Stevan Harnad
Université du Québec à Montréal
The T2/T3 Distinction and Robotic Grounding. There is a crucial distinction between T2 (the verbal-only version of the Turing Test — Turing-Indistiguishable from that of any normal human being) and T3 (the robotic version of the Turing Test, with the Turing-Indistiguishable verbal capacity grounded in sensorimotor capacity Turing-Indistinguishable from any normal human being). LLMs are T2 systems. T3 capacity — not just sensory capacity, but, critically, the motor aspect of sensorimotor interactionâis necessary for grounding. The “experience arrow” (x: H â W) seems a pale abstraction of what real grounding requires: the capacity to do things in the world with the referents of content-words, not just receive inputs from them and name them.
Direct vs. Indirect Grounding: Not Parasitism but Cheating. LLMs are indeed “epistemically parasitic.” Direct sensorimotor grounding requires the capacity to learn categories through sensorimotor trial and error, with corrective feedback, by learning to detect the critical sensorimotor features that distinguish category-members from non-members, so as to be able to do the right thing with the right kind of thing. Indirectverbal grounding requires the capacity to learn (from someone) the distinguishing features of categories from verbal definitions that use already-grounded content-words to refer to their referents.
Humans learning from indirect grounding aren’t “parasitic”âthey’re building on their own direct grounding foundation. Indirect grounding is dependent on prior direct sensorimotor grounding. LLMs cannot do indirect grounding at all. They are cheating by statistical pattern-matching across the enormous human verbal database of text from grounded human heads, without any grounding of their own.
Category Learning and Minimal Grounding Sets. The research on category learning and categorical perception (CP) concerns how organisms learn to detect category-distinguishing features through direct trial and error learning with corrective feedback (+/- reinforcement) from doing the right or wrong thing with members of the category. This is related to research on dictionariesâ âminimal grounding setsâ (âMinSetsâ). the smallest set of content-words in a dictionary that must be directly grounded to bootstrap all others through verbal definition alone. Jerrold Katzâs Katzâs âEffability Thesisâ and graph-theoretic analysis of dictionaries suggest that this MinSet can be surprisingly small, as few as 1000 content-words, among those that children learn earliest.
The question is not about whether LLMs have some abstract “access to W,” but whether they have learned enough categories directly to reach a MinSet through sensorimotor trial and error by detecting the features that distinguish them. (Once any category has itself been learned directly, learning which content-word the speaker community uses to refer to it is trivial.) Individual human learners who have approached or reached a MinSet for their language by direct grounding can then go on to ground the rest of the referring words of their language though indirect verbal grounding provided by verbal sources (such as teachers, dictionaries, text books â or LLMs) that can already name the distinguishing features of the referents of the rest of the words in the language and convey them to the learner through subject/predicate propositions (definitions and descriptions). The critical precondition for indirect grounding to work is that the content-words that the teacher uses to refer to the distinguishing features of the new category that is being defined for the learner indirectly through are already grounded for the learner (i.e., they are already grounded in the learnerâs MinSet or can be looked up by consulting a dictionary or a textbook or an LLM or a human teacher): They do not, however, need to be grounded for the source, whether dictionary, textbook, LLM, or human teacher. They need only be accessible to the learner from the source. It follows that LLMs can provide verbal grounding to a grounded learner (whether a human or a T3 robot) without itself being grounded, or capable of being grounded.
The Lexicon of a Language and Propositional Recombination. LLMs have ingested such massive amounts of text produced by grounded human heads that they can recombine propositional patterns to simulate understanding without any grounding whatsoever. The scale of training data allows statistical pattern-matching to mimic the outputs of grounded understanding, which LLMs do not, and cannot acquire, not even one MinSetâs worth, because, not being T3 robots, they do not have the sensorimotor means to acquire it. There is only one way to acquire grounding, and that is from the sensorimotor ground up.
The role of language’s combinatorial and expressive powerâgenerating infinitely many propositions from finite meansâis central here. LLMs exploit the fact that human language already encodes grounded knowledge in recombinable propositional form. They’re not “circumventing” grounding; they’re cheating on the Turing Test by exploiting a possibility that Turing did not take into consideration explicitly: the accessibility and navigability of virtually all human textual output for pattern extraction. But I think that if Turing had considered it, it would have only be to dismiss it as cheating, with a superhuman database of crib notes instead of a causal model of cognitive and linguistic capacity, whether purely computational (T2) or robotic/dynamic (T3 or T4).
Floridi et al.âs mathematical framework points out that LLMs operate on content C rather than world W: LLMs learn from text about the world, not from the world itself.
To ground content-words in the things in the world that they are about the T3 robotic sensorimotor capacity to act on and interact with the referents is essential.
Floridi et al. correctly note that LLMs lack genuine grounding. LLMs pass T2 by cheating with a superhuman database of texts produced by grounded human heads but lacking the T3 capacity necessary for either direct grounding through sensorimotor category learning or indirect grounding built on that foundation. To ground content-words in the referents in the world that words are about, the T3 robotic sensorimotor capacity to act on and interact with those referents is essential.
Floridi, L., Jia, Y., & Tohmé, F. (2025). A Categorical Analysis of Large Language Models and Why LLMs Circumvent the Symbol Grounding Problem. arXiv preprint arXiv:2512.09117.
Harnad, S. (2025). Language writ large: LLMs, ChatGPT, meaning, and understanding. Frontiers in Artificial Intelligence, 7, 1490698.
PĂ©rez-Gay JuĂĄrez, Fernanda, et al. (2019) Category learning can alter perception and its neural correlates. PLoS One 14.12 (2019): e0226000.
VincentâLamarre, P., MassĂ©, A. B., Lopes, M., Lord, M., Marcotte, O., & Harnad, S. (2016). The latent structure of dictionaries. Topics in cognitive science, 8(3), 625-659.