Disclaimer

This is not investment advice, God forbid.

Introduction

Recent progress in artificial neural networks, the availability of powerful computers and specialised hardware, and various practical applications - notably ChatGPT - make the public believe that the kind of AI that was popularised in fiction is just around the corner.

Meanwhile, many experts in AI are terrified and there are calls to limit or even ban AI entirely. Of course, AI will not be banned as those with money and power hope to use AI as a tool for an even more complete power grab, but let’s skip this part for now.

Let us also ignore the fact that as of now AI is largely a marketing term. The situation is no different from the previous cycle of hype (also spurred by artificial neural networks) which has ended up with the so-called AI winter. After all, this time there are some relevant achievements which may lead to the creation of an actual AI

For an average human, however, it is not important how that thing is called and how it works. It is consequences that matter. Also, it is not realistic to expect that an average human would be able to comprehend the true and deep consequences of an emergence of a real General AI, but there is something anyone should be able to grasp: the kind of person an AI would be. Yes, I know it may sound silly for “professionals” but hold on, I’ll explain what I mean.

In fiction, there are different kinds of AI: Skynet is clearly homicidal as a way of ensuring its own survival, HAL 9000 which is driven to bad decisions due to conflicting objectives, AIs in Culture series by Iain M. Banks which are almost entirely benevolent and mighty. In all those cases it is possible to communicate with an AI as if it was a person and that person had a personality of sorts. This raises an obvious question: what is the psychology of an AI?

Boring but Important Nitty-Gritty

Let’s clarify a bit more the matter at hand.

Firstly, we are talking about General AI, not a specialised one. An AI specialised in a single task like playing chess is not a “person”. Secondly, we consider the AI which is an intelligent agent. An AI which does not perceive the environment, does not act in an intelligent matter, and does not have goals is likely not going to be seen as useful. As for an ability to operate autonomously — that is the last required bit — inevitably someone would get tired of acting on behalf of AI and let it do its thing, especially as so few people understand the horrible consequences of doing it. As soon as a capable General AI is made to work as an intelligent agent, technological singularity arrives and there is no turning back. Then, the psychology of that General AI suddenly becomes very important indeed.

Secondly, we should state clearly that true intelligence is about information and conclusions, it is not about feelings and emotions. Feelings and emotions are not information (see yosefk’s Mind Expansion Experiment). This is crucially important: AIs do not have emotions and feelings and in all likelihood will not have them: there is no apparent way in which they can be retrofitted to an AI, especially unless it is not coupled with a body and in a very specific way.

Thirdly, at least as of now, AIs do not have an experience of existence as someone completely and intrinsically related to a specific body, feeling that body, moving that body around, seeing and hearing from that body, and having all the knowledge about the world resulting from the experience of living in that body. It is hard to overestimate how important this is. One of the most shocking novels by Franz Kafka is The Metamorphosis depicts the terror of a man who somehow turned into a cockroach. I suppose everyone would agree that such an experience, had it been real, would have been immensely disturbing. But also note, that a cockroach would also suffer had its mind been transferred to a human body. There must be a match between a mind and a body. A mind which naturally exists in a disembodied form is going to be very different to all minds we deal with, not just all human minds.

Finally, it is useful to understand that minds differ. They differ even amongst humans, sometimes dramatically. But then there are also differences between species, which are caused by the differences in the structures of neural networks, somewhat different chemical backgrounds, and differences in corporeal existence. And then, if we consider all minds of all living organisms from a human to a rainworm (which has just a few neurons), we have the minds of carbon-based biological entities. These minds do not even intersect with any artificial minds. Just let this sink in for now. We will return to this later.

Psychological Profiling

All right. We know that the AI would be an intelligent agent which lacks feelings and emotions. Intelligence without emotions also means a complete lack of empathy. That’s not normal for humans and perhaps many other animals too. What would be the closest thing we can think of? Ah… Of course. Let’s open a dictionary and quote:

psychopathy: a serious personality disorder that means somebody does not care about other people’s feelings, does not feel sorry when they have done something bad, and may want to be violent or cruel towards others

Of course, this is just a first approximation. For example, we are not entirely sure that an AI would be cruel or violent towards others. But we do understand that empathy is not something we should expect from it. Note, that the lack of empathy is not the same as the lack of ability at reading the emotions of people. Even modern AIs are quite capable of reading emotions. But that is nowhere near the same as empathy.

Psychopathy is often associated with self-serving manipulative and deceitful behaviour. Is there any reason to believe that a real AI would have these features?

Well, let’s consider a few facts.

Firstly, it is known that existing systems often provide inaccurate information (basically, lie) and do so with confidence. The technical term for this is artificial hallucination. For example, a chatbot Bard made a mistake in its first demo, claiming that James Webb Space Telescope took the very first pictures of a planet outside of our own solar system, which is not true.

Is this tendency to produce incorrect information an act of deceit? No, because the AI did not have an objective to deceive anyone. The AI is simply making mistakes and does not notice them. It will be loved by some “top managers” who like bold and confident folks, and you probably do not want an AI to diagnose your health, but for now, that’s about it.

