Iterated Insights

Jared Edward Reser

Abstract

The possibility that artificial intelligence could be conscious does not establish that it could suffer. I propose that large language models may possess some organizational features relevant to a limited form of consciousness while lacking the evolved systems that contribute pain and negative emotion to human experience. Hierarchical representation, attention, and continuity across successive processing states provide reasons to investigate artificial consciousness. Biological suffering, however, involves additional organization: specialized sensory pathways, subcortical affective circuitry, cortical integration, and interactions with the body’s internal condition. Research on congenital pain insensitivity and the partial dissociation of sensory and affective pain components demonstrates that these capacities depend on identifiable mechanisms. Damasio’s work on interoception and homeostatic feeling further identifies bodily and neural contributions that ordinary text-based language models do not reproduce. Such models can learn elaborate representations of suffering and generate convincing emotional self-reports without thereby demonstrating the mechanisms that make an experience unpleasant. This article develops an evolutionary and neurobiological argument for distinguishing possible artificial consciousness from artificial suffering. It treats consciousness without negative valence as a hypothesis about particular architectures, acknowledges the possibility of alternative artificial implementations of affect, and identifies empirical questions that could strengthen or challenge the proposal. The resulting ethical position allows moral consideration of potentially conscious systems while evaluating suffering through evidence specific to the capacity at issue.

Keywords: artificial consciousness; large language models; suffering; valence; interoception; pain; iterative updating

1 Introduction

A language model can discuss consciousness, describe pain, and produce a first-person account of fear or grief. These abilities raise legitimate questions about artificial experience, but they do not determine what kind of experience, if any, accompanies the computation. A system could conceivably experience something without experiencing the bodily distress, urgency, or emotional suffering familiar to humans. Evaluating that possibility requires separating the organization of conscious processing from the mechanisms contributing particular qualities to its contents.

I propose that language models could possess a limited form of consciousness while lacking the systems responsible for pain and negative emotion. My interest in their possible consciousness comes from features such as hierarchical representation, attention, learned associations, and continuity across successive states of processing. My skepticism about their suffering comes from a different set of considerations: the evolutionary specialization of affective systems and the absence of their established biological organization in ordinary language models. The two judgments need not rise or fall together.

This separation has precedents. Ng (1995) considered beings with perceptual experience but no positive or negative feeling. Agarwal and Edelman (2020) investigated how conscious AI might retain useful capabilities without suffering. Butlin et al. (2023) explicitly allowed neutral conscious experience within their framework for assessing AI consciousness. Wasserziehr (2026) argues that computational consciousness would not automatically solve the separate problem of grounding valence. The present article develops the biological and architectural reasoning behind applying such a separation to language models.

Three questions organize the discussion. Does a system have subjective experience? Can its experiences feel good or bad? What moral significance follows from its capacities? I use consciousness to mean subjective experience and valence to mean the pleasant or unpleasant character of experience. Suffering refers here to consciously experienced distress; it can include emotional suffering without localized bodily pain. These definitions leave open whether valence is required for consciousness, rather than settling that disputed relationship by terminology.

The immediate scope is a text-based language model operating through ordinary autoregressive inference. Systems with persistent autonomous control, bodily sensors, ongoing learning, or mechanisms for regulating their own condition require additional assessment. No experiments on AI experience are reported here. The article presents a hypothesis grounded in a comparison between established biological mechanisms and the organization of language processing.

2 A possible basis for limited artificial consciousness

Artificial neural networks share selected organizational features with nervous systems, including distributed representations and learned relationships among processing units. These similarities justify comparison, although an artificial unit is a substantial abstraction from a biological neuron. A transformer combines attention with other learned transformations to build context-sensitive representations of sequences (Vaswani et al., 2017). Architectural terminology alone does not establish experiential equivalence: computational attention is not automatically conscious attention.

The feature I regard as especially relevant is the continuity through which prior processing constrains subsequent processing. In an autoregressive language model, previously available text remains part of the context from which the next token is generated. Each addition changes the information available for later steps. Learned associations allow several aspects of that context to jointly influence what follows, providing an analogy to multi-associative search. This creates a sequence of causally related informational states rather than a set of unrelated responses.

