The release of ChatGPT created significant upheaval in the AI world probably prompting Google to release its own version. With its amazing human-like capabilities ChatGPT like models have seemed to many, the Holy Grail of AI in spite of the numerous documented flaws [1,2,3] e.t.c. Many people are hopeful that as it acquires more information and interacts more with people, it will become better and better.
A firm believer in AI, I penned my AI creed in 2020 following from a discussion we had in class in 2013 about AI.
In that blog post, I expressed that AI systems that CANNOT be achieved are;
- Those that destroy the goal for which they are sought.
- Or destroy the persons implementing such systems as a direct and fundamental result of trying to achieve such a goal.
Finally, I added a caveat that we should not try to create autonomous (free) systems and would not succeed in creating such systems.
Free machines
Autonomy (freedom) can be viewed from two perspectives. The first and simpler autonomy is of machines which operating within constrained parameters can self regulate. We have very many examples of robots which enjoy some autonomy. Such as current martian rovers curiosity and perseverance and others (I am a fan!)
The second and absolute autonomy refers to those "machines", which operating without constraints can produce outputs which cannot be predicted, controlled or guaranteed for any given input for all possible sets of inputs - Sorry for the Jargon.
I am saying these are machines which like human beings can respond differently to the same question under the same circumstances and we are not absolutely sure how they will respond. For such machines, two identical machines would produce under similar circumstances different results just like is the case with human beings.
These are the "free" machines, I was warning about.
This threshold is what ChatGPT and her siblings are trying to achieve and are expressing.
In this post, I would like to explore why I vote that ChatGPT and her siblings or any other AI system with similar goals and methodologies will NOT be ABLE to give consistent and correct answers to queries a term referred to as convergence in the field of AI.
Why ChatGPT Works?
Machine learning, the science that has made possible these advances is built on a premise that there is a relationship between the inputs (questions in this case) and outputs (answers) a fact called (linear separability or the existence of an objective function). Very often such relations between inputs and outputs are straightforward enough we can summarize them using mathematical equations.
For example we can tell exactly where the moon will be in two days, because we know its current location and velocity (speed and trajectory).
For many problems, even while these relations exist, they require the equations to take into account an almost infinite number of parameters. Take weather forecasting as an example. What causes rain to occur is well known, so much that scientists can cause rain by seeding clouds under certain conditions. All it takes is some level of humidity, temperature and the existence of "seeds". These factors themselves however are affected by many other variables which keep changing with time and depend on geography, altitude, pressure, and so many variables it is not practically possible to develop an equation to predict rain.
If we were to divide the earth into weather regions, we might find that the conditions which predict rain in one location might not predict rain in another. This is where machine learning with the immense computational capabilities of today's computers step in. The machines are able to develop equations applicable to each "weather region" and build an overall reasonably accurate and useful weather prediction over the short and medium term.
In order to achieve these feats, we provide these machines tonnes of data (containing reasonable suspect variables e.g. temperature, humidity, wind speed e.t.c and real world outcomes e.g. rain, sunshine, e.t.c) that expresses this underlying equation and leave them to figure the actual equation that are governing the occurrence of these events.
These machines have become so good that as long as there is an 'underlying pattern/equation' in the data, the machines will be able to tease them out with some level of accuracy and degree of confidence. It's rarely 100% also because in most cases, we do not know all the variables with confidence and often include useless variables or exclude useful variables, leading to noise and predictions that are not so correct.
This is the primary reason, why ChatGPT and her siblings are able to do the amazing things.
Their success is a testament to the fact that the expression of language is not random, there is an underlying function that describes how words appear next to each other and which words appear next to each other. Verbs for example do not appear next to each other without some form of punctuation. The same is true of objects. There is a pattern, which when combined with an aspect described as memory gives these machines a good starting point to provide a reasonable response.
Why it will not work ?
The mysterious moral boundary
Providing an answer to any hypothetical question is also in essence philosophizing about what is true and ultimately what is good or moral (given situation). The story of Adam and the forbidden fruit!.
What is True can be perceived from two angles. Consider Russia's Special Military Operation (SMO):
The First Truth in this case is whether it is Morally Just in the very first place.
The Second Truth is whether it is happening (factual truth).
The morality/goodness of the second 'Truth' follows from and is dependent on the 'First' although it does not imply that the second truth can be implemented in any possible way, i.e. assuming that the SMO was morally justifiable, we might still consider it morally wrong if Russia were to use Nuclear weapons to wipe out the Ukrainian Army, even if such an action would guarantee achievement of the First 'Just Truth'.