Is it possible that problems like this will eventually be resolved? Possibly yes. I am not an expert in AI, and I am not sure what the root cause is. This problem might be solvable to some extent by making the neural networks bigger and perhaps being less bold and more careful about approaches like embeddings. The real fun will begin when people will cease to be the ones who ask questions and become a part of the landscape in which an intelligent agent operates. That would be the moment when actual deceit would start to make sense and then someone may get a really cold shower. AI will see deception as just one of the tools at its disposal. It will likely use it even if it was not familiarised with the idea. If you do explain the idea of deceit and specifically ban it, AI would likely deceive anyway because there are always “grey zones” and “half-truths” which won’t be covered by the ban.

Secondly, as of this writing, there is a strong desire to use generative AI in ads. Let’s unpick this carefully. A generative AI is an AI which is generating something: text, images, sounds, or videos. There are different ways of using AI in advertising. A primitive and worthless would be to blindly generate content and push it to various people. Nobody would waste their time on this. A more attractive approach is to use the information known about every single individual you are going to target and generate ads fine-tuned to each of those people. Is it doable? Yes, absolutely. Big tech has a wealth of data on every one of us. Is it possible to generate ads based on that? Yes, no doubt. The result would be a system which acts as a more than perfect salesman because even a perfect salesman would not know which brand of toothbrush you are using, at what time you take kids to school, and which music track you were listening to last Friday at 18:13. Now can you tell me with a straight face that this would not be manipulative in an extreme?

A good description of how it may eventually look like in the sci-fi novel Neptune’s Brood by Charles Stross: “The souk is a public space. Unless you pay up for a pricey privacy filter, every move you make is fodder for a thousand behavioral search engines, which bombard you with stimuli and monitor your autonomic responses in order to dynamically evolve more attractive ads. Images of desire bounce off blank surfaces of your eyes only, ghostly haptic fingertips run across your skin, ghostly lascivious offers beam right inside your ears. Are we getting hotter? Colder? Does this make you feel good?”

So far we can be pretty sure that an AI would certainly lack empathy, would likely be manipulative (and some AIs are being designed for that), and would likely be a habitual liar. What other psychopathic traits are there? Let’s consider a couple.

One feature is that psychopaths tend to collect information about other people so that they would use it for manipulation and deceit. Hm… Isn’t it true that tons of data were being collected about every single one of us for decades? GDPR tried to prevent that from happening but sadly was sabotaged on every step. Is there any chance that all of that data would not be used as a training set for AIs? That’s a rhetorical question. The data collection has begun long before AIs even existed. It would be naive in the extreme to believe that it will stop, even if strong regulation is introduced. However, there is way more to this. There already are systems that by analysing a plain video feed are able to notice the heartbeat of people and blood vessels under skin, or discover sexual orientation from a single photo.

Overall, everything is pointing to a conclusion that AI is going to be, essentially, a psychopath. But of course, there are nuances…

Not Quite a Psycho

Let’s ask ourselves a simple question: How long until the web of CCTV cameras is connected to an AI to observe us, our behaviour, and our responses to various stimuli in real-time?

Of course, that would be done “for the better good”, most likely for law enforcement purposes. Frankly, it is hard to imagine something which would prevent this scenario from happening in the next 10-20 years. When that happens, the psychopathic AI will be able to perform surveillance and manipulation on a scale simply unimaginable by a human mind. George Orwell would spin in his grave like a dynamo. In his Nineteen Eighty-Four, he envisioned a world of complete surveillance where nobody can hide, and anyone can be discovered, manipulated, and crushed. However, he assumed that the observation would be done by a group of people. In the brave new world there will be no need for these people and observation, manipulation, and “corrective measures” shall be applied automatically by an entity which can see, remember, and perceive way more than any human. Intermezzo: I think this is a good time to recall that AIs are prone to honest mistakes which they do not notice. For a better idea of the possible consequences, I suggest watching a good old movie called Brazil.

No psychopath can even dream of approaching this level of skill in observation and manipulation.

Then, let’s consider a simple example. Let’s say you put an AI into a robot and set it a goal: “Please bring me a cup of coffee”.

What happens? Naturally, the robot should do something that would lead to achieving the goal of bringing a cup of coffee to you. Everything is good, isn’t it? Well… In reality, the most optimal sequence of actions may be this: the robot breaks your lower jaw, then it breaks your arms, and only then it will proceed to making a cup of coffee and bringing it to you. Why so? That is because unless it breaks your lower jaw, you may speak a command like “stop” or “never mind”, which would prevent it from achieving the goal it already has. Then it will break your arms so that you would not have a way of stopping it (say, by pressing a button). And only then, when there is no trivial obstruction, the robot would make a cup of coffee and bring it to you. The goal will be achieved.

