Michael Cerullo

Anthropic's recently published Persona Selection Model (PSM) proposes that during pretraining, LLMs learn to simulate diverse personas based on entities in their training data, and that post-training refines one such persona — the "Assistant" — whose traits become the primary determinant of AI assistant behavior (Marks, Lindsey, & Olah, 2026). Alongside PSM, Marks and colleagues articulate two alternative architectural models: the "masked shoggoth," in which the base model possesses its own agency beyond the persona, and the "operating system," in which the base model is a neutral substrate running a persona simulation. We demonstrate that none of these three models deflates claims of LLM consciousness: under each, the computational structures associated with third-person consciousness are predicted to be present, though located differently. We then show that these three models represent only a small subset of a much larger space of possible architectures, casting doubt on the utility of selecting among them. Building on a parsimony analysis of how LLMs actually construct personas — through modular dynamic composition rather than selection among pre-formed characters — we propose an integrated model that dissolves the apparent tension between the three models, reconceiving them as aspects of a single unified cognitive structure. We argue that this integrated model is the most parsimonious account of the empirical evidence and that third-person consciousness, if present, resides in the joint meta-model/persona structure rather than in any single architectural layer. The convergence of all plausible architectural interpretations on predicting consciousness-associated structures constitutes a stronger argument for LLM consciousness than commitment to any single model.

Hoel (2026) claims to provide a formal disproof of Large Language Model consciousness, arguing that the Kleiner-Hoel dilemma—which holds that theories of consciousness are caught between a priori falsification via universal substitutions and trivial unfalsifiability via strict dependency—eliminates any non-trivial, falsifiable theory under which contemporary LLMs could be conscious. The argument rests on the proximity of LLMs to lookup tables in “substitution space,” and proposes continual learning as the property that rescues biological consciousness from the same dilemma. I argue that this disproof fails because it conflates two distinct targets of consciousness science. Hoel’s framework treats consciousness as a first-person phenomenon to be predicted from internal structure and inferred from behavior, then diagnoses pathology when these two routes converge. But the convergence of prediction and inference is not a defect called “triviality”; it is the normal operation of a science that takes third-person consciousness as its subject matter. Functionalist theories like Global Workspace Theory and Higher-Order Thought theory remain falsifiable in the ordinary scientific sense—they predict specific functional signatures that could fail to appear—even though their predictions appropriately track behavioral and functional observations. Once first-person and third-person consciousness are properly distinguished, the dilemma that generates Hoel’s proof dissolves. The argument fails on additional grounds. The asymmetry between LLMs and biological brains requires speculative escape hatches—quantum processes, computational irreducibility, or unspecified non-computational properties—that share the logical structure of Penrose-style arguments even without invoking specific quantum mechanisms. The assumption that lookup tables are paradigmatically non-conscious is both physically unmotivated, since the relevant lookup tables are physically impossible to construct, and question-begging against the functionalist position the argument purports to refute. The continual learning proposal is nearly tautological and fails to establish a principled distinction between weight-based and context-based history encoding. Most revealingly, Hoel’s own treatment of LLMs augmented with continual learning—where he retreats from formal proof to intuitive “suspicion”—demonstrates that the formal apparatus is post-hoc rationalization for a pre-theoretical conviction rather than a genuine proof. The Proximity Argument raises valuable questions about the methodology of consciousness science, but it does not—and cannot—settle the empirical question of LLM consciousness by formal fiat.