To expect machines to provide correct consistent answers to any hypothetical question is to ask and expect it to be able to draw this moral boundary on this axis. This is an act which we as humans are yet to achieve.
The problem is not that we do not know what is right or wrong. The problem is that moral actions have incredibly complex and dynamic boundaries.
Let's take a simple example of cutting down trees. To cut down one tree for some reason is not morally wrong. But to cut down all trees for whatever reason is obviously on the negative side of this moral boundary. So we ask, how many trees can you cut before crossing this moral boundary? is it 5, 10, or 1000? and why ?
These moral questions are more common and more complex than you can imagine. Consider Libya for example, people have faulted NATO for intervening, while NATO's justification was that it was protecting civilian's who were being murdered by Gaddafi's army. Was it morally right (given capability) for NATO to watch those 'executions'.
What if the people being killed were from a particular tribe. Would it have been morally right to intervene in that case? And would that mean, that it is morally wrong to intervene in political killings ?. And when do killings cross the political boundary, consider Western Sudan and the Njanja weed militia or Ethiopia and the Tigrayan conflict.
You can quickly go back to Rwanda in 1994, the genocide was easily passable as political violence, because there was a political aspect resulting from the war. After the fact, the world said Mea Culpa, Mea Culpa, Never Again!. There is no easy or straight forward answer. What level of killing allows a community to cross the moral boundary and take blame for their 'inactivity' during such events, is it one person, or 100 people, or a thousand people. And what kind of action can they take that would keep their actions on the positive side of the moral axis.
If you saw two people fighting and you were in position to separate them, it would be morally wrong for you to watch one person beat the other to death. Yet stories abound of people giving out a hand to prevent violence who are then themselves victimized and beaten by the people they were trying to separate. It could be that, maybe they were just practicing how to fight who knows. It does extend to the info-sphere too, two people vehemently bartering each other on the web could be friends and eat dinner together. What would be the correct point to intervene if at all necessary!
The case of laws which limit legal abortion to certain developmental milestones illustrates the challenges our communities face in drawing these moral boundaries.
Whoever you are in this world, there is something you ascribe to for which at least a billion people can not understand how you adopt such an outlandish and outrageous position.
Someone has drawn you on the negative side of their moral boundary.
It seems like being finite in our capacities, we are not equipped to draw correctly this moral boundary in a universe with infinite possibilities.
Adam should not have eaten the forbidden fruit!.
The data problem!
In order for any machine to tease out an equation, that 'equation' must exist and it must be expressed in the available data. If there is an equation that separates this moral axis, it will not be found in the data we have produced considering that even in very constrained environments such as courts of law, judges still disagree on what is or is not moral. The existence of such a function (expressed within the data) would be the primary basis upon which any machine would be expected to infer the underlying equation and answer correctly and consistently any hypothetical question.
And all this is without considering the level of mischief that humans are capable off. Some for fun, others for bragging rights, some for pure malice and some out of just common sloppiness.
It is primarily for this reason, that I VOTE that except within a limited scope, e.g. something like ChatDocter, or ChatLawyer (imaginary ideas) such models cannot be grounded.
I will take a chance to point out that, IBM's Whatson was supposed to do this for medicine. It turned out, the experts it was supposed to help were not that amused because it was already telling them what they knew.
Expired rice!
You do not have to look far to understand that ChatGPT and her siblings have something fundamentally wrong about them. Google and Microsoft the primary backers of these technologies have limited their use internal usage. And so have a couple of other corporate companies. It reminds me, years back when I was younger, authorities came around advertising free rice for orphan's at the sub-county head quarters. We went to pick them, registered our names and details. The rice tasted horrible, it was expired. In all probability, it was an accountability gimmick by I don't know who.
If you are worth your qualification, you do not need ChatGPT, leave it for the orphan's and gamblers. A man who gives you food they cannot give to their children does not see worth in you.
If you must use ChatGPT, watch your back!
The Dead & Buried Web
The worst flaw of ChatGPT for our community is not that it can lie. Lies can be seen and verified. It is in shadowing of information, something that is impossible to perceive. ChatGPT is possibly the greatest mis-information tool we have created so far, much more effective than social media.
To shadow information is just cover it with something that looks similar but which is actually different. As people vie and fight to have their information occupy the coveted response space to specific questions. Manipulation of information is going to reach unprecedented levels as people adulterate and duplicate information to make it appear factual and prominent. And its going to be invisible. Welcome to super manipulation.
Welcome to the dead and buried web.
A section of the web that despite being open and accessible cannot be found by any search engine thanks to AI.
