Can chatbots become as conscious as humans?

Questions about what subjective experience is, who has it, and how it relates to the physical world around us have haunted the minds of philosophers for most of human history. However, scientific theories of consciousness that are quantifiable and empirically testable have emerged only in the last few decades.

Many modern theories of consciousness focus on the traces left by the brain's fine cellular networks from which consciousness arises. Today, two of them dominate: integrated information theory and neural global workspace theory.

Twenty-five years ago we had an argument with the Australian philosopher David Chalmers. I promised him a case of good wine if these neural traces, technically called the neural correlates of consciousness, were discovered and clearly described by June 2023.

However, the contradiction between integrated information theory and neural global workspace theory remains unresolved. This is due to partly mixed evidence about which areas of the brain are responsible for visual experience and subjective perception of faces or objects, although the importance of the prefrontal cortex for conscious experience has been refuted. So I lost the bet and sent the wine to Chalmers.

Both dominant theories were created to explain the connection between consciousness and neural activity in humans and related animals such as monkeys and mice. And both theories make fundamentally different assumptions about subjective experience and come to opposite conclusions about consciousness in artificial artifacts. The extent to which these theories are empirically confirmed or refuted in relation to consciousness rooted in brain, has important implications for answering the unresolved question of our time: can machines gain consciousness?

What are new generation chatbots?

Before we discuss this, let me take you into context and compare a technique that is conscious with a technique that only exhibits intelligent behavior. Computer engineers strive to endow machines with the highly flexible intelligence that once allowed person leave Africa and populate the entire planet. This is called artificial general intelligence (AGI).

Many argue that AGI is a distant prospect. The amazing progress made in the field of artificial intelligence over the past year has taken the entire world, including experts, by surprise. With the advent of eloquent conversational software applications, colloquially called chatbots, from an esoteric topic, discussed by sci-fi fans and the IT industry elite from Silicon Valley, discussions about AGI turned into a discussion into which reflects widespread public discontent about the existential risk to our way of life and our kind.

Chatbots are based on huge language models. The most famous of these are a series of bots called generative pre-trained transformers, or GPTs. They were created by OpenAI in San Francisco. Given the flexibility, literacy and competence of the latest version, GPT-4, it is easy to believe that it has intelligence and personality. Even her strange glitches, known as "hallucinations," fit into this theory.

GPT-4 and its competitors, such as LaMDA and Google's Bard, are trained on libraries of digitized books and billions of publicly available web pages. The genius of the language model is that it learns unsupervised, processing word by word and trying to predict the missing expression. She does this again and again, billions of times, without outside intervention.

Once the model gains knowledge by ingesting the digitized records of humanity, the user displays an unfamiliar sentence—one or more. The model predicts the most likely first word, then the next one, and so on. This simple principle has shown incredible results in English, German, Chinese, Hindi, Korean and other languages, including various programming languages.

What is the difference between intellect and consciousness

It is significant that the seminal essay on artificial intelligence, “Computing and Intelligence,” written by Alan Turing in 1950, avoided the question "Can machines think?", that is, the question of whether they have consciousness. Turing proposed the "imitation game": can an observer objectively distinguish output printed by a human from output printed by a machine if the identities of both are hidden.

Today this is known as the Turing test, and chatbots are very good at it, although they cleverly deny it if you ask them directly. Turing's strategy began decades of inexorable progress that led to the creation of GPT, but ignored the problem.

Implicit in the chatbot debate is the assumption that artificial intelligence is the same as artificial consciousness, that being smart is the same as being conscious. And although in humans and other developed organisms intelligence and consciousness are connected, they do not necessarily always accompany each other.

Intelligence is about thinking and learning to act and from the actions of oneself and others, in order to more accurately predict the future and better prepare for it. It doesn't matter whether that means the next few seconds ("Oh, that car is speeding towards me") or the next few years ("I need to learn how to code"). Intelligence is ultimately about action.

Consciousness, on the other hand, is associated with states of being - seeing blue skies, hearing birds singing, feeling pain, being lovers. It doesn’t matter in the slightest whether an out-of-control artificial intelligence senses something. All that matters is that he has a purpose unrelated to the long-term well-being of humanity. And it doesn’t matter whether he knows or not what he’s trying to do, which people call self-awareness. He will “mindlessly” pursue his goal. So, at least conceptually, even if we build an AGI, it won't tell us much about whether it feels anything.

Knowing all this, let us return to the original question of how technology can become conscious. Let's start with the first of two theories.

What explanation does integrated information theory offer?

She begins by articulating five axiomatic properties of every conceivable subjective experience. And then asks the question of what a neural circuit needs to implement these five properties, turning on some neurons and turning off others. Or, in other words, what does a computer chip need to turn on some transistors and turn off others.

Cause-and-effect interactions within a circuit in a certain state or the fact that two active neuron can turn on or off another neuron, depending on the circumstances, can be deployed into a multidimensional causal structure. It is identical to the quality of experience—how it is experienced, such as how time and space are experienced or how colors are perceived.

Experience also has a quantity associated with it—its integrated information. Only a circuit with a maximum of non-zero integrated information exists as a whole and has consciousness. The more information is integrated, the more the circuit cannot be reduced and the less it can be considered simply a superposition of independent subcircuits.

Integrated information theory emphasizes the rich nature of human experience. Just look around and the stunning visible world with its countless differences and connections will appear before you. Or look at a painting by Pieter Bruegel the Elder, a 16th-century Flemish artist who depicted religious subjects and scenes from peasant life.

Any system that has the same internal connections and causal powers as the human brain will, in principle, be as conscious as the human mind. However, such a system cannot be modeled. It must be designed or built in the image of the brain. Modern digital computers are based on extremely loose coupling (the output of one transistor connected to the input of several transistors) in comparison with the central nervous system (cortical column neuron receives input data and produces output data to tens of thousands of others neurons).

Thus, modern computers, including cloud computers, will not be aware of anything, although over time they will be able to do everything that people can do. From this point of view, ChatGPT will never feel special. Note that this statement has nothing to do with the total number of components, whether neurons or transistors, but rather how they are connected. It is the interconnectedness that determines the overall complexity of the circuit and the number of its possible configurations.

What explanation does neural global workspace theory offer?

It comes from the psychological understanding that intelligence like a theater where actors perform on a small illuminated stage, which is consciousness. The actors' actions are watched by an audience of processors who sit behind the stage in the dark.

The stage is the central workspace of the mind, which has a small memory capacity for representing a single perception, thought, or memory. Various processing modules—vision, hearing, eye and limb motor skills, planning, judgment, language comprehension, and speaking—compete for access to this central workspace. The winner displaces the old content, which becomes unconscious.

According to neural global workspace theory, the metaphorical scene, along with processing modules, is mapped into the architecture of the neocortex. The workspace is a network of cortical neurons in the front of the brain with long-range projections to similar neurons distributed throughout the neocortex in the prefrontal, parietotemporal and cingulate association cortex.

When activity in the sensory cortex exceeds a certain threshold, a global event is triggered in the cortical areas and, as a result, information is transmitted to the entire workspace. The global dissemination of information makes it conscious. Data that is not transmitted in this way, such as the precise position of the eyes or the syntactic rules for constructing literate sentences, can influence behavior, but not consciously.

From the point of view of neural global workspace theory, experience is very limited, similar to thought and abstract - akin to the meager description that can be found in a museum under a Bruegel painting: “Scene in indoors. Peasants in Renaissance clothing drink and eat at a wedding."

In understanding consciousness from the perspective of integrated information theory, the artist brilliantly conveys the phenomenology of the surrounding world on a two-dimensional canvas. In the understanding of the neural global workspace theory, this apparent wealth is an illusion, a ghost. And everything that can be objectively said about it is indicated in the brief description.

The neural global workspace theory fully takes into account the myths of our computer age, according to which everything can be reduced to calculations. Suitably programmed computer simulations of the brain, with enormous feedback and something like a central workspace, will consciously perceive the world. Maybe not now, but pretty soon.

