Are AIs actually intelligent?
If you have been on any social media platform in the last few months, I’m sure you have heard about ChatGPT, Bing’s search engine, Bard, and all these new technologies that can probably write better than you and me. Their creators call them Artificial Intelligence (AI), although some tend to question whether we should name it that way.
I agree that AI is what we should call these new advancements, considering the terminology used in this field for over 70 years. When we talk about AI, we are referring to a computer program, an artificial one, capable of autonomously solving tasks or performing functions that are typically associated with human behavior or reasoning. In other words, something we could refer to as intelligent behavior. That’s why when a program is used to aid decision-making, object identification, or generate information from a given context, it is referred to as Artificial Intelligence: because it is something that we humans usually do using our own human intelligence, but we have automated and delegated it to a machine. And I cannot think of many tasks that are more human than using properly any human language.
You can refer to the first two chapters of the famous book by Russell and Norvig on Artificial Intelligence [1] to understand what AI is. If you are interested in this topic, it is an essential read to provide you with the necessary context.
But what I want to discuss in this article is something that many people are starting to question, given the incredible results achieved with these language models in a very short time. Are these computer programs truly intelligent? Behind the impressive paragraphs generated by GPT4 or any other Large Language Model (LLM), is there real intelligence?
It’s not a new question, but its impact on society has been dramatically updated in just a few months because this field has progressed rapidly. It’s a complex topic that requires establishing many concepts, but if all the technological advancements of the moment impress you as much as they impress me, I encourage you to join me on this journey from the human mind to the thinking core of a computer.
I want to clarify that I am currently not involved in the development of these types of Artificial Intelligence models. However, I am a telecommunications engineer with a master’s degree in artificial intelligence, and I am passionate about this topic. So, although this is an opinion article and not a scientific one, everything I know and believe comes from extensive study in this field. It is still important to note that everything is open to discussion. And now, let’s get into matter…
What is intelligence?
The first question we need to address is, what is intelligence? And I must admit, in this text, we won’t be able to fully resolve this eternal question. So, disappointingly, everything we argue from this point forward will be based on uncertain ground — the definition of intelligence. Without a solid foundation, we cannot make a strong and definitive argument about whether something is intelligent or not. However, we can form an opinion on the matter.
Interestingly, we all have some intuition about the meaning of the word “intelligence.” We believe we could all identify the most intelligent person in our class and those who lagged behind a bit. We all agree that when a baby learns to imitate their parents and starts operating a smartphone screen with their hand, it demonstrates a certain level of intelligence.
However, if we stop and think about it, I’m sure very few of us have ever received a precise definition of that same word. But somehow, we have all come to the conclusion that intelligence is a quality related to the mind, to our capacity to think, understand, and create.
This is a possible definition of intelligence:
“A set of cognitive and behavioral skills that enables efficient adaptation to the physical and social environment. It includes the ability to solve problems, plan, think abstractly, comprehend complex ideas, and learn from experience. Intelligence is not identified with specific knowledge or specific skills, but rather it refers to a general cognitive ability that encompasses specific capacities.”
I’m not exactly sure who authored that definition as I found it published in several places, but what’s interesting is that it emphasizes skills and abilities. If we look at other formal definitions that have emerged over time [2], we indeed see a similar focus on the capacity to do something:
- Lewis Terman (1921): Capacity for abstract thinking.
- David Wechsler (1944): Capacity to act with a specific purpose, think rationally, and relate effectively to the environment.
- Jean Piaget (1952): Capacity for adaptation to the environment.
- Sternberg and Salter (1982): Capacity to adapt behavior to the achievement of a goal. Includes the abilities to benefit from experience, solve problems, and reason effectively.
- Papalia and Wendkos-Olds (1996): Active interaction between inherited abilities and environmental experiences, resulting in the individual’s capacity to acquire, remember, and use knowledge, understand concrete and abstract concepts, comprehend relationships between objects, facts, and ideas, and apply and utilize all of this with the specific purpose of solving everyday life problems.
Nowadays, we know that there are even tests or assessments to measure intelligence, although many experts still don’t agree on what exactly those tests actually measure or how to confirm it.
Is there intelligence beyond us?
When we talk about intelligence, we usually think exclusively of our own minds, the minds of Homo sapiens. However, if you have a special fondness for your dog or work in a veterinary clinic, you might have already considered the following: Is there intelligence beyond human beings? Are other animals intelligent? The answer seems to be yes. Dogs, primates, dolphins, elephants, even octopuses — there are many animals that you may be familiar with because they are considered particularly intelligent in one way or another. In other words, we can confirm that humans, despite considering themselves the pinnacle of intelligence on Earth, have already acknowledged intelligence at other levels beyond Homo sapiens.
As Ignacio Crespo graciously and accurately comments in his book “A Jungle of Synapse” [3], human beings tend to think that their minds and their intelligence are special and unique because we have no evidence of anything else in the universe that has developed such complex cognitive abilities. Therefore, our brain, as the receptacle of that marvelous intelligence, must also be special. This has given rise to a major current in neuroscience that seeks to discover what makes our brains so special.
But what if, despite the incredible nature of the human brain, our intelligence is nothing more than an emergent effect resulting from the interconnection capacity of millions of neurons, producing differences in electrical potential in a complex biological network? What if we replace that network composed of living cells with silicon-based structures that also exchange electrical potentials and can give rise to emergent effects? Would it still be intelligent?
