UNDERSTANDING IS GRASPING
UNDERSTANDING IS GRASPING is the conceptual metaphor that structures a significant portion of the debate about whether AI systems "really" understand what they process. English speakers grasp ideas, hold concepts, cannot get a grip on a difficult argument, let go of old assumptions, carry knowledge, pick up new skills, turn ideas over in their minds. Every expression maps the cognitive domain of comprehension onto the motor domain of manual manipulation — reaching, grasping, holding, releasing, rotating. The mapping exists because human beings are creatures with hands, for whom the primary mode of engaging with the physical world is to reach out, grasp objects, manipulate them, examine them from different angles. When critics say that a large language model does not truly understand language, they are — usually without awareness — activating the GRASPING metaphor and noting that the system does not physically grasp anything.
In the AI Story
The circularity is precise: the metaphor defines the standard for understanding, the standard is applied to the system, the system fails the standard, and the conclusion is drawn that the system does not understand. The conclusion was contained in the metaphor, not in the evidence. A different metaphorical frame for understanding would produce a different evaluation of identical behavior. Consider an alternative: UNDERSTANDING IS CONNECTING. English speakers also say they see the connection, ideas are linked, an argument is well-constructed, concepts are tied together, a theory holds together. In this frame, understanding is not grasping a single object but perceiving relationships among multiple objects — and connecting patterns is precisely what large language models do, at a scale and speed that no individual human mind can approach.
The point is not that the model understands or does not understand. The point is that the question "Does the model understand?" is not a question about the model. It is a question about the metaphor through which understanding is conceptualized. Different metaphors produce different answers to the same empirical question about the same system. The metaphor is doing the cognitive work, and the work is invisible to the people relying on it. This is the general pattern Lakoff's framework reveals: disputes about what AI systems can or cannot do are often disputes about which conceptual metaphors frame the comparison, with the metaphors operating beneath the level of explicit argument.
The GRASPING frame is not wrong. It captures something real about human comprehension: that understanding involves taking hold of ideas, examining them, manipulating them in relation to other ideas. The capacity to do this is partly constituted by the embodied experience of physical grasping, which provides the image-schematic structure through which conceptual manipulation becomes tractable. But the frame is partial. It does not exhaust what understanding is. A system that connects patterns across vast conceptual territories without grasping any of them in the image-schematic sense may be performing something that deserves to be called understanding, even if the specific grasping-based form of understanding is not what it does. The framework's contribution is not to declare AI understanding real or fake but to show that the declaration depends on which frame is operating, and that the choice of frame is an analytical move rather than a neutral description.
For AI research and policy, the recognition that understanding-debates are frame-dependent has practical consequences. Benchmarks designed to test whether systems "understand" inherit the conceptual metaphors of their designers, which may build in assumptions that predetermine the answers. Evaluation frameworks that presuppose a single standard for understanding may not capture the kinds of cognitive operations AI systems actually perform. The productive path forward, Lakoff's framework suggests, is not to answer whether AI understands within inherited frames but to ask what kinds of cognitive operations AI systems perform and what relationship those operations bear to the various human capacities the word "understanding" has been used to describe.
Origin
The UNDERSTANDING IS GRASPING metaphor is among the most ancient and widespread in Indo-European languages, with cognates and parallels across languages whose speakers share the basic embodied experience of manual manipulation. Its specific application to debates about AI comprehension emerged in the 1960s with early AI systems and has intensified with each subsequent wave of AI development.
Key Ideas
Source domain: manual manipulation. Understanding is mapped onto reaching, grasping, holding, rotating.
Embodied grounding. The metaphor draws on the universal bodily experience of using hands to engage with physical objects.
Frame-dependent conclusions. Whether a system "understands" depends on which metaphorical frame is applied to evaluate it.
Alternative frames exist. UNDERSTANDING IS CONNECTING and other frames generate different evaluations of identical system behavior.
Research implications. Benchmarks and evaluation frameworks inherit the frames of their designers, potentially building in predetermined conclusions.
Appears in the Orange Pill Cycle
Further reading
- George Lakoff and Mark Johnson, Metaphors We Live By (University of Chicago Press, 1980)
- Mark Johnson, The Body in the Mind (University of Chicago Press, 1987)
- John Searle, "Minds, Brains, and Programs," Behavioral and Brain Sciences 3 (1980)