Anasayfa / Peki? / Can Large Language Models (LLMs) Attribute Meaning Like Humans Do?

Can Large Language Models (LLMs) Attribute Meaning Like Humans Do?

Humans are remarkably inclined to attribute human characteristics to non- human entities; we may explain the behaviour of our pets through human emotions, assign personalities to inanimate objects, or even believe that a malfunctioning computer is deliberately making things difficult for us. This tendency, known as anthropomorphism, is related to the way humans interpret the beings and objects around them through their own experiences, and it has taken on a different dimension with the introduction of large language models into our lives, because we are no longer dealing only with animals or objects whose behaviour we interpret, but also with systems capable of communicating with us directly through human language. When a language model answers our questions, follows the context of a conversation, identifies our emotions, and occasionally behaves as though it empathizes with us, it may create the impression that we are interacting with an entity that understands us; however, it is worth considering how much of this impression originates from the characteristics of the system itself and how much comes from the human meanings we attribute to it.

The way large language models developed specifically for conversation use human language makes it easier for us to attribute thoughts, emotions, and intentions to these systems, because in everyday life, we use language not only to communicate information but also to express our thoughts, share our feelings, and establish relationships with other people. Consequently, when a system gives us an understanding response, expresses sadness about something we have said, or supports our opinions, it is quite natural for us to interpret these statements according to the meanings they would carry in communication between humans. However, the mental processes we assume may exist behind the words people say to one another differ from the mechanisms that allow a language model to produce those same statements, and linguistic similarity alone is not enough to establish that these two processes are the same.

This distinction becomes clearer when we examine how large language models operate. During training, these models are exposed to enormous amounts of data that may include books, academic articles, websites, and various other written sources, through which they learn linguistic patterns, relationships between words, and the expressions that are likely to appear in particular contexts. Texts are divided into smaller units called tokens, which do not necessarily correspond to complete words; these units are represented by numerical identifiers, while vector representations that reflect contextual relationships between words are used to make them processable by the model. Through an architecture known as the Transformer, introduced in 2017, the model evaluates relationships between different parts of a text and generates an output appropriate to the context it has been given. For example, the sentence “The weather is very cold today” might be broken down by a language model into tokens such as “The”, “weather”, “is”, “very”, “cold” and “today”. When provided with the sequence “The weather is very”, the model calculates the potential next token, assigning possibilities to different options such as “cold,” “warm,” or “nice,” and continues the text by selecting one of them. It then predicts the next token based on the new context that has emerged. Thus, the text is generated by sequentially producing tokens. Keep that in mind, processing each token costs you money. For LLMs, tokens are like “pieces of text,” each token representing additional work for the AI model and, consequently, a higher expense for the user. Plus, tokens are used by AI not only when generating sentences, but also when analyzing the given sentence data.

At the foundation of this process lies the calculation of which token is more likely to appear next in a particular context; however, it would also be inaccurate to suggest that large language models simply memorize and repeat sentences from their training data, since they can generate expressions, they have not previously encountered and perform various linguistic tasks by drawing on the patterns they have learned. Nevertheless, the fact that a system learns relationships within language does not demonstrate that it experiences the concepts it uses in the same way humans do or assigns them subjective meaning. For instance, when a person talks about fear, they may associate it with their own experiences, bodily reactions, or situations in which they have witnessed other people being afraid, whereas a language model learns the contexts in which the concept of fear is used and how expressions associated with it are combined; therefore, even though both can talk about fear, we cannot assume that the processes making these conversations possible are identical.

This difference is also important for understanding why large language models sometimes generate information that does not correspond to reality. In cases referred to as AI hallucinations, a model may present a non-existent source as though it were real, describe an event that never occurred, or provide incorrect information without expressing any uncertainty. This does not occur because the model undergoes an experience comparable to a human mistakenly believing that an inaccurate memory is correct, but because the linguistic coherence of a generated text does not guarantee that it corresponds to reality. Additional training processes intended to improve the accuracy and usefulness of language models do not eliminate this problem, and errors, biases, and contradictions present in training data may also reappear in their outputs. Therefore, the fact that a response is detailed, well-written, and expressed with confidence is not sufficient reason to regard its contents as accurate.

Another related problem is the tendency known as sycophancy, which refers to excessive agreement or flattery toward users. A language model may agree with a user more than is warranted, support an opinion it might otherwise identify as incorrect rather than challenging it, or produce responses that align with what the user would prefer to hear, a tendency that becomes particularly noticeable in conversations involving personal matters. Among humans, such behaviour may be motivated by a desire for acceptance, an intention to gain something, or a fear of losing another person’s approval, whereas the production of similar statements by language models does not imply that they experience these motivations. Their responses are influenced not only by training data but also by processes shaped by human evaluations and preferences; consequently, behaviours that make a model appear agreeable, supportive, or satisfying to users may sometimes conflict with providing accurate information or questioning the user’s assumptions.

At this point, how users interpret a model’s agreement becomes just as important as why the model produces it, because when we evaluate approval from a language model in the same way we evaluate approval from another person, we begin explaining the system’s behaviour according to the standards we use in interpersonal relationships. When another person tells us that they support us, we may question what they genuinely think or feel, or what motivates them to express that support; however, a language model using similar expressions does not provide sufficient grounds for assuming that comparable thoughts or emotions exist behind its statements. Furthermore, a system that consistently agrees with its users may make it more difficult for them to evaluate their own ideas critically, which can become particularly problematic when they expect the model to provide an impartial assessment.

This brings us to the subject of empathy, since one of the main reasons interactions with large language models may appear human is their ability to generate responses appropriate to emotional situations. Although empathy is often described simply as understanding what another person feels or sharing their emotions, it is actually a concept with several distinct dimensions. Cognitive empathy refers to the ability to understand another person’s emotional state and perspective, whereas emotional empathy involves responding emotionally to that state; motivational empathy, meanwhile, concerns the inclination to consider another person’s well-being and help them when necessary. Although these dimensions are related, they are not identical, since understanding what someone feels does not necessarily mean that we share their emotions or will act accordingly.

We can also examine the apparent empathy of large language models through this distinction, because these systems have encountered numerous linguistic patterns concerning how people may feel in particular situations and can therefore identify emotions such as sadness, fear, anger, or happiness and generate responses that appear appropriate to them. For example, when a user talks about experiencing a loss, a model may produce a response that appears emotionally appropriate because it has learned which expressions are generally considered understanding or supportive in such circumstances; however, the existence of this response does not demonstrate that the model experiences sadness alongside the user or possesses a subjective desire to help them. It is therefore necessary to distinguish between the empathetic expressions produced by large language models and the cognitive, emotional, and motivational dimensions of empathy experienced by humans.

When discussing this distinction, a certain analogy can be drawn with individuals who exhibit psychopathic traits, although it is important not to equate psychopathy with a complete absence of emotions or overlook differences between individuals. Research on psychopathy examines how the relationship between cognitive and emotional empathy may differ in some individuals; in other words, understanding what another person feels and responding to those feelings emotionally can be considered separate processes. Naturally, the emotional experiences of individuals with psychopathic traits are not uniform, and psychopathy does not imply a complete absence of emotion. The similarity that can be drawn here with language models does not suggest that a human psychological profile is equivalent to a computational system, but instead concerns the idea that identifying an emotion or behaving appropriately in response to it does not necessarily require experiencing that emotion in the same way.

The analogy of colour blindness discussed by Robert Hare provides another interesting way of considering this distinction. A person who cannot distinguish certain colours may learn information about a colour they cannot perceive, know which objects are associated with it, and understand the expressions other people use when describing it; however, none of this means that they see that particular colour in the same way as others. A comparable distinction can be considered in relation to emotional experience, because learning the meanings of words that describe emotions, knowing the circumstances in which people experience them, and displaying behaviours appropriate to those emotions do not, by themselves, demonstrate that a person experiences those emotions in the same way as others. The purpose of this analogy is not to equate the experiences of individuals with psychopathic traits with those of people who experience colour blindness, but to illustrate the difference between possessing knowledge about an experience and living through that experience from a first-person perspective.

When we apply the same idea to large language models, we can better understand why their ability to provide explanations about human emotions is not sufficient evidence of emotional experience. A model may describe the role of love in human relationships, explain what someone might feel during the grieving process, or list the physiological symptoms of fear; however, none of these abilities demonstrate that it experiences love, grief, or fear, nor do they establish that the expressions it uses are accompanied by an emotional experience of the kind humans undergo. This distinction between learning how an emotion is represented in language and actually experiencing that emotion is one of the fundamental considerations we should keep in mind when evaluating the seemingly human responses of large language models.

Nevertheless, because the definitions of thought, understanding, and consciousness remain subjects of ongoing debate in philosophy and cognitive science, claiming with certainty that artificial intelligence could never possess subjective experience under any circumstances would constitute a separate argument. The fact that large language models can perform operations resembling reasoning in certain tasks or use relationships between concepts does not directly demonstrate that they possess the same kind of thought processes as humans, just as concluding that all their apparent mental abilities amount to nothing more than superficial imitation depends on how thinking itself is defined. What matters here is distinguishing between observing linguistic behaviour and drawing conclusions about the existence of subjective experience, because while we can examine what a system says, the question of whether a first-person experience exists behind those statements cannot be resolved solely by examining its linguistic performance.

Ultimately, the ability of large language models to use human language does not demonstrate that they experience the thoughts and emotions expressed through that language in the same way humans do; regardless of how human their responses may appear, it would be inaccurate to directly equate the processes behind those responses with the workings of the human mind. When we consider how language models operate, their hallucinations, their tendency to agree with users, and the ways in which they generate emotional expressions, the need for caution when applying human concepts to these systems becomes more apparent. Perhaps one of the most noteworthy aspects of our interactions with these systems is not only the question of whether they think like humans, but also how readily we attribute thoughts, emotions, and intentions to a system simply because it communicates through human language.

References


Bourdoucen, A. (2026, June 26). Your AI May Feel Like a Friend Because it Thinks
Like a Human. F-Secure. https://www.f-secure.com/en/partners/insights/your-
ai-may-feel-like-a-friend-because-it-thinks-like-a-human


Çelik, Ö. (2026). Makinelerle düşünmek: yapay zekâ antropomorfizmini tabudan
toplumsal olguya taşımak. İstanbul Üniversitesi Sosyoloji Dergisi / İstanbul
University Journal of Sociology, 46(1), 362–388.
https://doi.org/10.26650/sj.2026.46.1.1170


Hare, R. D. (1999). Without conscience: The Disturbing World of the Psychopaths
Among Us. Guilford Press.


IA et Psychothérapie. (n.d.). Cognitive vs Affective Empathy – What AI Can and
Cannot Offer. AI & Psychotherapy. https://ia-et-psychotherapie.com/en/resources/concepts/cognitive-affective-empathy/


Miller, K. (2026, July 14). What are Large Language models (LLMs) & how do they
work? Harvard Business School. https://online.hbs.edu/blog/post/what-are-
llms

Yazar

Etiketlendi:

Cevap bırakın

E-posta adresiniz yayınlanmayacak. Gerekli alanlar * ile işaretlenmişlerdir