Large Language Models (LLMs) are profoundly transforming our lives, yet the understanding of their essence remains highly contentious. A recent perspective article published in Communications Psychology argues that current research heavily relies on metaphors—comparing LLMs to brains, markets, or complex systems. While this strategy can highlight certain characteristics, it also fosters cognitive fragmentation and a recursive loop of anthropomorphism.
The authors, from a research team in China, categorize metaphors into three levels: mechanical, behavioral, and interactive, pointing out that each has its explanatory boundaries. For instance, the physical metaphor of "emergent abilities" cannot predict LLMs' impact on social equity; while psychological tests reveal cognitive biases, they often blur the line between machine capabilities and human projections.
To break this deadlock, the authors propose a framework of "machine empiricism." Drawing on the experientialist philosophy of cognitive linguists Lakoff and Johnson, they argue that LLMs' "understanding" is not an imitation of human cognition, but a unique form of knowledge constructed from massive textual environments within their Transformer architecture. This shift requires us to move from asking "How human-like are LLMs?" to "What is the inherent logic of LLMs?"
The article calls for interdisciplinary researchers to carefully consider the applicability of metaphors and avoid unconsciously inheriting inappropriate assumptions. Only by distinguishing "what LLMs are" from "what we think they resemble" can we move toward a more comprehensive understanding.