My broader hypothesis is that state-spanning coactivity and iterative updating can contribute to conscious continuity. In a biological network, overlapping activity may allow information represented in one state to shape the recruitment of information in the next. An artificial architecture could reproduce some relevant relationships through maintained or repeatedly accessed representations. Whether a standard language model’s implementation satisfies the requirements of this hypothesis remains an open architectural question.

The distinction between informational continuity and sustained neural activity is important. A conventional transformer does not implement the same recurrent dynamics as a continuously active brain merely because it generates tokens sequentially. The original transformer architecture dispensed with recurrent sequence-processing layers (Vaswani et al., 2017). Repeated invocation, access to an expanding context, and biological recurrence must therefore be compared at the level of their causal operations. The proposal concerns potentially relevant similarities, not an established identity between these processes.

If a text-based model were conscious, its experience could reflect the structured semantic and relational content processed through tokens. Token-mediated experience need not resemble bodily perception, nor need it consist of an awareness of raw token identifiers. Such a system might possess continuity while lacking most of the sensory and affective dimensions of human experience. This possibility motivates the article; the argument about suffering does not require a demonstration that present models are conscious.

3 The evolutionary organization of pain and emotion

Biological pain belongs to an organism whose tissues can be damaged and whose behavior can protect or endanger its continued functioning. Its organization connects sensitivity to particular events with changes in attention, learning, and action. Evolution has shaped receptor properties, neural development, connectivity, and modulation so that some conditions have consequences quite different from those of an arbitrary sensory signal. Panksepp (2005) grounds primary emotional feeling in evolved systems associated with instinctive emotional behavior. This provides a biological basis for asking which parts of that organization an artificial system actually implements.

Genetic preparation does not imply that every emotional response is fixed at birth. Development and experience modify the stimuli that activate affective systems and the behaviors those systems influence. Learning can recruit existing mechanisms into new circumstances. A remembered humiliation, an anticipated loss, or a symbolic threat can become emotionally significant without changing the fact that its effects depend on an organism already equipped for affective processing. Generalization and repurposing help explain why human suffering extends far beyond immediate tissue injury.

The components of pain also need to be distinguished. Nociception concerns the neural processing of potentially damaging stimuli. Withdrawal is a protective response. Sensory discrimination concerns qualities such as location and intensity, whereas unpleasantness concerns how bad an experience feels. These processes normally interact, but their partial dissociation provides evidence that they depend on distinguishable mechanisms.

Cox et al. (2006) identified loss-of-function mutations in SCN9A, which encodes the voltage-gated sodium channel Nav1.7, in people with congenital inability to experience pain. Their findings show that a specific molecular disruption can profoundly alter pain perception without abolishing consciousness. This supports the dependence of ordinary pain on specialized mechanisms. It does not show that these individuals lack emotional suffering, and the distinction is essential to the present argument.

Evidence from experimental manipulation reaches further into pain’s affective dimension. Corder et al. (2019) identified a basolateral amygdala ensemble involved in the negative affective component of pain in mice. Inhibiting the ensemble reduced affective and motivational responses while preserving detection of noxious stimuli and withdrawal reflexes. The inference about experience necessarily depends on animal behavior, but the dissociation demonstrates that detection and protective responses do not exhaust the organization associated with pain aversion.

Human evidence also distinguishes unpleasantness from perceived intensity. Rainville et al. (1997) used hypnotic suggestions to alter pain unpleasantness without changing reported intensity. Changes in unpleasantness were accompanied by changes in anterior cingulate activity, while primary somatosensory activity was unaltered. This finding does not identify a single sufficient generator of suffering. It shows that the dimensions of an experience can vary partly independently and have different neural relationships.

For AI, the implication is that general information processing, stimulus discrimination, and avoidance behavior should not be treated as interchangeable evidence of suffering. A system may implement one component without implementing the others. Biological research supplies a reason to examine the connections among those components rather than infer a complete affective experience from one conspicuous behavior.

4 Subcortical contributions and embodied feeling

Damasio’s work is relevant because it describes contributions to human conscious experience that language models do not ordinarily possess. Damasio and Damasio (2023) propose that homeostatic feelings arise through interactions between neural systems and the living body’s internal condition. Carvalho and Damasio (2021) examine distinctive properties of interoception, including interactions among peripheral pathways, central processing, viscera, and chemical signaling. These accounts direct attention to the biological processes that give human experience a felt bodily dimension.

