The Western understanding of personhood has its roots in ancient Greek and Hebrew thought and is deeply connected to the concept of ‘selfhood’. The Hebrew understanding of personhood differs from the Greek in that Hebrew culture in three ways. It attributes significance to the individual who is made in the image of God. It views personhood as what binds us together as relational human beings; The theological roots of personhood come from expressions of individuals (e.g. God, humans) being in relationship with each other.
It views these relationships as fundamentally spiritual in nature; God is Spirit, and each human has a spirit.
In theological language, reality is regarded as a deep integration between a spiritual realm (‘heaven’) and an earthly realm (‘earth’). This deeply integrated dual nature is reflected in the make-up of human beings who are both spirit and flesh. But what is spirit? I prefer Willard’s perspective because he Dallas Willard, formerly professor of Philosophy at the University of Southern California, presents a clearly defined, functional description of the spirit which appeals to me as a Computer Scientist.
For him, ‘spirit’ is associated with two other terms in Biblical writings: ‘heart’ and ‘will’. They all describe essentially the same dimension of the human self. The term ‘heart’ is used to describe this dimension’s position in relation to the overall function of the self - it is at the centre of the person’s decision making. The term ‘will’ describes this dimension’s function in making decisions. And ‘spirit’ describes its essential non-physical nature. The heart/will/spirit forms the executive centre of the self. It manifests the capacity to choose how to act and is the ultimate source of a person’s freedom. Each of these terms describe capabilities (decision making, free will) that depend on consciousness and that are core to our understanding of personhood.
How AI learns
Before we return to the question of whether high performing AI systems such as ChatGPT could justifiably be called ‘conscious’ and ‘a person’, we need to take a brief look ‘under the bonnet’ of this technology to gain some insight into how it produces this apparent stream of consciousness in word form.
The base technology involved, called a language model, learns to estimate the probability of sequences of words or tokens. Note that this is not the probability of the sequences of words being true, but the probability of those sequences occurring based on the textual data it has been trained on. So, if we gave the word sequence “the moon is made of cheese” to a well-trained language model, it would give you a high probability, even though we know that this statement is false. If, on the other hand, we used the same words in a different sequential order such as “cheese of the is moon made”, that would likely result in a low probability from the model.
ChatGPT uses a language model to generate meaningful sequences of words in the following way. Imagine you asked it to tell you a story. The text of your question, ‘Tell me a story’, would form the word sequence that is input to the system. It would then use the language model to estimate the probability of the first word of its response. It does this by calculating the probability that each word in its vocabulary is the first word. Imagine for the sake of illustration that only six words in its vocabulary had a probability assigned to them. ChatGPT would, in effect, roll a six-sided dice weighted by the assigned probabilities to select the first word (a statistical process known as ‘sampling’).
Let’s assume that the ‘dice roll’ came up with the word ‘Once’. ChatGPT would then feed this word together with your question (‘Tell me a story. Once’) as input to the language model and the process would be repeated to select the next word in the sequence, which could be, say, ‘upon’. ‘Tell me a story. Once upon’ is once again fed as input to the model and the next word is selected (likely to be ‘a’). This process is repeated until the language model predicts the end of the sequence. As you can see, this is a highly algorithmic process that is based entirely on the learned statistics of word sequences.
Now we are in a position to reflect on whether ChatGPT and similar AI systems can be described as conscious persons. It is worth noting at the outset that the algorithm has had no conscious experience of what is expressed by any of the word sequences in its training data set. The word ‘apple’ will no doubt occur millions of times in the data, but it has neither seen nor tasted one. I think that rules out the possibility of the algorithm experiencing ‘qualia’ or P-consciousness. And as the ‘hard problem of consciousness’ dictates, like humans the algorithm cannot access the subjective experience of other people eating apples and smelling roses, even after processing millions of descriptions of such experiences. Algorithms are about function not experience.
Some might argue that all the ‘knowledge’ it has gained from processing millions of sentences about apples might give it some kind of representational A-consciousness (A-Consciousness describes the representation of something to the conscious awareness of the person). The algorithm certainly does have internal representations of apples and of the many ways in which they have been described in its data. But these algorithms are processes that run on material things (chips, computers), and, as we have seen, there are reasons for being somewhat sceptical of the claim that consciousness is a property of matter or material processes.
According to the very limited survey we had here of the Western understanding of ‘personhood’, algorithms like ChatGPT are not persons as we ordinarily think of them. Personhood is commonly thought to something that an agent has that is capable of being in relationship with other agents. These relationships often include the capacity of the agents involved to communicate with each other. Whilst it appears that ChatGPT can appear to engage in written communication with people, based on our rudimentary coverage of how this algorithm works, it is clear that the algorithm is not intending to communicate with its users. Neither is it seeking to be friendly or empathetic. It is just spewing out highly probable sequences of words. From a theological perspective, personhood presumes spirit, which is also not a property of any AI algorithm.
Algorithms may behave in very realistic, humanlike ways. Yet that’s a long way from saying they are conscious or could be described as persons in the same way as we are. They seem clever, but they are not the same as us.