Large language models such as ChatGPT have rapidly moved from research laboratories into everyday life. They now assist doctors, academics, journalists, and students with writing, analysis, and explanation. Their ability to produce fluent, context-sensitive language often gives the impression of genuine understanding. Yet a fundamental question remains unresolved. Do these systems understand meaning, or are they simply producing convincing imitations of human language?
A new philosophical analysis by Professor Emma Borg from the Institute of Philosophy at the School of Advanced Study, University of London, directly addresses this question. Published in the peer reviewed journal Inquiry, the article is titled “LLMs, Turing tests and Chinese rooms: the prospects for meaning in large language models”. Borg examines whether current artificial intelligence systems can use language meaningfully, and what kind of meaning might be involved if they do.
A debate shaped by Turing and Searle
Modern discussions of artificial intelligence remain deeply influenced by two classic thought experiments. The first is Alan Turing’s Imitation Test, often referred to as the Turing test. Proposed in 1950, it suggested that if a machine can carry on a text-based conversation indistinguishable from that of a human, it should be treated as intelligent.
The second is John Searle’s Chinese Room argument. Introduced in 1980, it challenged the idea that behavior alone is sufficient for understanding. Searle argued that a system could manipulate symbols according to rules and still lack any grasp of meaning. In his thought experiment, a person inside a sealed room follows instructions which allow them to respond appropriately to Chinese text but without understanding the language.
Large language models appear to bring these two positions into direct conflict. Systems like ChatGPT can easily pass many informal versions of the Turing test. Yet their internal operations resemble the symbol manipulation Searle described. Borg’s article asks whether it is possible to move beyond this stalemate by examining what sceptics and supporters of AI implicitly assume about meaning itself.
Determining whether AI systems ‘mean what they say’ is crucial for deciding the role such systems can play in our lives. Currently, however, it seems that neither the complete cynics nor the total optimists are getting things right.
-Emma Borg
How large language models work
At a technical level, large language models are trained to predict the most likely next word in a sequence. They learn from enormous datasets containing billions of words drawn from books, articles and online sources. During training, words are represented as mathematical vectors based on statistical patterns of co-occurrence.
Modern systems employ transformer architectures that incorporate self-attention mechanisms. These enable models to dynamically adjust word representations based on context. As a result, the same word can assume different roles depending on the surrounding text. This enables large language models to exhibit impressive fluency and flexibility.
Importantly, these systems are not explicitly trained to understand meaning. They are optimized to minimize prediction error. This has led many critics to argue that any appearance of understanding is illusory. According to this view, language models are advanced pattern-matching systems rather than genuine semantic agents.
Rethinking meaning beyond the world
Borg explores several responses to the sceptical challenge. One option is to reject the idea that meaning must be grounded in direct reference to the external world. Internalist theories of semantics argue that meaning arises from relationships between words and concepts rather than from direct contact with objects.
Distributional semantics, often summarized as knowing a word by the company it keeps, fits naturally with how large language models operate. On this view, meaning emerges from patterns of usage across language. Inferential role semantics offers a related approach, defining meaning in terms of the inferences a system can make.
From this perspective, large language models may already encode substantial semantic information. Their ability to draw inferences, answer questions and generalize across domains suggests sensitivity to conceptual structure rather than mere surface patterns. Borg notes that whether this counts as genuine representation remains an open philosophical question.
The idea of derived meaning
The most distinctive contribution of Borg’s article is its defense of derived meaning. Derived meaning refers to semantic content that depends on human intentions and social practices rather than on intrinsic understanding within a system.
Computers have long been understood as possessing derived intentionality. A calculator’s output means four because humans interpret it that way. Large language models take this a step further. They employ the same words and grammatical structures that humans use to communicate with one another.
Borg argues that when a language model produces a well-formed English sentence, that sentence inherits meaning from the broader linguistic community. The meaning does not originate within the machine, but it is nonetheless real. In this sense, large language model outputs can be genuinely meaningful even if the system itself does not understand them in the way humans do.
Why this mirrors human language use
An important implication of Borg’s analysis is that much of human language use is also deferential. Speakers routinely use terms they only partially understand. They rely on experts and communal norms to fix meaning. Social externalism has long recognized this feature of language.
From this perspective, the gap between human and machine language use may be smaller than often assumed. Both rely on inherited linguistic structures. Both can produce meaningful sentences without a full understanding of all implications. The crucial difference lies in agency, intention and responsibility rather than in semantic content itself.
This insight challenges the idea that meaning must always originate in individual cognition. It suggests that linguistic meaning is a fundamentally social phenomenon, one that artificial systems can participate in under certain conditions.
Meaning without assertion or understanding
Despite defending the meaningfulness of language model outputs, Borg draws an important boundary. She argues that large language models do not make assertions. Assertion involves a commitment to truth. It carries epistemic and ethical responsibilities that current AI systems cannot bear.
Language models generate text based on probability rather than belief. They do not aim at truth in the way human speakers typically do. This explains why fabricated information can appear alongside accurate statements without any internal conflict.
For this reason, Borg urges caution in how AI-generated content is presented and used. Output should be clearly marked as machine-generated. Human oversight remains essential, particularly in scientific and medical contexts where accuracy is critical.
The limits of original intentionality
The final section of the article addresses original intentionality. This refers to the capacity to form mental states that are intrinsically about the world. Many philosophers associate original intentionality with consciousness, agency, and moral status.
Borg argues that current large language models lack original intentionality. They lack embodied interaction, persistent goals, and subjective experience. While future systems may integrate language models with sensory and motor capacities, it is unclear whether this would be sufficient to generate genuine understanding.
Reference
Borg, E. (2025). LLMs, Turing tests and Chinese rooms: the prospects for meaning in large language models. Inquiry. https://doi.org/10.1080/0020174X.2024.2446241