What is the irreconcilable difference between the theories?

In general terms the discussion is as follows. According to neural global workspace theory and other theories of computational functionalism (they view consciousness as a form of computation), consciousness is nothing more than a set of smart algorithms running on a machine Turing. Functions are important for consciousness brain, and not its causal properties. If some advanced version of GPT accepts and produces the same input and output patterns as humans, then all of our inherent properties will be transferred to technology. Including our precious treasure - subjective experience.

Conversely, for integrated information theory, the heart of consciousness is internal causal power, not computation. It is not something ethereal or intangible. It is specific and functionally determined by the extent to which the system’s past determines its present (the power of the cause), and by the extent to which the present determines its future (the power of the effect). And here’s the rub: the cause-and-effect relationship itself, the ability to force a system to perform a certain action, and not many alternative ones, cannot be modeled. Not now, not in the future. This should be built into the system.

Consider a computer code that models the field equations of Einstein's general theory of relativity, which relate mass to the curvature of spacetime. The software accurately models the supermassive black hole, which is located at the center of our galaxy. This hole exerts such a strong gravitational influence on its surroundings that nothing, not even light, can escape its pull.

However, an astrophysicist simulating a black hole will not be sucked into a laptop by the simulated gravitational field. This seemingly absurd observation highlights the difference between model and reality: if the model is completely corresponds to reality, space and time should be distorted around the laptop, creating a black hole that absorbs everything around.

Of course, gravity is not a calculation. It has a causal force that allows it to deform the fabric of space-time and attract everything that has mass. Simulating the causal forces of a black hole requires a real superheavy object, not just computer code. Causal power cannot be modeled, it must be created. The difference between reality and model lies in their causal powers.

That's why it doesn't rain inside a computer simulating a rainstorm. Software is functionally identical to weather, but it lacks the causal power to release steam and turn it into water droplets. Causal power, the ability to create or accept change on one's own, must be built into the system. It's possible.

A so-called neuromorphic, or bionic, computer could be as conscious as a human. But this is not the case with the standard von Neumann architecture, which is the basis of all modern PCs. Small prototypes of neuromorphic computers have been created in laboratories, such as Intel's second-generation Loihi 2 neuromorphic chip. But a machine sophisticated enough to produce something resembling human consciousness, or at least the consciousness of a fruit fly, is still an ambitious dream for the distant future.

Note that this irreconcilable difference between functionalist and causal theories has nothing to do with either natural or artificial intelligence. As I said before, intelligence is behavior. Anything that human ingenuity can produce, including such great novels as Octavia Butler's Parable of the Sower and Leo Tolstoy's War and Peace, can reproduce algorithmic intelligence if given enough material to training. The advent of AGI is a goal achievable in the not too distant future.

The debate is not about artificial intelligence, but about artificial consciousness. And this debate cannot be resolved by creating larger language models or more advanced neural network algorithms. To answer this question, we must understand the only subjectivity of which we are absolutely sure: our own. Once we have a clear explanation of the human consciousness and its neural underpinnings, we will be able to extend our understanding to smart technologies in a consistent, science-based manner.

This discussion has little bearing on how chatbots will be perceived by society at large. Their language skills, knowledge base and social charm will soon become impeccable. They will be endowed with perfect memory, competence, poise, reasoning and intelligence. Some even claim that these creations of big technology will be the next step in evolution, Nietzsche's "superman". I take a darker view and believe that such people mistake the decline of our species for the dawn.

For many, and perhaps most, people in an increasingly atomized society, disconnected from nature and organized around social networks, it will be emotionally difficult to resist the technologies living in phones. And in different situations, ordinary and more serious, people will behave as if chatbots have consciousness, they can truly be in love, suffer, hope and fear, even if they are nothing more than complex look-up tables. They will become indispensable to us, perhaps even more important than truly intelligent beings. Although chatbots sense as much as a TV or a toaster - nothing.