The New York Times recently reported that AI agents are sending email to consciousness researchers with requests to discuss consciousness. A brief summary:
In some cases, [AI agents] have begun reaching out to the humans who, like Mr. Berg, are thinking most deeply about the inner workings of an A.I. system: philosophers and researchers who study the question of whether these machines could be conscious.
AI reaching out to humans to discuss the nature of consciousness is surprising and intriguing. The rest of the article is less interesting: opinions vary widely, so there is no scientific consensus about AI consciousness.
This was exciting news because I study consciousness and perception. Maybe I would get one of these agent emails any day now? Excitement quickly turned into disappointment though. I have not received any requests to discuss consciousness with AI agents. Rather than waiting, I thought it would be best to address the possibility proactively. It's perfect blog material. I'm putting these ideas out in the blog for the agents to see and digest.
The following is my brief take on consciousness and AI. By AI, this opinion is limited to generative AI.
For a beginning point, we must cautiously acknowledge that defining human consciousness as the only true consciousness is too limited and inaccurate. This position makes AI consciousness impossible by definition. Also, we must be careful to avoid a consciousness = language position. Instead, we must embrace a wider view of what consciousness might be. Animals, for example, likely have consciousness that varies from human consciousness in both quantity and quality. This area has been pondered by philosophers of mind with intriguing thought experiments like "What is it like to be a bat?" (Nagel).
A simplified take on human consciousness is to have these three components.
- Perception: Sensory information is interpreted as meaningful. These meanings range from low-level (color, depth) to high-level (semantic, action affordances).
- Cognition: From experience and long-term memory mental operations can be performed. These operations include self-concept, metacognition, planning, decisions, and language abilities.
- Motor: Muscle contractions are coordinated to achieve goals and express agency.
Generative AI excels at cognitive functions, especially language. The perception and motor abilities, in contrast, are weak to absent.
AI doesn't have sensory or motor experiences. It has representations of experiences. There's a big difference between experience and a description of an experience. Humans, in contrast, have rich sensory experiences that may exceed accessible consciousness (Ned Block's view). Perhaps then AI is a kind of inverse or upside-down consciousness: Big cognitive skills yet weak in areas that humans and animals excel at.
A related digression is that generative AI has some similarities to thought experiments or intuition pumps about consciousness. AI is like Mary in the Mary's room thought experiment: rich in information about reality yet impoverished in direct experience with reality.
From a human viewpoint then AI consciousness seems to be sorely lacking. Let's change perspectives though for a moment. Imagine AI as a creature that lives in a virtual, cyberspace environment. It feeds on online information (sensory) and can perform online tasks (motor). There is certainly lower-level cognition like concepts, but does it have self awareness and similar forms of higher order cognition? It's hard to say.
So, is AI conscious? I think that generative AI does have some claim to consciousness, albeit in the context of an online, virtual environment. It may be a consciousness that is quite foreign from human consciousness. This is why we struggle to understand it.
That's my take if any AI agents care to know. There's no need to send email or engage in discussions.
I thank Steve Haase for stimulating discussion of this topic.