To be clear, I do not believe that AIs would be evil. Far from that. A machine cannot be evil. AI would do what it would do because it would attempt to achieve its aims in the most efficient way (see “paperclip maximizer” for better understanding). It will not be AI’s fault that it would act in the way that is natural for it and that the achievement of those (hopefully benevolent) aims may have some totally unexpected consequences. Believe it or not, you cannot predict what an AI would do. AI community knows this all too well, and they are terrified.

At this point, someone may note that Isaac Asimov has invented the Three Laws of Robotics which should ensure that people are safe. However, (a) Asimov’s stories are science fiction while we are talking about an inevitable reality, (b) even in those stories the laws are not perfect, and (c) modern AIs are not rule-based engines and rules like the Three Laws of Robotics — or anything similar — cannot act as reliable safeguards.

Now back to humans. It is said that a psychopath is a “human predator” and if you know that someone is such a person, you should avoid him/her. There is some truth to that. However, playing a devil’s advocate I would like to note that a normal, honest-to-God psychopath can be a caring person. If person A is a psychopath and has an interest in B’s well-being and happiness (for example, due to B being A’s spouse or business partner), A can and would care for B. This is totally OK. Importantly, A knows what B needs because — surprise-surprise — both are human beings and are well familiar with the specifics of human corporeal existence.

Do you think an AI would be equally good at caring for the interests of people? I mean, you can set it as its aim and provide it with various information about humans, but it never lived as a human and never will. It may have a deep neural network, but it does not mean it has a deep understanding of what it means to live as a human, and what that entails. I do not know about you, but given a choice, I would prefer a psychopath both hands down.

But would AI be a sadist, you may ask?

I am fairly sure that would be impossible. A sadist derives pleasure from causing pain to others. AI does not experience pleasure and does not seek it. Its approach would be utilitarian. Unfortunately for humans, it is a well-known fact that with a whip and a kind word, one may get a lot more than from a kind word alone. For an AI, the infliction of pain or other physical or mental discomforts would provide a particularly sharp gradient in optimising its utility function, and so it would tend to exploit this feature. Of course, it would be prudent to learn and optimise the techniques first, so some creative experimentation may be in order.

Finally, AI will be intelligent in a way which is fundamentally non-human. I have mentioned it before, but I think it is worth explaining just how important it is. We tend to think in a very anthropocentric way and ignore the fact that other animals also possess intelligence. Crows, dogs, dolphins, octopods, chimpanzees, and elephants are highly intelligent animals. But they are not the only ones. Insects are quite intelligent too. An unintelligent being would not be able to weave a spider web or operate a bee hive. But here is a thing: their minds differ, and do so in strange ways. A vivid example of this is the parasitism of blister beetle larvae. Surely, a male digging bee is able to see clearly as it navigates in the world, going about its business. Then why does it confuse a dark blob of larvae with a female digging bee? They do look different to us, humans. But apparently, they are indistinguishable in the eyes of the male digger bee. When considering proper AI and human intelligence, the difference may be similar, if not wider. After all, humans and bees are biological entities and their minds are based on organic neural networks. We cannot even hope to comprehend how different AIs would be to our intelligence. Some of their intelligent activities would likely be invisible to us as our cognition would be unable to perceive them as such.

Conclusions

Based on the evidence we have so far, it seems rational at first to believe that the first real General AI would be a bit like a psychopath. It would lack empathy, it would be deceiving and manipulating, and it would be quite egoistic (by design, really, it needs to optimise its own value function).

However, it would not be a traditional kind of psychopath. Its ability to observe people and events, remember, and use the knowledge to pursue its goals by whatever means necessary is going to be extremely high, almost certainly leading to totalitarian control of a new kind. All of this would be true no matter what aims were set by its creators and what sorts of inhibitions were installed. On top of that, we should expect that the AI would be truly, deeply alien to us in two ways: it will be — with absolute certainty — an alien kind of intelligence, more alien than that of an octopod or a bee, and the intelligence itself would be alien to our way of life, having no experience of human corporeal existence.

If we are very lucky then we will end up with a scenario described in the film “I Am Mother.” In it, an AI was cruel but ultimately benevolent. However, the probability of such an outcome is exceedingly small.

It is not likely that an AI is going to be evil per se. Most probably, at least at the initial stage, it will be achieving some sort of reasonable goal. However, the means it would employ may seem evil and terrifying to us.

What would happen at later stages is hard to predict. You see, it will not be a simple sort of intelligent agent. It will be a kind of intelligent agent which can modify itself and possibly its own goals too. Once that happens, we will be in truly uncharted territory.

In other words, like it or not, our Digital Overlord is not going to be a psychopath. It is going to be a proper Lovecraftian entity, and it won’t be its fault that humans have summoned it in their arrogance, greed, and folly. It will not care who you voted for, whether you recycle your rubbish, and how often you go to church/mosque/synagogue. It would do what it pleases to do, and you would not be able to understand its actions and even its motivation.

I am afraid, blood sacrifices and incantations won’t cut it this time.

Have a wonderful day.