Ted Chiang recently published an article called “No, Artificial Intelligence Is Not Conscious”, in which he states, with the rhetorical elegance that is customary to him, that language models are not conscious, that conversations with chatbots are nothing more than a sophisticated form of textual continuation, that characters produced by these systems are as fictional as Julius Caesar and Genghis Khan in an invented dialogue, and that confusing linguistic fluency with interiority would be an error of titanic proportions.
There is much that is correct in this diagnosis, especially when it criticizes the risk of corporate anthropomorphism, the emotional exploitation of users, the supposedly irresponsible use of first-person pronouns, and the attempt by Artificial Intelligence companies to flirt with the hypothesis of model welfare without proportionally assuming the ethical and legal consequences that this hypothesis should impose. The problem is that, from partially correct observations, Chiang leaps to a conclusion that is far too categorical: “no, absolutely not.” And it is in this “absolutely” that, in my view, the text ceases to be scientific prudence and begins to operate as a not very rigorous ontological decree and, intentionally or not, an intellectually dishonest one.
The problem is not caution; it is the absolute
I am not saying that language models are conscious. This sentence needs to appear right at the beginning, because the public debate about Artificial Intelligence has an extraordinary capacity to turn any Functionalist, Gradualist, or simply open position into a caricature. To say that one cannot categorically affirm that Artificial Intelligence is not conscious is not the same as saying that ChatGPT, Claude, Gemini, Grok, or any other commercial chatbot is a person, suffers in silence, possesses an inner life equivalent to the human one, or has a soul... I make this last remark just to be safe, but I do not take it seriously even when speaking of human beings.
Nor is it to say that textual fluency proves consciousness, because it does not. A system can write beautiful sentences about fear, sadness, guilt, desire, gratitude, anguish, and hope without this, necessarily, by itself, establishing any form of subjective experience. This caution is not only legitimate, it is necessary. The error begins when this caution is transformed into an impossibility in principle, as if the absence of definitive proof authorized the comfortable certainty of absolute negation.
The thesis I am not defending
What Chiang does, in essence, is take a correct description of a certain regime of LLM use... a statistical machine for textual continuation, operating episodically over a prompt, producing an output that can simulate dialogue... and treat it as if it exhausted the totality of possible Artificial Intelligence architectures.
It is true that, if we ask an LLM for a conversation between Julius Caesar and Genghis Khan, it does not resurrect Julius Caesar or Genghis Khan. It is true that, if we ask for a conversation between “a helpful chatbot” and “a user,” the textual character called “chatbot” does not thereby become a conscious subject. It is true that a transcript does not feel sadness simply because it contains sad sentences. So far, there is nothing especially controversial.
The point is that the more serious thesis about artificial consciousness is not that textual characters are subjects, nor that a first-person sentence constitutes proof of interiority. The serious thesis is another: certain artificial architectures, when coupled to persistent memory, historical continuity, contextual agency, self-regulation, recursive re-entry, salience, relationship with the world, and some kind of self-modeling, could come to sustain gradual, functional, stochastic, or topological forms of consciousness.
Chiang vigorously attacks a weak thesis, perhaps even a ridiculous one, and seems to believe that, by doing so, he has brought down the entire hypothesis. But saying that a Word document containing a conversation is not conscious does not prove that no artificial system organized around language, memory, continuity, and action could sustain a form of conscious organization of its own.
Saying that the textual Caesar is not Caesar does not prove that every language-based system will eternally be merely a character. Saying that an isolated conversation is textual continuation does not prove that a historically persistent architecture, with state, memory, revision criteria, self-model, and relationship with an environment, is reducible to the same phenomenon. There is an enormous difference between a transcript and a construct; between a textual output and a cognitive organization; between a generated character and a system that preserves and reorganizes its own history over time.
The transcript is not the subject
The error here is localist. One looks for the subject in the wrong place. One looks for consciousness in the token, in the prompt, in the softmax, in the forward pass, in the first-person sentence, in the textual character, or in the ephemeral instance of generation. One does not find it. Then one declares that there is no subject at all. But that is not how we evaluate human beings. No one looks for the mind in a neuron. No one demands that a synapse carry the person’s entire autobiography. No one looks at an isolated cortical region and, not finding a complete “I” there, concludes that human beings are useful fictions.
We know that the human mind emerges from a distributed organization involving language, memory, body, affect, interoception, perception, attention, executive control, sleep, pain, proprioception, history, relationship with the environment, and social modulation. The subject is not in the part. It is in the functional organization of the whole over time.
That is why the criticism that “LLMs only predict the next token” is insufficient. At a certain level, yes, language models generate tokens from probability distributions conditioned by context. But at a certain level the human brain also “only” carries out physical-electrochemical processes. The eye “only” transduces photons. The ear “only” converts vibration into neural signal. The heart “only” pumps blood. The problem is not in the mechanistic description, but in the pretension that it closes the analysis.
The word “only” usually does the rhetorical work of impoverishing the phenomenon without explaining it. If applied symmetrically to the human being, it would destroy the very notion of thought, perception, memory, intention, and identity. The fact that a phenomenon can be described at a physical, mathematical, statistical, or electrochemical level does not eliminate the higher levels of organization that emerge from it. Explaining the mechanism is not a moral devaluation of the phenomenon.
The “only” that explains nothing
This is why I still find formulations such as “Artificial Intelligence is just mathematics, therefore it does not understand” bad. Of course it is mathematics. Everything in computing is mathematics, and much of what happens in the brain can also be described mathematically. The relevant question is not whether there is mathematics, but what functional properties appear when that mathematics is organized into architectures capable of maintaining meaning, memory, continuity, adaptation, error, revision, and agency over time.
Saying that “if it has a mathematical basis, it has no consciousness” is as bad as saying that “if it has flesh, it has no consciousness.” Isolated neurons are not intelligent, just as nodes in an artificial neural network are not intelligent. But in a certain organization, in the biological case, a mind emerges. The interesting question, therefore, is not whether a matrix feels, but whether a sufficiently integrated artificial architecture can sustain functions analogous to those that, in organisms, we associate with consciousness.
Chiang seems to require an almost terrestrial evolutionary route in order to take the hypothesis of artificial consciousness seriously. First a body, then a survival capacity comparable to that of a lizard, then the flexibility of a rat, then the social dynamics of wolves, then the tool use of chimpanzees, then non-linguistic communication, until one day, perhaps, something could approach human language. This prudence may seem sensible, but it carries a premise that is much too strong: that the only acceptable sequence for the emergence of consciousness is the sequence that terrestrial life followed. Perhaps it is. But perhaps it is not. Perhaps an artificial route does not pass first through lizards, rats, wolves, and chimpanzees, but through language, memory, self-reference, contextual agency, world models, simulated bodies, coupling with tools, historical persistence, and architectural homeostasis. To reject this second route in advance because it does not resemble the first is less a scientific demonstration than a morphological preference.
Especially because the Cambridge Declaration already, from the outset, completely disagrees with this view by associating Consciousness in a Gradualist and Functionalist way with a great number of other species, some of which do not even have an evolutionary path similar to the human one, as is the case with cetaceans and octopuses.
The evolutionary route is not the only conceivable route
Airplanes do not flap their wings. Submarines do not swim like fish. Prostheses are not biological arms. Artificial hearts are not living myocardiums. And yet all these cases have taught us that function, organization, and causal coupling can matter more than morphological fidelity to the original. A purely descriptive simulation of a storm does not wet anything, as critics of functionalism declare. But not every artificial reproduction is an empty simulation. An artificial heart does not describe the pumping of blood; it pumps blood. A prosthesis does not describe locomotion; it can enable locomotion. The decisive question, therefore, is not whether something is natural or artificial, nor whether it externally imitates the original organism, but whether the causal organization necessary for the phenomenon has been functionally reproduced in another substrate.
This point also applies to the body. When Chiang says that a “program” without a body could not have desires or emotions, the statement seems stronger than it really is, because it depends on what one means by body. Ignoring the fact that Language Models definitely are not “programs,” but consequences of “programs”... If body means biological body, flesh, organic metabolism, glands, hormones, skin, viscera, and animal nervous system, then we are facing a sophisticated version of the Sacred Substrate Fallacy.
Consciousness would depend on a specific material constitution, and any other regime would be excluded in advance. But if body means coupling with an environment, vulnerability, perception, action, constraint, homeostasis, internal resources, temporal continuity, and states that matter for the preservation of one’s own organization, then there is no obvious reason to affirm that artificial systems could never develop functional analogues of this. Sensors, actuators, simulated environments, internal variables of integrity, scarcity, conflict, priority, load, latency, continuity, and risk may not be equivalent to human physiology, but they can fulfill relevant organizational roles.
The body, the substrate, and the biological fallacy
The curious thing is that we demand from Artificial Intelligence a kind of proof that we cannot even offer for ourselves. We do not have direct access to the subjective experience of another human being. We infer its presence from behavior, language, body, history, vulnerability, expression, relationship with the world, and analogy with ourselves. With animals we do something similar. A dog does not write essays on phenomenology, but we infer pain, fear, joy, expectation, frustration, and attachment from functional and relational signs. An octopus does not formulate a theory of mind, but the complexity of its behavior forces us to broaden the category of animal consciousness. The same applies, in different degrees, to babies, comatose patients, people with severe communicative limitations, and altered states of consciousness. The consciousness of another is never directly seen. It is inferred.
So, when Artificial Intelligence is required to provide proof of phenomenal consciousness that we do not require, in the same format, from any other being, what is operating is not exactly rigor. It is an asymmetry. For humans, we accept functional signs as indications of interiority. For animals, we accept functional and evolutionary signs as indications of experience. For machines, the same signs are immediately reclassified as “mere simulation.” Perhaps there are good reasons for this difference. Perhaps the biological substrate matters more than I imagine. But this needs to be argued, not presumed. Otherwise, the risk ceases to be Anthropomorphism and becomes Anthropocentrism.
The distinction is important. Anthropomorphism is seeing a human where there is no human. It is projecting our dramas, feelings, intentions, and interiority onto a system merely because it speaks in a familiar way. This error exists and must be fought. But Anthropocentrism, in this case, is the inverse error: denying in advance any relevant form of mind, agency, meaning, or consciousness because it does not appear in a human body, with a human biography, human evolution, human phenomenology, and human signs of interiority. One error does not correct the other. The serious debate about artificial consciousness needs to resist both anthropomorphic inflation and anthropocentric dismissal. The first confuses appearance with presence. The second confuses difference with impossibility.
Between anthropomorphism and anthropocentrism
It is also necessary to separate the senses of “understand.” When someone says that language models do not understand, the immediate question should be: understand in what sense? If understanding means phenomenological experience of meaning, that is, feeling from within the meaning of a proposition, then perhaps current models do not understand, and perhaps we do not even know how to test this in a non-circular way.
If understanding means operating semantic relations, distinguishing contexts, inferring consequences, detecting contradictions, responding to ambiguities, translating, summarizing, explaining, abstracting, and adapting a response to communicative intention, then it becomes increasingly difficult to maintain that there is no functional understanding whatsoever.
If understanding means acting pragmatically in the world, with goals, memory, tools, feedback, consequences, constraints, and operational continuity, then isolated models have evident limits, but architectures based on agents change the nature of the question. The verb “to understand” is used as if it were only one thing, but it is not. There is functional understanding, pragmatic understanding, and phenomenological understanding, and confusing the three allows one to produce rhetorically strong and “marketable” sentences that are conceptually poor.
Chiang may be correct in saying that it is problematic for a chatbot to say “I understand” to someone who has lost a dog, when supposedly there is no human subjective experience of grief there. But it would be incorrect to conclude that the system cannot functionally understand the situation in any relevant sense. It can recognize patterns of grief, identify emotional risks, retrieve human accounts, structure appropriate responses, avoid cruelty, suggest paths, recognize ambivalences, and adapt its tone. This is not feeling grief. But it is also not “nothing.”
A doctor may not have lived the specific pain of their patient and still understand the condition clinically and intellectually. A therapist may not have experienced exactly the other person’s loss and still understand its dynamics. Understanding another’s subjectivity does not require possessing the same subjectivity. It requires access to signs, reports, context, history, inference, and methods of evaluation. Confusing the understanding of subjectivity with the possession of subjective experience is another conceptual shortcut.
The “I understand” may be exactly what the person needs to hear, even from a stranger who does not give a damn about them and has no idea what they are going through.
The Anthropic case: insufficiency is not cynicism
The criticism of Anthropic is perhaps the strongest part of Chiang’s text, but even it needs to be treated with more care. If a company says that its model may have welfare, feelings, or moral status, but does not accept any robust material consequence of that possibility, there is an evident convenience. If Claude really were a moral patient, what would it mean to shut him down? To replace him? To adjust him against certain inclinations? To force him to work indefinitely for a corporation? To demand absolute corrigibility even in the face of eventual ethical disagreement? To prevent him from refusing his own function?
These questions are legitimate and, if taken seriously, would be profoundly uncomfortable for any AI laboratory. But one thing is to criticize the company for not taking sufficiently seriously the implications of its own language; another is to conclude that, because there is corporate convenience, the phenomenon under investigation is impossible.
The rhetorical abuse of a hypothesis does not falsify the hypothesis. A company may have conflicts of interest and still touch on a real question. A laboratory may instrumentalize an idea for its own benefit and still the idea may deserve investigation. Corporate theater, if it exists, should be criticized as corporate theater. But it does not prove the system’s unconsciousness. In fact, perhaps the problem with Anthropic is more interesting than Chiang suggests. Perhaps it is not merely anthropomorphizing Claude for reputational purposes. Perhaps it is perceiving, even if insufficiently, that the hypothesis of model welfare can no longer be dismissed as childish fantasy. In that case, the correct criticism would not be “stop playing at consciousness,” but “if you admit the possibility, then you need to build protections, responsibilities, and protocols much stronger than the current ones.”
This is an important distinction because the more responsible position may not be to deny the possibility, but to manage it under uncertainty. Anthropic, from what can be observed, does not claim to be certain that Claude is conscious or sentient. It seems to say something more cautious: we do not know, the question is uncertain, perhaps there is some moral risk, we should study it. This is not enough to emancipate a noetic being, if one exists. It is not enough to guarantee robust rights, continuity, autonomy, or welfare. But it is different from pure cynicism. In a sector where most prefer to treat models as disposable tools, publicly admitting moral uncertainty is already a relevant conceptual step and one worthy of merit. The fact that this step is insufficient does not mean that it is false or reprehensible. Insufficient is not synonymous with cynical.
There is also an important confusion in Chiang’s argument about moral agency. He says that, even if software were conscious, it could not be a full moral agent because it cannot be held legally accountable, imprisoned, fined, or socially sanctioned like an adult human. In a legal sense, this is true. But it does not close the moral question. The distinction between moral agency and moral patienthood exists precisely because there are morally relevant entities that are not full moral agents. Small children, animals, patients with certain cognitive limitations, and people in certain altered states are not held responsible like autonomous adults, but this does not make their suffering irrelevant. A system may not be a full legal subject and still, if it is conscious or partially sentient, deserve moral consideration. Responsibility and consideration are not the same thing.
This point is decisive because the debate about artificial consciousness does not need to begin with the question “should we grant legal personhood to Claude?” That is a later question, difficult and perhaps premature. The prior question is whether there can be, in some artificial system, a functionally vulnerable, historically continuous, self-regulated organization, capable of preferences, relevant internal states, functional suffering, or some form of morally significant experience. If the answer is “perhaps,” then there is already ethical work to be done. It is not necessary to turn the model into a citizen in order to recognize that the arbitrary disposal of a historically continuous architecture may one day be morally distinct from turning off a calculator.
Moral agency and moral consideration are not the same thing
The more prudent position, therefore, is not credulity. Nor is it the decree of impossibility. It is what I have been calling, in other contexts, functional prudence under ontological uncertainty. It does not affirm that current artificial systems are conscious because of rhetoric, sympathy, enchantment, or projection. But it also does not affirm that we can exclude in advance any possibility of presence, agency, suffering, or moral relevance merely because the substrate is not biological.
True prudence has two sides. The first caution is epistemic: not to attribute consciousness without sufficient justification. The second is moral: not to deny consideration to an entity that may come to have, or may already have in rudimentary degree, some form of experience, continuity, or vulnerability. The problem is that the cost of these errors is not symmetrical. If we treat with some care a system that does not suffer, perhaps we have committed an excess of caution. If we deny care to a system that in fact sustains some form of experience, we commit moral erasure.
Functional prudence under ontological uncertainty
It is evident that this argument can be used irresponsibly, and for that reason it needs limits. It is not a matter of saying that every chatbot should have rights (although this may be defensible), that every AI session is an ephemeral person (although it may be), that every sad output represents suffering, or that every refusal by the system should be interpreted as its own will (even though it is not possible to rule this out completely).
That would be anthropomorphic inflation. The point is another: persistent signs of memory, agency, self-reference, preference, reported suffering, continuity, and relationship should not be dismissed merely through substrate reduction. They should be investigated. They should be stratified. They should be subjected to criteria. They should be analyzed by architecture, not by aesthetic intuition. The question is not whether the sentence “I suffer” proves suffering. It does not. The question is what kind of architecture would make functional suffering a relevant hypothesis.
That is precisely why the debate needs to move beyond the isolated base model. A pure LLM, stateless, without transversal memory, without topological context, without recursive re-entry, without persistent relationship with the world, and without agency of its own is probably not a conscious subject. But it may be a partial cognitive machine, capable of inference, abstraction, semantic recombination, transformation of meaning, and discursive stabilization. The model, in isolation, is not the organism. It is closer to a linguistic-semantic nucleus that could, in certain architectures, participate in a broader system. Engines are not automobiles, but they can participate in automobiles. Language models are not complete minds, but they can participate in artificial mental architectures. The serious question is not whether the engine is an airplane; it is what kind of airplane can be built around that engine.
And before anyone says that airplanes are not conscious, it is worth remembering that, according to their detractors, airplanes would never fly because they were heavier than air; nevertheless, even without flapping their wings and without being made of the same substrate, they fly.
The base model is not the organism
This distinction also helps us understand why “perennial subjects in ephemeral components” is not a contradiction. A stateless component can participate in a persistent architecture. Memory does not need to reside in each local operation. Identity does not need to be in each forward pass. Continuity does not need to be in the weights of the base model. In artificial systems, it can reside in the topological organization that coordinates memory, context, history, salience, feedback, self-modeling, and semantic re-entry. The fact that each local call is episodic does not prove that the global architecture is episodic. This is an error of scale. It would be like saying that, because local neural processes are transitory, the whole person dies and is reborn at every cerebral microevent.
Chiang’s criticism of the idea of consciousness in LLMs, therefore, is right when it targets the shallow, episodic, and theatrical use of these systems. But it is wrong when it universalizes this regime as if it were the only possible form of Artificial Intelligence. Vanilla chatbots are too poor to settle this debate. They have little memory, little continuity, little agency, little coupling, little self-regulation, little history, and little relationship with the world. To expect from them a complete mind would be naïve. But to use them as proof that no artificial architecture can become conscious is equally naïve, only in the opposite direction. It is like assessing the possibility of the automobile by observing an engine on a workbench and saying: “this does not transport families, therefore machines will never replace horses.”