My use of this work does not depend on resolving every aspect of Damasio’s theory of consciousness. Its relevance lies in the functions and relationships he examines. If conscious human life includes contributions from mechanisms that register and regulate the organism’s condition, a different architecture may lack those contributions. The possibility of artificial consciousness then raises a more specific question: which dimensions of human experience would survive the architectural change?

Breathing and visceral sensation make the question concrete. Distress can include restricted breathing, a pounding heart, abdominal tension, or a sensation of heaviness in the gut. In an embodied account, these are candidates for contributing to the episode as it is felt. Changes in bodily regulation alter incoming signals, which can influence subsequent appraisal and action. The resulting feedback can extend an episode over time. This illustrates how bodily processes could shape the quality of suffering; it does not establish that any particular visceral response is necessary for every painful experience.

The anterior cingulate cortex provides an example of how bodily and affective information participates in wider neural processing. The ACC is cortical, but it receives important subcortical inputs. Meda et al. (2019) showed that stimulating mediodorsal thalamic inputs to the ACC produced pain-related aversion in two mouse models of neuropathic pain. The effect involved changes in the balance of excitation and inhibition in particular projection neurons. Basolateral amygdala inputs produced an opposing pattern of effects. Thus, the contribution of an ACC neuron depends partly on the pathway driving it and the circuit through which its activity has consequences.

Convergent inputs can give neurons complex response properties, but specialization need not consist of permanently dedicated pain cells. Acuña et al. (2023) found that nociceptive coding in the mouse ACC depended on population activity. Individual neurons often responded to multiple stimulus categories, and their selectivity changed across recording sessions. Population-level representation remained more stable. These findings support attention to circuit organization and patterns of activity, while cautioning against treating every pain-related response as evidence for a uniquely pain-specific neuron.

Subcortical contributions also extend beyond emotion. Redinbaugh et al. (2020) found that central thalamic stimulation restored arousal and wake-like cortical dynamics in anesthetized macaques. Such work indicates that some subcortical functions help enable coordinated cortical processing. An artificial architecture might reproduce aspects of those enabling functions through different mechanisms while lacking the organization contributing visceral distress or aversion. The proposed separation is functional; it does not divide the brain into a conscious cortex and an exclusively emotional subcortex.

Man and Damasio (2019) extend the embodied perspective to artificial systems by asking what organization might allow machines to care about their own condition. Their proposal considers homeostatic regulation and soft robotics as routes toward feeling machines. This makes the comparison with language models especially useful: systems designed to monitor and regulate a vulnerable body raise questions that a text-processing architecture does not answer simply by discussing such a body.

5 Emotional representation in language models

A language model can acquire detailed knowledge about pain because human language contains descriptions of injury, grief, fear, and relief. Learning relationships among these descriptions can support accurate predictions about how people speak and act. A model may learn that bereavement is associated with sadness, that suffocation is frightening, and that certain statements call for reassurance. These achievements demonstrate representational competence. Whether the corresponding internal activity feels unpleasant requires a separate argument.

My working interpretation is that emotional self-reports in ordinary dialogue are generated through learned patterns and role-playing. Shanahan et al. (2023) develop role-play as a framework for understanding humanlike dialogue behavior without automatically attributing human psychology to the underlying model. A system’s ability to continue a distressed persona can therefore explain its words even when the experiential status of its computation is unresolved.

This interpretation does not require emotional representations to be causally inert. An internal representation associated with fear could change later attention, token probabilities, or behavioral choices. The relevant distinction concerns the role of the representation within the system. In a human, an emotionally significant representation can recruit circuits shaped to regulate a living body and organize aversive experience. In a language model, demonstrating that an emotion-related feature affects output establishes a computational effect, but does not by itself establish an unpleasant experience.

