You ask:
What would it take for an AI to have beliefs?
It's an open question if it's possible for AI to have human-like beliefs and experience, but on a broad reading of belief as a spontaneous intentional internal representation of the self or the world, it is arguable they already do. But the vagueness of terms possess a challenge, and there are actually several related readings of this question:
- Are AI systems capable of having belief? What exactly is belief?
- Are AI systems capable of having representations? What exactly is a representation?
- Are AI systems capable of replicating human belief? What does it mean to replicate human belief?
This sort of question is fraught with confusion, so immediately, the SEP offers us resources to suss out a response:
From the first SEP article:
Anglophone philosophers of mind generally use the term “belief” to refer to the attitude we have, roughly, whenever we take something to be the case or regard it as true. To believe something, in this sense, needn’t involve actively reflecting on it.
Computer systems are capable of having internal representations even about the external world and judging something to be true or false, and they needn't actively be aware or reflect on that process, so on the shallowest reading, yes, AI software can have belief. But this does not make it human belief, or even dog belief. (This, of course, is a generalization of the use of belief, whose necessary and sufficient conditions for many include it to be a property of sentient, biological beings unconditionally if one subscribes to prototypical definitions.) Belief and representation comes in many forms, and there is not valid reason to exclude non-biological electromechanical systems where biological systems are concerned as general principle.
Let us also acknowledge that mere belief is not knowledge if one accepts that judgment often plays some role in transforming belief to knowledge. Even with these resources, thinkers are going to have a strong set of opinions based on their personal definitions of 'belief', 'representation', 'thought', etc. Clearly, AI systems are not capable of belief and knowledge in the same way as people, period, because the richness of the complexity of the human brain in terms of using and understanding symbols is unmatched by even the most powerful computers such as Deep Blue, Watson, and so on. This is why modern computer science recognizes a distinction between the concepts of AI and AGI.
But systems are built that mimic human belief, desire, and intention, conveniently called Belief-Desire-Intention (BDI) software, and therefore, while software belief does not match the belief of humans, or arguably Great Apes generally, it might be on par with simpler organisms who are sometimes held to have cognition by ethologists. That is to say, a complicated software system that implements a formal representation of belief can produce robotic behavior that mimics the abilities of simple animals. Boston Dynamic's Spot is capable of generating a real-time representation of the world, generating a plan, and prioritizing and implementing action.
Combining robotic BDI, integrating natural language processing (NLP), and packaging it in a humanoid form not only can be done, but creates the effect called uncanny valley. These systems certainly do not fool anyone when time is taken to examine them, but they do reflect enough human-like properties to give pause, and certainly to form the foundation of philosophical discourse, which is arguably a form of experimental philosophy (SEP). Some thinkers faithfully imbue neurons with an elan vital or attribute quantum woo to consciousness, in a defense that somehow neurons are capable of leading to the emergent properties of consciousness in a way that transistors or not. But, personally, I've never seen a compelling argument, and functionalism (SEP), as flawed as it is, creates compelling lines of argument.
So, current systems arguably have beliefs, desires, and intentions if operationalization is the standard by which abstractions are deemed to exist. But they are shallow facsimiles of human forms of belief, desire, and intention in the same way even the most sophisticated human prosthetic hand resembles the biological human hand. And you'll observe strong knee-jerk reactions to the question in the same spirit of the uncanny valley. No! Computer can never think! Yes, computers are already self-aware! Some people react quite emotionally in the face of asserting that robots and computers can possess human-like qualities in the same way people will literally get angry if confronted about the their religious beliefs.
Ultimately, it would require a book-length work to provide a thorough response to differentiate how modern AI systems fall short of AGI, and thus human-level belief, desire, and intention. In fact, there are a lot of highly intelligent thinkers devoting energy to this issue now, and we see lines of debate that mimic the Searle/Dennett discourse surrounding minds, language, and consciousness. So it continues to be the job of philosophers, to wade through these conceptual questions.And continues to be the job of scientists and computer scientists to build bigger and more powerful formal systems of belief, desire, intention, and action. Who knows what quantum computers and other future technologies will bring.
See also:
