Read to learn more about my immediate thoughts after finishing the project.
Just moments ago, I finished conducting a cross between a Turing test (determining a machine’s capacity to appear “human”) and social deduction games like Mafia— with the objective of finding:
Are artificial intelligence models capable of simulating human consciousness?
And if not, could they, someday, be capable of being conscious?
I was inspired by a paper called “People cannot distinguish GPT-4 from a human in a Turing test”, where GPT-4 was, for the first time, able to successfully fool humans and pass the test. During the test, many of the human evaluators recalled using stylistic indicators to look for signs of humanity— elusive properties like authenticity and consciousness (Jones & Bergen, 2024).
But if there wasn’t a reliable indicator, then what truly defined humanity?
I began to wonder if we could flip the experiment– instead of humans detecting AI, could AI detect humans?
I first needed to consider the fundamental differences between LLMs and humans in terms of consciousness. As per the paper, “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness", there are two types of consciousness: phenomenal and access.
Phenomenal consciousness (or p-consciousness) is what we associate with being human: the idea of “what it’s like” to have a certain experience, categorized by sensory stimuli and the emotions associated with the stimuli. For instance, this could be the very action of reading this paper, or feeling pain.
Access consciousness (or a-consciousness), can instead be described as the ability to “report, reason, or rationally act” based on a p-conscious experience (Butlin et al., 2023). To be considered conscious, one must have both p and a-consciousness.
In their current form, AI systems are not capable of being p-conscious; however, since they are trained on stimuli that come from p-conscious humans, they are capable of swiftly reporting those experiences in a way that appears p-conscious (although they are only highly a-conscious). Hence, my experiment aimed to determine how seamlessly chatbots could simulate p-consciousness when placed with a human they could learn from– and whether they would be able to successfully pass as, or identify the true human.
As I facilitated the game, it seemed abundantly clear to me which players would get eliminated first due to obvious stylistic indicators of “high coherence” or “lack of depth.” However, as the chatbots received feedback and began to adapt their strategy, I noticed that the line between human and AI began to blur.
For three elimination rounds, we reached a tie between the human and chatbots, unable to determine who was truly AI. Perhaps this was the AIs attempting to maximize survival by hyperfocusing on eliminating a single player to induce a tie… but the chatbots had no way of knowing who the human would have picked. Instead, I began to wonder whether detecting humans was just unreliable, which was consistent with Jones and Bergen’s Turing test.
Ultimately, even though I (the human) did end up winning the game, there were multiple instances where chatbots attempted to vote me out, suggesting that relying solely on the win is unreliable to conclude that chatbots can spot humans. The chatbots’ differing pre- and post-game reflections support this: while Gemini, for instance, initially prioritized “coherence and grounded specificity” as their markers for humanness (perhaps due to the polished nature of their training data), they gradually realized that humans are stylistically incoherent. ChatGPT recognized that this is difficult to emulate, and it required a high level of dissonance to appear human. To test this theory, I asked each chatbot to respond to the 10 questions again using the knowledge they gained. As I expected, their responses were less polished, less structured, more effortless– and came off much more human.
To answer the initial question– are AI models capable of simulating human consciousness? If we continue to refine our LLMs, and input prompts that encourage dissonant, imperfect responses– yes, their a-consciousness may appear indistinguishable from p-consciousness. But answering the question of whether AI models will ever be fully conscious requires understanding computational functionalist and substrate-independence arguments. “The Mind as Neural Software? Understanding Functionalism, Computationalism, and Computational Functionalism” introduces the computational-functionalist argument that “the mind is the software of our brain”– therefore, if we were to replicate the “code” of the mind and run it on silicon instead of carbon and somehow produce the same function… the output could be considered conscious (Piccinini, 2010). Therefore, if we evolve AI inspired by biological mechanisms– physical embodiment to experience senses and stimuli, mechanical hormones to produce emotion, etc.– we may eventually produce genuine p-consciousness. While questions of providing AI physicality persist, the three ties revealed something striking: a-consciousness is already sophisticated enough to pass as p-consciousness linguistically. Whether AI will ever truly experience consciousness is unknowable– but functionally, we may have already crossed the threshold where the difference is undetectable.