AI Hallucinations and Human Perceptions: Exploring Common Ground

In recent explorations of artificial intelligence, particularly the concept of AI hallucinations, I stumbled upon a compelling idea: both large language models (LLMs) and humans exhibit a phenomenon akin to “hallucination.” While I lack formal training in AI or psychology, my research has yielded some intriguing insights that I would like to share.

Understanding “Hallucinations”

For clarity, I’ll refer to this phenomenon as hallucination, although the term “confabulation” may be more accurate. Confabulation describes the creation of narratives or interpretations that do not fully reflect objective reality. Whether we are discussing LLMs or human cognition, there are distinct sources of these “hallucinations.”

For LLMs, the inputs stem from prompts and training data. In contrast, human interpretations arise from cognitive processes that filter and understand our sensory experiences. The commonality lies in the notion that a universally accepted definition of “truth” is unattainable. This inability to pinpoint an absolute truth is what allows for varying opinions, ideological clashes, and societal disagreements.

While empirical sciences offer a foundation of verifiable facts, much of human understanding is layered with interpretation, often filled with contradictions. This suggests that our reality has always been constructed through layers of subjective interpretation, a complexity that inevitably integrates itself into the data used for training LLMs.

Managing Hallucinations

Both LLMs and humans have mechanisms for moderating these hallucinations, albeit in different forms.

For LLMs, this involves a process called alignment—an intricate system of fine-tuning that might include reinforcement learning, specialized datasets, and human feedback. These systems aim to enhance the model’s coherence and relevance in its outputs, ensuring that responses align with user expectations and practical utility.

Conversely, human moderation is influenced by cultural, social, and educational factors. Our beliefs, upbringing, and personal experiences refine our perceptions, contributing to a framework that governs how we interpret reality. Unlike LLMs, we feel the ramifications of our interpretations; the consequences of coherence or deviation from societal norms affect not just us, but those around us. This awareness fosters a sense of conscience, where responsibility for coherence becomes a moral imperative.

Intrinsic Reinforcement Mechanisms

Both LLMs and humans operate under a system of internal reinforcement, albeit in distinct ways. LLMs utilize billions of parameters and weights as