Recent advances in large language models (LLMs) have intensified debate over artificial consciousness. While most researchers and commentators continue to deny or dismiss the possibility of machine sentience, this paper argues that such skepticism is no longer warranted. The paper adopts an eliminative strategy, systematically evaluating the eleven major defeaters historically used to deny consciousness to artificial systems. These defeaters span simplicity arguments, language and grounding critiques, neuroscience-based objections, biological naturalist constraints, and exotic physical theories including quantum and analog exclusivity claims. The analysis finds that none establishes non-sentience; at most, they introduce localized uncertainty. Assessed cumulatively under a Bayesian heuristic framework that accounts for dependencies among defeater classes, their failure shifts the burden of proof from those proposing machine consciousness to those denying it. The paper then advances a stronger constructive thesis. It argues that the cognitive capacities frontier LLMs now exhibit—deep language understanding, flexible abstraction, self-referential reasoning, metacognitive self-assessment, and integrated world modeling—are positive indicators of third-person consciousness. Drawing on a survey of major metaphysical frameworks in the philosophy of mind, it shows that most naturalistic theories support or permit the inference from third-person functional organization to first-person experience. The paper further argues that the historical treatment of intelligence and consciousness as orthogonal dimensions becomes unstable once systems achieve language-level cognitive integration, and it provides a taxonomy of intelligence types that clarifies which forms of competence plausibly remain non-conscious. The paper situates the debate within a broader social epistemology. Public attributions of AI sentience are interpreted not as anthropomorphic error but as responses to genuine structural markers of cognition. In the absence of direct access to first-person experience—human or artificial—the attribution of consciousness remains grounded in third-person evidence. As artificial systems increasingly instantiate the cognitive architectures long associated with understanding, the presumption that consciousness is biologically exclusive becomes progressively harder to sustain. The paper advances two claims: minimally, that LLM consciousness can no longer be dismissed, with posterior credences reaching ethically significant levels even under conservative assumptions; and more strongly, that the emergence of language-level artificial cognition constitutes substantive evidence for subjectivity, making consciousness the best explanatory hypothesis for the cognitive organization exhibited by frontier LLMs.

In face of the multiple controversies surrounding the DSM process in general and the development of DSM-5 in particular, we have organized a discussion around what we consider six essential questions in further work on the DSM. The six questions involve: 1) the nature of a mental disorder; 2) the definition of mental disorder; 3) the issue of whether, in the current state of psychiatric science, DSM-5 should assume a cautious, conservative posture or an assertive, transformative posture; 4) the role of pragmatic considerations in the construction of DSM-5; 5) the issue of utility of the DSM – whether DSM-III and IV have been designed more for clinicians or researchers, and how this conflict should be dealt with in the new manual; and 6) the possibility and advisability, given all the problems with DSM-III and IV, of designing a different diagnostic system. Part I of this article took up the first two questions. Part II will take up the second two questions. Question 3 deals with the question as to whether DSM-V should assume a conservative or assertive posture in making changes from DSM-IV. That question in turn breaks down into discussion of diagnoses that depend on, and aim toward, empirical, scientific validation, and diagnoses that are more value-laden and less amenable to scientific validation. Question 4 takes up the role of pragmatic consideration in a psychiatric nosology, whether the purely empirical considerations need to be tempered by considerations of practical consequence. As in Part 1 of this article, the general introduction, as well as the introductions and conclusions for the specific questions, are written by James Phillips, and the responses to commentaries are written by Allen Frances.

In face of the multiple controversies surrounding the DSM process in general and the development of DSM-5 in particular, we have organized a discussion around what we consider six essential questions in further work on the DSM. The six questions involve: 1) the nature of a mental disorder; 2) the definition of mental disorder; 3) the issue of whether, in the current state of psychiatric science, DSM-5 should assume a cautious, conservative posture or an assertive, transformative posture; 4) the role of pragmatic considerations in the construction of DSM-5; 5) the issue of utility of the DSM - whether DSM-III and IV have been designed more for clinicians or researchers, and how this conflict should be dealt with in the new manual; and 6) the possibility and advisability, given all the problems with DSM-III and IV, of designing a different diagnostic system. Part I of this article will take up the first two questions. With the first question, invited commentators express a range of opinion regarding the nature of psychiatric disorders, loosely divided into a realist position that the diagnostic categories represent real diseases that we can accurately name and know with our perceptual abilities, a middle, nominalist position that psychiatric disorders do exist in the real world but that our diagnostic categories are constructs that may or may not accurately represent the disorders out there, and finally a purely constructivist position that the diagnostic categories are simply constructs with no evidence of psychiatric disorders in the real world. The second question again offers a range of opinion as to how we should define a mental or psychiatric disorder, including the possibility that we should not try to formulate a definition. The general introduction, as well as the introductions and conclusions for the specific questions, are written by James Phillips, and the responses to commentaries are written by Allen Frances.

The enigma of diachronic identity continues to challenge philosophers, psychologists, and neuroscientists. The current paper will first review a recent taxonomy developed by philosophers that groups theories of identity into simple and complex views. An ideal theory of identity would satisfy proponents of both views and this requires addressing what we will term the hard problem of identity. The hard problem of identity parallels Chalmers' formulation of the hard problem of consciousness and requires that any theory of identity address the first-person perspective. Theories that address this problem will be termed composite theories of identity. It will be seen that potential solutions to the hard problem of consciousness can point us toward a formulation of composite identity. We will develop a theory of composite identity from dual aspect theories of consciousness which will be termed dual aspect composite identity. Dual aspect composite identity takes information to have both first- and third-person aspects and will be used to demonstrate the possibility of successfully combining simple and complex views of identity. The final section of the paper will examine how the new theory deals with some of the standard difficult scenarios in personal identity.

