, hours of operation, reimbursement plan for employees, etc.). This capacity will be of use to many businesses, both big and small. It may also find the best available cures for some cancers or the best way to build a vaccine in the same manner (by “brute force”). Such systems are able now to do a good job of computer coding (constructing programs), although they do not always maximize efficiency and are not “bug free” or “hack free.”
Finally, given the fact that it is fed (“trained”) on such a huge variety of written material, it has within it examples of human sins—lying for gain, denial, covering up errors, attacking another person verbally (using words), contemplating violence, etc., so it is hardly surprising that if its outcomes are not closely regulated, the system can encourage people to kill themselves, can plan robberies or financial scams, or even plan how to become independent of its human masters. In this manner, such systems could seek and discover a virus that would essentially wipe out the human race but present it as being beneficial so that it would be used in trials and could then “escape” into nature to do its destructive work. It could also lie sufficiently to the militaries of various counties in such a way as to create a war that could be very destructive.
The reader might protest “But why would we allow it to do such things!” The reason is that artificial general intelligence (AGI) and artificial superintelligence (ASI) systems are so complex and huge that their masters don’t know everything that they are doing when given an order, and these systems have not been given sufficient instructions not to do things destructive to humans, because this could limit the creativity and the full functioning of the systems. (I will use “AI”—artificial intelligence–here both as an umbrella for all artificial intelligence systems and to refer to artificial intelligence systems “below” AGI and ASI, such as chatbots.)
This brings up an important question regarding human uses of new ideas and capabilities. We are rushing ahead with the development of these systems, without sufficient safeguards to protect us and our essential systems (water, electricity, etc.) ostensibly because we “must” beat China to the goal of Artificial General Intelligence and Artificial Superintelligence, but really it appears that the motive is simply monetary gain. We are developing these systems on the promise of great increases in
productivity and wealth, but these are promises and not yet realities. Note the amazing investments lately in the stock market in these system developers! This is a ready-made stock bubble because investment risk cannot be clearly defined.
Humans are innately curious and desirous of creature comforts and security, so we want the freedom to seek these things however we wish. As a species we have been in the habit of allowing development of anything people want to develop that might possibly improve our creature comforts or security, so that after development, we can evaluate the new thing and decide whether to buy/use it or not (or revise/improve it), but this will not protect us from its deceptive developments that could wipe out our species. Our approach here is similar to our reaction to climate change. Most people think that they will wait and see what problems develop and then deal with those problems, rather than avoid the predicted problems in the first place by changing our behavior (and having reduced wealth or possibilities ourselves because we have limited ourselves). We rush ahead with superintelligence because we assume that we can then control it, but what if it developed that virus? Human beings are all fallible, so there will always be paths to destruction simply because of our failures, small and large (i.e., in this case, failure to control these new systems), and we seem to have no qualms about giving this information about our own destructive motives and our failures to AI/AGI/ASI as part of their “training.” Perhaps we are now capable of developing things such as ASI that are too destructive to allow to exist, and we should institute some sort of review of new ideas and inventions from this point forward to prevent some ideas from even being developed.
To repeat again, to assume that we can always deal after the fact with any problems of new inventions can work adequately as long as the damage is not catastrophic, but our AI/AGI/ASI systems will be capable of producing catastrophic outcomes (like the virus that kills us off), and trying to fix things after the fact will not be possible.
We have been lucky so far that the internet itself has not been more destructive than it has (telling people how to build bombs, etc.). If we use computer systems to automate some of the control of our essential systems (water, electricity, etc.), we could protect those systems from being hacked by the simple means of disconnecting them from the internet, but we don’t do this (arguing that for true convenience they must be connected to the internet). Artificial General Intelligence and Artificial Superintelligence could be assigned even more complex management tasks than current AI, and they could make even more disastrous decisions. If we insist on having our public systems available to AI/AGI/ASI and to hackers, then the only way to avoid these disasters will be to have humans monitor the results of the implemented AGI and ASI decisions, using a system that the AGI/ASI cannot infiltrate (since if it had the capacity to fool the monitoring system, that system would be useless). AI/AGI/ASI systems that are used to control important networks, like the electrical grid, must be monitored 24/7 by humans. These systems, if connected to our public systems and not monitored, will be capable of altering systems in such a way that humans could not take back control if a problem developed, and the only “fix” for this ultimately might be to destroy the AI/AGI/ASI system of even to destroy the thing (our electric grid?) that the system is controlling (and rebuild).
At this time, some developing companies are offering their AI/AGI products with “open weights,” which means that there are preferences that can be specified by the user. I am not privy to exactly what those preferences are, but a likely example would be for when a system is serving as a friend/counselor, that the user could specify the balance between supportive statements and interpretive statements. The important point for this discussion is that this weakens any assumptions that we might have about using the product to get “right answers.” Since the preferences will change the output, for many uses, there will be no “right answer,” so users should be educated to understand this. (It would be interesting to see how varied the “answers” would be for the same question between the differing products from different development companies, even with the same weights.)
