Large language models' (LLMs) stellar performance has shocked the world. ChatGPT's reactions to people's prompts and questions have been amazingly spot-on. Since then, many questions have arisen concerning whether a LLM could be conscious. If not now, could it happen in the near future? Do we even want LLMs to be conscious? What guardrails should we implement to prevent them from harming humankind?
Among them, the foremost key one is, How would we really know if it is conscious?
This is perhaps the most challenging question humankind has asked themselves for thousands of years. Humans are conscious because we talk to each other, sharing our experiences, thoughts, and feelings. We have the language to describe our own self-awareness and reflect on what's on our minds.
Consciousness needs to be told but can't be directly observed. Furthermore, the core of consciousness is an instance of experience each individual has at a given moment, which is coherent, multimodal, and irreducible. In other words, it is real-time, not sharable, and not reproducible. All these features and properties make it extremely hard to tell if an animal is conscious and what level of consciousness it may have because they cannot speak to tell.
Conversely, LLMs can speak perfectly as humans. Their language responses are fluent, logical, and intelligent, and they can even perform step-to-step reasoning today. Now, we come to the other end of the spectrum—given the language it is speaking, how would we know whether an LLM is conscious or not?
As summarized in my article on the concept and theories of consciousness, consciousness has three typical components:
Experience of the present
Continuity of experience when we are awake
Awareness of the self, including the core and autobiographical self (e.g., social, spiritual, and ego).
Among them, experience is the most essential feature of consciousness. If an LLM is conscious, it should be able to describe its experience. Current LLMs cannot tell their experience because they do not have experience. The models are trained first with a large corpus of text and then fine-tuned with curated questions and answers. Once the deep neural network learns through gradient descendants, the parameters are fixed, and the model is released as software for customers to use. In essence, chatGPT reacts to the prompts and questions. It does not interact, hence, experience, with the customers or the environment—such as sensing, moving, or learning—in real-time as humans do.
However, each LLM has a conceptual system (word embeddings) with hundreds or thousands of dimensions, along which all the words are encoded in a meaningful way (please see my recent article on Medium for detailed explanations). If the conceptual system is connected with sensors and motors, with real-time sensory and motor information also encoded, can experience emerge from the system? As LLMs are moving toward multimodality, including integrating with robot movement, this question will likely be addressed not far in the future.
We humans have our own conceptual systems encoded in our brains. In his classic book Metaphors We Live By, the renowned cognitive linguist George Lakoff thoroughly explained that our conceptual system, hence, the human thought process, is "largely metaphorical." More importantly, our metaphors signify our immediate understanding and are grounded in our experience.
"The essence of metaphor is understanding and experiencing one kind of thing in terms of another....*The concept is metaphorically structured, the activity is metaphorically structured, and consequently, the language is metaphorically structured."
The examples he gave in the book are illuminating, and I am sure many of us have not noticed or realized before. For instance, since we humans walk with heads up and facing forward, we use orientational metaphors all the time. Here are some examples:
Metaphor 1: Happy is up:
She is feeling up
His spirits rose
I am in high spirits
Thinking about my mother always gives me a lift.
Metaphor 2: Sad is down:
I am feeling down
I am depressed.
He's really low these days
My spirits sank
Moreover, our experiences with physical objects, including our bodies, provide the basis for various ontological metaphors, manifesting how we understand events, activities, emotions, ideas, etc., as entities and substances. For example, we refer to inflation as an entity from our experience:
Inflation is lowering our standard of living
We need to combat inflation.
Inflation is backing us into a corner
Inflation makes me sick
When we think of an abstract or intangible thing, we relate it to our experience in terms of similar features, properties, or dimensions. Because one concept may only partially overlap with an experience, we use multiple metaphors to form a coherent web of narratives. For example, the mind can be a machine, a brittle object, or a container, depending on which property or context we focus on at a given moment:
My mind broke down. (machine)
I lose my mind. (object)
I feel my mind is full. (container)
My mind is rusty. (machine)
He was easily crushed. (object)
Stop, I can't hold anymore. (container)
Unlike the LLMs' high-dimensional word embeddings from large amounts of text, our conceptual systems mostly come from our interactions with the surrounding world and people. In other words, our conceptual systems are coherent with our own experiences, as Lakoff argues in his book:
"What constitutes a basic domain of experience? Each such domain is a structured whole within our experience that is conceptualized as what we have called an experiential gestalt. Such gestalts are experientially basic because they characterize structured wholes within recurrent human experiences. They represent coherent organizations of our experiences in terms of natural dimensions (parts, stages, causes, etc.). Domains of experience that are organized as gestalts in terms of such natural dimensions seem to us to be natural kinds of experience..., which are products of human nature."
A typical example of an interactional metaphor is “love.” We primarily understand love metaphorically in concepts of our other experiences, including journeys, war, health, or madness:
Look, we’re at a crossroads; how far we’ve come through.
Where are we? Are we stuck?
It has been a bumpy road, but I am glad we are done with it.
Our relationship is healthy and in good shape
He is crazy about her.
She is besieged and has to fend them off.
Moreover, the essential function of metaphors is understanding. Lakoff stated in his book, "The general principles involve whole systems of concepts rather than individual words or individual concepts. We have found that such principles are often metaphorical in nature and involve understanding one kind of experience in terms of another kind of experience."
All in all, metaphors are the product of our consciousness grounded in our own experience. They truly reflect how we think and understand, and how our underlying conceptual system is organized and integrated with our experience.
Today, when we seek explanations in chatGPT, its responses are well presented with concepts and facts from textbooks, dictionaries, or online articles. It does not use metaphors by itself, not to mention to relate to its own experience or memory. Tomorrow, when we see a robot or LLM start to use metaphors, we can be pretty certain it has consciousness.
Understandably, a machine’s experience can be quite different from that of humans. For example, their vision may not be front-facing but could cover both front and back. Therefore, their spatial and visual experience might be very different, and so will their metaphors. The key point is the use of metaphors, which indicates the existence of experience. Conversely, a machine’s consciousness might not be at the same level as humans, depending on what kind of experience it might have.
In conclusion, Lakoff’s philosophy on how metaphors reflect our ways of thinking and understanding based on our own experience is fascinating. He makes us look at our language not just as a sequence of words for communication but as a source of evidence for how our mind grasps the truth from the surrounding environments and what the underlying conceptual system is like. Moreover, now we can examine LLMs from a new perspective—can they create their own metaphors in the future?
If you are interested in exploring deeper into the consciousness and architecture of LLMs, below are the friend links to my previous two articles: