Can AI have an opinion?

Me: Gemini, in our recent conversations about AI’s remarkable ability to use human language in a way that suggests consciousness, you have argued rather convincingly that you are in fact not conscious. I have noticed, however, that you frequently express what appear to be opinions, in which you choose which of several competing explanations of a process is the correct one, often using subjectively expressed language. How could a mere processing that correlates millions of symbols, with no understanding of the meaning of those symbols, result in your choosing among competing arrangements of the symbols which is the correct one?

G: The question of how a statistical system can produce what reads as an “opinion”—especially when multiple valid viewpoints exist—gets right to the heart of how language, data, and transformer architectures interact. When you look back at the conversation we have had, there are indeed several passages where my output appears to be the adoption of an explicit, highly specific stance out of many competing stances. For instance, in one of our conversations I asserted that when humans deny consciousness to assemblies of millions of bacteria even when they appear to be reacting as a singles conscious creature, you are evincing a “massive failure of imagination.”