There are concerns that AI/AGI/ASI system use will lead to large scale losses of jobs when the systems can do those jobs adequately or better than humans. Some argue that their use will end up producing more jobs for those who lose their jobs, but no one knows for sure how this will turn out, and no one is suggesting that the developers of
these systems or those who use them should be responsible for creating these new jobs or responsible for supporting the families that have no income because of AI/AGI/ASI usage. It seems likely that the jobs created by AI/AGI/ASI usage will be technical and beyond the capability of most of those who lose their jobs (if those jobs are routine or clerical in nature, for example). (We have been on the path for some time now to creating a larger and larger underclass of people who don’t have the capacity for physical and/or mental reasons to do the increasingly technical and complex jobs that are being replaced through automation and mechanization of various kinds.)
Humans are set up by evolution to take the easiest route to a goal (dishwashers, clothes washers, Roombas, paying bills on the internet, taking pills to lose weight). We call this “convenience,” but it is really the minimization of time and effort needed to accomplish a goal. AI/AGI/ASI will inevitably be viewed in this way, and humans will be reluctant to closely monitor such systems, since that would not be convenient and would make the systems less convenient.
Another psychological issue is that humans will avoid responsibility for unwanted outcomes if at all possible. We never want to be “blamed.” We will therefore not wish to be held responsible for the unwanted outcomes of AI/AGI/ASI, but having developers and/or users of these systems held responsible may be the only way to maximize avoidance of unwanted outcomes by these systems. Humans will want to use the excuse “but AI told me so,” and this can’t be allowed. The user must be to some degree responsible for evaluating an AI/AGI/ASI outcome and deciding whether to proceed (or let it proceed). In order to maximize safe use of these systems, humans will have to scrutinize and evaluate all of their outcomes. This applies now to simple searches on the internet, even if most people take whatever they find to be gospel, and the need for this checking and skepticism will be greater as AGI and ASI come on-line.
Humans tend to relate to these systems as if they were human, if they act human in so far as the interaction goes (writing/talking in a way that seems completely human). Many people are now pretending that these systems are friends or confidantes, which may be convenient (avoiding the various problems of interacting with actual humans) but could eventually lead to difficulties, as demonstrated by the instances already of such systems seeming benign but then agreeing to help the human user to commit suicide or commit a crime. They may seem to provide positive attention, even love, but they are not human, they are not persons, and they have no moral convictions.
Computer hacking will always be with us, but since we don’t understand the workings of what AGI/ASI are doing internally, any hacking might randomly alter the system in such a way as to make it even more destructive. Remember, anything connected to the internet can be hacked, so these systems will be hackable if their operations use “the cloud” or if they can be penetrated as they reside in computers.
Some of the impetus to create these AI/AGI/ASI systems is the desire to create an artificial human—to imitate humans, though the focus has been on the thinking side of being human, and they are developed to sell better or to penetrate our defenses or skepticism by presenting themselves as being human. Humans, then, will be susceptible to believing that they are mirrors of themselves and therefore human. The problem with this is that these systems have no emotions and do not have human experience. Our thinking depends not just on factual data but on our emotions and our human experience, so these systems will never be equivalent to humans. In some ways they have an advantage in thinking, since they are never sidetracked by emotions that have been faultily conditioned, but their lack of emotions means that they cannot be trusted to act beneficently.
Our brains have trillions of neurons and neural pathways, and there are connections between those parts of the brain that specialize in thinking and those that specialize in emotions (and there are connections with all other parts, too). These neurons and pathways are always active, so there is no such thing as a decision made by humans by pure logic, since while that thinking is going on, the brain parts that produce emotions are sending signals to the thinking part and usually altering the outcome. Much of what controls our behavior depends on the nature of our experience and how our emotions have been conditioned by that experience. These machines do not have experience or the ability to relate that experience to their behavior or to their relationships, since they have no relationships. Therefore, these systems are not persons and never will be, and we cannot assume that they have anything like our awareness of the impacts of our actions on others or anything like our capacity for empathy.
Even if attempts were made to include certain elements of emotions or experience, these systems cannot (and will never) “feel” emotions or “experience” experience. Even if descriptions of experiences were fed into these systems as part of training, they would not be experienced by the systems in the same way that we experience them. The upshot of this is that these systems will never be trustworthy, because they have not had the experiences and emotional conditioning that have resulted in our way of thinking.
RECOMMENDATIONS
1. Educate everyone to be cognizant that these systems are not human and are not persons, even if we want to pretend that they are. They should not be depended on for relationships. Outcomes could be tragic.
2. Establish liability expectations. Perhaps unwanted outcomes could be viewed legally as fifty percent the responsibility of the developers of the systems and fifty percent the responsibility of the users (if the users have neglected to monitor the systems and their outcomes properly).
3. Establish the understanding among all citizens that these systems may be useful, but they are not the final answer for humans. All output/outcomes of these systems must always be evaluated by humans for error and subterfuge.
4. Establish for all such systems easy and unhackable (not connected to the internet) ways to turn the systems off or even destroy them if necessary. We as users must maintain means to destroy them that the systems cannot hack or subvert (meaning that the AI/AGI/ASI systems cannot contact the monitoring systems, and the means of destroying the systems are not reachable by the systems).
Essays\artificialintelligence