Category: artificial-intelligence
Type: puzzle
Associated thinker(s): Stevan Harnad
Original publication or approximate period: Harnad, “The Symbol Grounding Problem” (Physica D, 1990)
Difficulty: Intermediate
Themes: meaning, symbols, perception
Related experiments: Chinese Room; Blockhead; Beetle in a Box
Harnad gives two versions of a Chinese–Chinese dictionary exercise. In the “difficult” version, someone who already knows another language tries to learn Chinese using only Chinese definitions. In the more demanding version, the learner is treated as learning Chinese as a first language from the dictionary alone. Each definition leads to more symbols, without a connection to the world.
Harnad uses the second version as an analogy for a purely symbolic model of cognition, whose symbols are manipulated according to their forms. It is not a claim about every computer or AI system.
How can the symbols of a formal system have meaning that does not depend on someone interpreting them?
Harnad connects the problem to Searle’s Chinese Room, then presents the dictionary exercise as his own analogy. He sketches one candidate account: ground some elementary symbols in nonsymbolic sensory and categorical representations, then let higher-order symbols inherit grounding through composition. He presents this as a proposal, not a demonstrated solution, and says it does not guarantee subjective meaning.
A contemporary question is whether text-trained language models acquire meaning from linguistic patterns alone. Harnad’s 1990 paper does not assess current language models. He also notes possible applications to robotics and machine vision.
Harnad’s account is one proposed way to address the problem. Taddeo and Floridi (2005) argue that the strategies they review, including Harnad’s hybrid proposal, fail their “zero-semantic-commitment” test. This is one published critique, not a consensus or a final verdict.
Harnad describes a dictionary go-round; he does not claim dictionaries are literally circular. He argues that adding more symbolically codified material does not by itself ground interpretation. His proposal does not prove that machines cannot understand or settle what is required for meaning. Whether a particular system has grounded representations needs a separate argument.
As a teaching prompt, imagine a system that receives camera images or image-linked language input. Such inputs add perceptual data; whether they ground the system’s symbols is a further question. These are not historical variants attributed to Harnad.
The relevant full-text sections of Harnad’s 1990 paper were consulted in its arXiv-hosted version: §§1.2, 2.1–2.3, 3.2–3.3, 4–5, and related footnotes. The publisher record was checked for journal metadata. Relevant sections of Taddeo and Floridi’s 2005 review were read; their critique is attributed to them and is not presented as consensus. The Chinese/Chinese dictionary go-round is Harnad’s own example (§2.2). The arXiv record is a 1999 deposit of the 1990 article, not its publication date. This is source checking, not expert review.
Research check: 2026-11-01. Contemporary applications and camera examples above are teaching questions, not findings about present-day AI systems.