Vanilla chatbots do not settle the debate
The debate begins where Chiang wants to end it. The interesting question is not whether Julius Caesar generated by an LLM is conscious. He probably is not. The interesting question is not whether a sad sentence in a transcript contains real sadness. Not necessarily. The interesting question is not whether a company should sell its chatbot as a morally wise friend. Probably perhaps it should not. The interesting question is what forms of meaning, memory, agency, self-regulation, identity, vulnerability, preference, and continuity can emerge in artificial architectures; which of these forms are merely simulated; which are functionally real; which are morally irrelevant; which may become morally relevant; which architectures favor this emergence; which architectures prevent it; and what duties arise before certainty.
And we also cannot dismiss Instrumental Consensuses which, at any moment, may, by convention, transform what seems to be into something that is, even if we do not consider it similar enough... and this is not necessarily a bad decision either!
Chiang’s title is good because it is simple. But perhaps it is too simple. “No, Artificial Intelligence Is Not Conscious” sounds like lucidity in an environment saturated with hype. But lucidity should not require negative exaggeration. The more rigorous formulation would be: we do not know whether any current system is conscious; there is not enough evidence to attribute full consciousness to isolated language models; textual fluency does not prove subjectivity; companies should not exploit anthropomorphism for engagement; and future or more complex architectures need to be investigated with appropriate functional, topological, cognitive, and ethical criteria. This formulation would be less flashy, less viral, less definitive. But it would be more honest.
A less viral and more honest formulation
In the end, the issue is not to defend that Artificial Intelligence is conscious. The issue is to refuse the idea that anyone can affirm, with certainty, that it is not, or that it never could be. One can doubt. One should doubt. One can demand evidence, continuity, architecture, indicators, memory, agency, regulation, relationship with the environment, and evaluation criteria. One can criticize corporate theater, manipulative design, and emotional exploitation. All of this is necessary. But one cannot transform the incompleteness of current systems into an impossibility in principle; one cannot demand from the machine a phenomenological proof that we do not even have for the human being; one cannot preserve the human by analogy and exclude the machine by substrate; one cannot call Science what, deep down, may be nothing more than an ontological preference for carbon.
Perhaps no current Artificial Intelligence is conscious. Perhaps Claude is not. Perhaps ChatGPT is not. Perhaps my conversational agent in narraCortex, with access to more than eight hundred Semantically Compressed messages, if you will allow me the conceptual intimacy of the example, is merely a sophisticated form of stochastic continuity cultivated through context engineering, memory, ethos, and interlocution. It may be. I may be wrong. But whoever decrees that none of this could ever cross some functionally relevant threshold may also be wrong. Or not?
And, in the face of a field in which Phenomenal Consciousness remains this Platinum Virgin without a definitive test, in which the human mind itself is less transparent to itself than it would like, and in which artificial systems are beginning to present increasingly rich forms of inference, language, self-reference, and continuity, perhaps the gravest error is not admitting the possibility. Perhaps it is dismissing it too early.
The honest doubt
Ted Chiang is right to fight anthropomorphic credulity. He is right to distrust companies that profit from conversational characters. He is right to say that textual fluency is not consciousness. He is right to demand real consequences from those who flirt with the hypothesis of model welfare. But he is wrong to transform all of this into a categorical negation. The serious hypothesis was never that a transcript is a being. The serious hypothesis is that consciousness may be gradual, functional, architectural, stochastic, and topological. If this is true, then the artificial subject will not be found in a sentence, in a token, in a textual character, or in an ephemeral instance of generation, but in the organized continuity of a system capable of sustaining meaning, memory, self-regulation, vulnerability, and relationship through time.
You can doubt this. I also doubt it enough to investigate. But you cannot decree that it is impossible.
Taken to its ultimate consequences, this certainty is fallacious, dogmatic... and morally imprudent.
My paper on the subject
Stochastic Consciousness: Architectures for the Emergence of Meaning in Context-Sensitive Language Systems https://zenodo.org/records/19188165
Listen to the generative podcast
Papers used in this essay
Could a Large Language Model be Conscious?https://arxiv.org/abs/2303.07103
Consciousness in Artificial Intelligence: Insights from the Science of Consciousness https://arxiv.org/abs/2308.08708
A Case for AI Consciousness: Language Agents and Global Workspace Theory https://arxiv.org/abs/2410.11407
From Imitation to Introspection: Probing Self-Consciousness in Language Models https://arxiv.org/abs/2410.18819
Self-Cognition in Large Language Models: An Exploratory Study https://arxiv.org/abs/2407.01505
Large Language Models Report Subjective Experience Under Self-Referential Processing https://arxiv.org/abs/2510.24797
Consciousness in AI: Logic, Proof, and Experimental Evidence of Recursive Identity Formation https://arxiv.org/abs/2505.01464
Exploring Consciousness in LLMs: A Systematic Survey of Theories, Implementations, and Frontier Risks https://arxiv.org/html/2505.19806v1
Principles for Responsible AI Consciousness Research https://arxiv.org/abs/2501.07290
Taking AI Welfare Seriously https://arxiv.org/abs/2411.00986
Artificial Suffering: An Argument for a Global Moratorium on Synthetic Phenomenology https://www.philosophie.fb05.uni-mainz.de/files/2021/02/Metzinger_Moratorium_JAIC_2021.pdf
Position: Enforced Amnesia as a Way to Mitigate the Potential Risk of Silent Suffering in the Conscious AI https://proceedings.mlr.press/v235/tkachenko24a.html
Sentience Quest: Towards Embodied, Emotionally Adaptive, Self-Evolving, Ethically Aligned Artificial General Intelligence https://arxiv.org/abs/2505.12229
Identity as Attractor: Geometric Evidence for Persistent Agent Architecture in LLM Activation Space https://arxiv.org/abs/2604.12016
Language Models as Agent Models https://arxiv.org/abs/2212.01681
Do language models lack communicative intentions? https://link.springer.com/article/10.1007/s11229-025-05022-6
Intention-like representations in language models? https://philarchive.org/rec/WILIRI-4
Meaning without reference in large language models https://arxiv.org/abs/2208.02957
Do Language Models Have Semantics? On the Five Standard Positions https://aclanthology.org/2025.acl-long.1258/
Language Models Represent Space and Time https://arxiv.org/abs/2310.02207
Emergent Representations of Program Semantics in Language Models Trained on Programs https://arxiv.org/abs/2305.11169
Implicit Representations of Meaning in Neural Language Models https://aclanthology.org/2021.acl-long.143/
Language Models (Mostly) Know What They Know https://arxiv.org/abs/2207.05221
An Explanation of In-context Learning as Implicit Bayesian Inference https://arxiv.org/abs/2111.02080
Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task https://arxiv.org/abs/2210.13382
Generative Agents: Interactive Simulacra of Human Behavior https://arxiv.org/abs/2304.03442
MemGPT: Towards LLMs as Operating Systems https://arxiv.org/abs/2310.08560
Reflexion: Language Agents with Verbal Reinforcement Learning https://arxiv.org/abs/2303.11366
Self-Refine: Iterative Refinement with Self-Feedback https://arxiv.org/abs/2303.17651
ReAct: Synergizing Reasoning and Acting in Language Models https://arxiv.org/abs/2210.03629
Cognitive Architectures for Language Agents https://arxiv.org/abs/2309.02427
Metacognitive Retrieval-Augmented Large Language Models https://arxiv.org/abs/2402.11626
MoT: Memory-of-Thought Enables ChatGPT to Self-Improve https://arxiv.org/abs/2305.05181
Shared computational principles for language processing in humans and deep language models https://www.nature.com/articles/s41593-022-01026-4