What I mean to say is that although we are still far from being able to fully model and understand the overwhelming functioning of the human brain, we can imagine that what we call intelligence arises from a highly complex structure of smaller entities, which we call neurons, that enable us to perform certain tasks or abilities collectively. Just as neurons are small parts of something larger that functions very well, deep learning models like the one used in GPT follow the same idea. This doesn’t mean that we can automatically expect a deep learning model to behave like a brain, but perhaps it can approach or emulate certain aspects of a brain, yielding potentially better results as the complexity of these models increases.
Nobel laureate Kazuo Ishiguro wrote brilliantly about this topic in “Klara and the Sun” [4], which is narrated from the perspective of a robot, an artificial intelligence in a possible future. Highly recommended and highly relevant to the ideas discussed in this article, but I won’t spoil the novel for you. Just wanted to share it for your own delightment!
Some experts in Artificial Intelligence, such as Professor José Ángel Olivas from whom I have learned a lot, advocate for an approach that emulates human cognitive processes, while others argue for the emulation of the biological structure of the brain to recreate an intelligence governed by a computer program. Perhaps neither of these approaches alone is sufficient; we don’t fully understand either one — the cognitive and the biological.
However, today we do recognize that similar to the brain, we can already observe emergent effects, skills, or abilities that arise in the training of artificial neural networks used in these artificial neural network (ANN) models. Emergent effects are one of the foundations of Systems Engineering, and our earliest reference to human recognition of emergent effects can be found somewhat tangentially in Aristotle’s Metaphysics [5], where he discusses what “have several parts and where the totality of them is not like a heap, but the whole is something beyond the parts.” In plain words: when you combine lettuce, tomato, onion, oil, and vinegar, you don’t just have those separate ingredients, but you have a salad — something more, something different from the parts that make it up. Similarly, the interconnectedness of millions of neurons in our brain gives rise to intelligence, and the combination of millions of encoded numerical values in an artificial language model allows ChatGPT to make your grocery list.
It seems that little by little, we are approaching an answer to the conundrum of intelligence.
Narrowing the search: computer intelligence is real?
Terrence J. Sejnowski, from the Salk Institute for Biological Studies, published in March 2023 a very interesting article on this topic in the Neural Computation journal from MIT [6]. To begin with, he highlights that different specialists do not agree on whether the latest conversational programs are intelligent or if they truly understand what they say. However, he presents some important considerations: these programs can demonstrate abilities such as creativity, abstraction, synthesis, and adaptation through their language. They still lack others, such as initiative or long-term continuous learning. It is worth remembering at this point that intelligence is often defined as a set of abilities or a general ability that encompasses specific ones. It’s important to note that LLMs like GPT4 not only use our language but also use it generally correctly based on what is being asked or requested. It is, they show some certain ability in the use of language. Until recently, everyone agreed that this kind of response from any being requires an understanding that necessitates intelligence. However, the controversy arises: ChatGPT and their equals are not made of flesh and bone like any of us.
After considering all of this, in my opinion, the most reasonable approach seems to be avoiding absolute judgments. We cannot categorically say whether the new AI systems are or are not intelligent, just as we should not always assume who is the most intelligent in the class because life surprises us. We understand that we cannot claim that a dog is not intelligent simply because it cannot speak, and that’s because we subconsciously identify intelligence in different aspects.
What I believe is that we can say with reasonable confidence that current artificial intelligence models have developed a set of skills that are closely related to certain aspects of human intelligence. They are not human; they do not have a brain, and we have not been able to computationally simulate one. However, they use an equally complex structure of elements that have given rise, in an emergent manner, to intelligent behaviors in language usage. And many other usages, actually.
It is widely accepted today that general intelligence as a singular concept may not exist or may not be particularly relevant. Instead, the mind is believed to be based on different types of intelligence. There are numerous models [7] that describe different kinds of intelligence that complement each other, ranging from Sternberg’s three types of intelligence to Guilford’s 150 types, as well as Gardner’s widely recognized model of eight different intelligences. This could suggest that AI systems may be developed in a way that enhances specific types of intelligence while disregarding others entirely.
Conclusions
There is still a long way to go before machine intelligence becomes similar to human intelligence, and I don’t even think that’s something that makes sense. In the early days of AI, developers were obsessed with making their programs pass the famous Turing Test (which, of course, has been surpassed in conversational abilities for years). However, gradually the trend shifted towards developing AI systems that are more useful and better adapted to complement, facilitate, or sometimes replace human actions, rather than pretending to be something they are not.
In this same vein, it wouldn’t be truly useful for ChatGPT to have the exact conversational level of a human. Imagine if you ask it about the foundational principles of quantum physics and it responds, “Well, that’s super complicated. I’ve read something about it, but honestly, I don’t understand it.” What good is an AI that is so human-like in that regard? I already have my own self with my own limitations for that.
That’s why we need to start thinking that human intelligence is not as unique and special in the universe as we may think. It is composed of a series of truly incredible abilities that can be partially matched or even surpassed by other organisms, sometimes through natural evolution and sometimes artificially created by ourselves.
References
[1] Artificial Intelligence: A Modern Approach. Stuart Russel and Peter Norvig. Pearson.
[2] INTRODUCCIÓN A LA PSICOLOGÍA, TEMA 8 LA INTELIGENCIA. Universidad de Alicante.
[3] Una Selva de Sinapsis. Ignacio Crespo. Paidos Iberica.
[4] Klara and the Sun. Kazuo Ishiguro.
[7] Universal Intelligence: A Definition of Machine Intelligence. Shane Legg and Marcus Hutter.
