Fluid Concepts
Human concepts are not fixed categories with rigid boundaries. They are living structures that reshape themselves continuously in response to new encounters. The concept of 'last letter' shifts when applied to 'iijjkk' — it might mean the last character, the last group, or the last unique letter, depending on how the perceiver chooses to parse the string. The choice is not arbitrary; it is guided by a sense of elegance, of structural fit, of what makes the analogy feel right rather than forced. That felt quality of analogical rightness is precisely what Hofstadter spent his career trying to understand and precisely what he argues current AI systems do not possess.
In the AI Story
The distinction that matters is between activation and reshaping. Activation is the retrieval of a pre-existing representation and its application to a new situation; the representation remains fixed, only its deployment is novel. Reshaping is the modification of the representation itself in response to the demands of the situation; the representation changes, the conceptual space expands, new thoughts become possible that had not been possible before.
Large language models operate through activation. Their representations — the high-dimensional vectors encoding semantic relationships — are determined during training and remain fixed during inference. A prompt activates these vectors in novel combinations, producing outputs that can be combinatorially new. But the vectors themselves do not reshape. The conceptual space is frozen at training time. This is the structural reason why the AI-era engine produces combinatorial novelty but cannot produce structural novelty — the creation of new concepts that expand the space of possible thought beyond what the pre-existing elements could generate through any combination.
The practical consequence of this architectural difference is that the collaboration between human and machine, however productive in the moment, does not produce lasting cognitive change in the machine. A human who works with Claude for six months develops new concepts, new intuitions, new ways of parsing problems. The machine that works with the human for six months is, at the level of its representations, exactly where it started. The conversation enriches the human's conceptual repertoire. It enriches the machine's not at all.
This asymmetry is invisible in any single interaction. A conversation with Claude feels dynamic — feels like a process of mutual exploration, of ideas evolving, of understanding deepening. But the feeling is produced by the human's development, not the machine's. The human's concepts are reshaping in response to the machine's outputs. The machine's representations are activating in response to the human's prompts. The dynamism is real, but it is one-sided. Only one participant is actually changing.
Origin
The concept was developed through Hofstadter's long research program at the Fluid Analogies Research Group, beginning in the 1980s at the University of Michigan and continuing at Indiana University. The 1995 book Fluid Concepts and Creative Analogies collected the group's central findings, including detailed accounts of Copycat and related programs designed to model the fluid character of human thought in narrow microdomains.
Key Ideas
Living representations. Human concepts change under pressure, reshaping their boundaries and internal structure in response to novel encounters.
Activation vs reshaping. The pivotal distinction: machines activate fixed representations; humans reshape representations through the act of thinking.
Felt rightness. The perceiver has a bodily, pre-verbal sense of when a conceptual mapping fits and when it forces — a sense AI systems lack.
Frozen training. LLM conceptual space is determined at training time and does not evolve through conversation.
Asymmetric development. In human-AI collaboration, only the human's concepts actually reshape; the machine's remain static.
Appears in the Orange Pill Cycle
Further reading
- Douglas Hofstadter and the Fluid Analogies Research Group, Fluid Concepts and Creative Analogies (Basic Books, 1995)
- Melanie Mitchell, Analogy-Making as Perception: A Computer Model (MIT Press, 1993)
- Douglas Hofstadter, 'The Copycat Project: An Experiment in Nondeterminism and Creative Analogies' (AI Memo 755, 1984)