Many people have seen the video of a bear in a forest stumbling across its own reflection in a mirror and destroying the mirror in the mistaken belief that the mirror is a threat to itself. Cats and dogs sometimes respond aggressively to their own reflections in a mirror.
Some animals however show positive responses to the mirror test—dolphins, magpies, elephants and nonhuman apes—suggesting a degree of self-awareness in these animals[1]. Young children are able to pass the mirror test by the age of 18-30 months.
In an LLM [Large Language Models] we face a mirror test of our own, but instead of light it reflects our own written language back at us. [2]
Michael Timothy Bennett (2023)
This quote from Bennett captures the essence of large language models – AI text generators such as ChatGPT and Gemini – that they are essentially mirror tests of ourselves. Not tests of our ability to recognise our own physical self but test of our ability to recognise that we are language. Language is our DNA and when we look at the output from AI we see ourselves reflected back at us.
We may spurn the output AI produces, mock the production process, and in some cases physically attack the computers that host such software. But if we do, are we not in effect like the bear in the forest? Do we not fail to see ourselves reflecting in the linguistic output?
I have argued elsewhere that sentience (consciousness) is in the ‘eye of the beholder’. It is for me to ascribe sentience or not to machines such as ChatGPT. Sentience is not something that a machine ‘has’. Sentience is a first-person phenomenal awareness of language in my mind that is not transferable.
So when I read the output from LLMs, I need to remember that it is me in that language, as it is me in the language right here.
[1]Plotnik, J.M., de Waal, F.B.M., Reiss, D., 2006. Self-recognition in an Asian elephant. PNAS 103, 17053–17057.
[2] Michael Timothy Bennett (2023) On the Computation of Meaning, Language Models and Incomprehensible Horrors. in Hammer, P., Alirezaie, M., & Strannegård, C. (2023). Artificial General Intelligence: 16th International Conference, AGI 2023, Stockholm, Sweden, June 16-19, 2023, Proceedings (1st ed., Vol. 13921). Cham: Springer International Publishing AG.