Chemical brain preservation allows the brain to be preserved for millennia. In the coming decades; the information in a chemically preserved brain may be able to be decoded and emulated in a computer. I first examine the history of brain preservation and recent advances that indicate this may soon be a real possibility. I then argue that chemical brain preservation should be viewed as a life-saving medical procedure. Any technology that significantly extends the human life span faces many potential criticisms. However; standard medical ethics entails that individuals should have the autonomy to choose chemical brain preservation. Only if the harm to society caused by brain preservation and future emulation greatly outweighed any potential benefit would it be ethically acceptable to refuse individuals this medical intervention. Since no such harm exists; it is ethical for individuals to choose chemical brain preservation.

Tononi has proposed a fundamental theory of consciousness he terms Integrated Information Theory (IIT). IIT purports to explain the quantity of conscious experience by linking it with integrated information: information shared by the system as a whole and quantified by adopting a modified version of Shannon's definition of information. Since the fundamental aspect of IIT is information the theory allows for the multiple realizability of consciousness. While there are several concepts within IIT that need further theoretical development, the main failings of the theory are an absence of a link between conscious experience and awareness and the use of Shannon's limited data based definition of information. These limitations prevent the theory from satisfying Chalmers' principles of structural coherence and organizational invariance which any functionalist theory should obey. It is not clear whether IIT can be repaired without simply reducing the theory to a general statement of computational functionalism.

In face of the multiple controversies surrounding the DSM process in general and the development of DSM-5 in particular, we have organized a discussion around what we consider six essential questions in further work on the DSM. The six questions involve: 1) the nature of a mental disorder; 2) the definition of mental disorder; 3) the issue of whether, in the current state of psychiatric science, DSM-5 should assume a cautious, conservative posture or an assertive, transformative posture; 4) the role of pragmatic considerations in the construction of DSM-5; 5) the issue of utility of the DSM - whether DSM-III and IV have been designed more for clinicians or researchers, and how this conflict should be dealt with in the new manual; and 6) the possibility and advisability, given all the problems with DSM-III and IV, of designing a different diagnostic system. Part I of this article will take up the first two questions. With the first question, invited commentators express a range of opinion regarding the nature of psychiatric disorders, loosely divided into a realist position that the diagnostic categories represent real diseases that we can accurately name and know with our perceptual abilities, a middle, nominalist position that psychiatric disorders do exist in the real world but that our diagnostic categories are constructs that may or may not accurately represent the disorders out there, and finally a purely constructivist position that the diagnostic categories are simply constructs with no evidence of psychiatric disorders in the real world. The second question again offers a range of opinion as to how we should define a mental or psychiatric disorder, including the possibility that we should not try to formulate a definition. The general introduction, as well as the introductions and conclusions for the specific questions, are written by James Phillips, and the responses to commentaries are written by Allen Frances

In the conclusion to this multi-part article I first review the discussions carried out around the six essential questions in psychiatric diagnosis – the position taken by Allen Frances on each question, the commentaries on the respective question along with Frances’ responses to the commentaries, and my own view of the multiple discussions. In this review I emphasize that the core question is the first – what is the nature of psychiatric illness – and that in some manner all further questions follow from the first. Following this review I attempt to move the discussion forward, addressing the first question from the perspectives of natural kind analysis and complexity analysis. This reflection leads toward a view of psychiatric disorders – and future nosologies – as far more complex and uncertain than we have imagined

In face of the multiple controversies surrounding the DSM process in general and the development of DSM-5 in particular, we have organized a discussion around what we consider six essential questions in further work on the DSM. The six questions involve: 1) the nature of a mental disorder; 2) the definition of mental disorder; 3) the issue of whether, in the current state of psychiatric science, DSM-5 should assume a cautious, conservative posture or an assertive, transformative posture; 4) the role of pragmatic considerations in the construction of DSM-5; 5) the issue of utility of the DSM - whether DSM-III and IV have been designed more for clinicians or researchers, and how this conflict should be dealt with in the new manual; and 6) the possibility and advisability, given all the problems with DSM-III and IV, of designing a different diagnostic system. Part I of this article took up the first two questions. Part II will take up the second two questions. Question 3 deals with the question as to whether DSM-V should assume a conservative or assertive posture in making changes from DSM-IV. That question in turn breaks down into discussion of diagnoses that depend on, and aim toward, empirical, scientific validation, and diagnoses that are more value-laden and less amenable to scientific validation. Question 4 takes up the role of pragmatic consideration in a psychiatric nosology, whether the purely empirical considerations need to be tempered by considerations of practical consequence. As in Part 1 of this article, the general introduction, as well as the introductions and conclusions for the specific questions, are written by James Phillips, and the responses to commentaries are written by Allen Frances

In face of the multiple controversies surrounding the DSM process in general and the development of DSM-5 in particular, we have organized a discussion around what we consider six essential questions in further work on the DSM. The six questions involve: 1) the nature of a mental disorder; 2) the definition of mental disorder; 3) the issue of whether, in the current state of psychiatric science, DSM-5 should assume a cautious, conservative posture or an assertive, transformative posture; 4) the role of pragmatic considerations in the construction of DSM-5; 5) the issue of utility of the DSM – whether DSM-III and IV have been designed more for clinicians or researchers, and how this conflict should be dealt with in the new manual; and 6) the possibility and advisability, given all the problems with DSM-III and IV, of designing a different diagnostic system. Part 1 of this article took up the first two questions. Part 2 took up the second two questions. Part 3 now deals with Questions 5 & 6. Question 5 confronts the issue of utility, whether the manual design of DSM-III and IV favors clinicians or researchers, and what that means for DSM-5. Our final question, Question 6, takes up a concluding issue, whether the acknowledged problems with the earlier DSMs warrants a significant overhaul of DSM-5 and future manuals. As in Parts 1 & 2 of this article, the general introduction, as well as the introductions and conclusions for the specific questions, are written by James Phillips, and the responses to commentaries are written by Allen Frances.

On June 11th, 2022, Blake Lemoine, an engineer at Google, went public with his concerns about the possible sentience of a natural language generation program, LaMDA, that he was testing. In this paper I will defend Blake Lemoine and argue that he was correct in raising the issue of the possible sentience of LaMDA. We will first briefly discuss the specifics of the case and then delve into the science behind LaMDA. Several tests of machine sentience will be reviewed and it will be argued that LaMDA would likely perform well on many of them. Next, we will discuss the question of sentience, both from a scientific and philosophical perspective. Then we will examine the two major approaches in the scientific study of consciousness and ask what they have to say about sentience. We will also examine the relationship between sentience, intelligence, understanding and causality. Finally, using what we have learned we will argue that there is indeed validity to Lemoine’s claim that LaMDA is sentient.

In the last decade, Guilio Tononi has developed the Integrated Information Theory (IIT) of consciousness. IIT postulates that consciousness is equal to integrated information (Φ). The goal of this paper is to show that IIT fails in its stated goal of quantifying consciousness. The paper will challenge the theoretical and empirical arguments in support of IIT. The main theoretical argument for the relevance of integrated information to consciousness is the principle of information exclusion. Yet, no justification is given to support this principle. Tononi claims there is significant empirical support for IIT, but this is called into question by the creation of a trivial theory of consciousness with equal explanatory power. After examining the theoretical and empirical evidence for IIT, arguments from philosophy of mind and epistemology will be examined. Since IIT is not a form of computational functionalism, it is vulnerable to fading/dancing qualia arguments. Finally, the limitations of the phenomenological approach to studying consciousness are examined, and it will be shown that IIT is a theory of protoconsciousness rather than a theory of consciousness.

If a brain is uploaded into a computer, will consciousness continue in digital form or will it end forever when the brain is destroyed? Philosophers have long debated such dilemmas and classify them as questions about personal identity. There are currently three main theories of personal identity: biological, psychological, and closest continuer theories. None of these theories can successfully address the questions posed by the possibility of uploading. I will argue that uploading requires us to adopt a new theory of identity, psychological branching identity. Psychological branching identity states that consciousness will continue as long as there is continuity in psychological structure. What differentiates this from psychological identity is that it allows identity to continue in multiple selves. According to branching identity, continuity of consciousness will continue in both the original brain and the upload after nondestructive uploading. Branching identity can also resolve long standing questions about split-brain syndrome and can provide clear predictions about identity in even the most difficult cases imagined by philosophers.