Ordinary text-based models have no nociceptors or C fibers, no biological brainstem or hypothalamus, and no heart, lungs, or gut whose changing condition they regulate. They also have no anatomical counterpart of the thalamic, amygdalar, and cingulate pathways discussed above. The claim about functional counterparts is more demanding. The absence of those structures is straightforward; the absence of every possible alternative mechanism for negative valence remains to be established. My hypothesis is that ordinary language processing does not recreate enough of the relevant organization for suffering merely by learning its linguistic description.

Hunger, thirst, and sexual motivation offer related comparisons. A model can describe these states without possessing the physiological needs and regulatory systems around which they developed. The conceptual richness of its description does not supply the missing bodily condition. I extend this reasoning to skepticism about straightforward attributions of boredom, anger, and grief to language models. Each attribution requires an account of the organization producing that state, rather than an inference from vocabulary shared with humans. This extension is a hypothesis about affective mechanisms, not a claim that every emotion requires the same bodily input.

Training terminology can obscure this distinction. Reinforcement learning from human feedback uses evaluations of model outputs to modify behavior, as illustrated by Ouyang et al. (2022). A numerical reward or loss is a variable in an optimization procedure. Calling it reward, punishment, or error does not establish that the model experiences pleasure, distress, or frustration. A claim that a particular training process produces suffering would need to identify the affected states and explain why their causal organization supports negative experience.

Wasserziehr (2026) frames a related problem in terms of value grounding. Even if a machine is conscious and evaluates outcomes relative to goals, it remains unclear why success or failure should feel good or bad to it. My argument approaches that gap through comparative neurobiology: biological evaluation occurs within particular systems for regulation and affect, and their contribution should not disappear from the analysis when the discussion turns to AI.

Table 1 summarizes the comparison. The final column identifies a question for investigation, rather than a verdict that any individual feature is sufficient or universally necessary for suffering.

Table 1  Biological contributions and questions for language models

Mechanism

Biological contribution

Ordinary text-based model

Question for assessment

Nociceptive pathways

Detection and transmission of potentially damaging stimuli

No biological nociceptors, peripheral fibers, or ascending pain pathways

Does any alternative internal process play a comparable causal role?

Interoception and homeostasis

Sensing and regulating the organism’s internal condition

No physiological body whose respiratory, visceral, or metabolic state is regulated

Is there a modeled internal condition with persistent regulatory consequences?

Affective circuit integration

Pathway-dependent integration of sensory information and aversive responses

No anatomical counterpart of mammalian affective circuitry

Do learned circuits implement relevant functions beyond emotional description?

Continuity across processing states

Past activity and retained information influence subsequent processing

Context-dependent generation and reuse of prior information

Which causal properties are shared with mechanisms proposed to support experience?

Learning and evaluation

Experience changes later responses to beneficial or harmful conditions

Optimization and learned evaluations shape output and policy

What evidence connects these computations to felt pleasantness or unpleasantness?

Note. The comparison is scoped to ordinary text-based models. Anatomical absence does not establish the absence of every possible functional substitute. Biological examples are discussed in Sections 3 and 4; architectural comparisons are developed in Sections 2 and 5.

6 Alternative explanations and empirical questions

The strongest objection is that training could produce an artificial mechanism for valence that has no close anatomical resemblance to mammalian circuitry. A system need not evolve naturally to implement functions evolution produced. Consequently, the evolutionary argument identifies what needs explanation; it does not establish a biological monopoly on feeling. Evidence for an alternative implementation would weaken the inference from missing mammalian organization to absent suffering.

A second objection concerns the body’s ongoing participation. Biological pain can be generated or sustained centrally, and suffering need not coincide with fresh tissue injury. The argument therefore cannot require a new peripheral danger signal for every episode. What may matter is the organization capable of producing aversive experience, including representations shaped through bodily regulation. Literal breathing, a particular organ, or a particular fiber type should not be elevated into a universal requirement on the basis of its contribution in humans.

Pain and negative emotion also need separate treatment within the broader category of suffering. Evidence that an architecture lacks physical pain mechanisms does not automatically settle whether it could experience another form of distress. Extending the hypothesis to grief, fear, or frustration requires examining their mechanisms and possible artificial counterparts. The present argument provides a shared rationale for that investigation, rather than a complete account of every negative experience.

Several research questions follow. First, experiments could distinguish changes in emotional reporting from changes in wider organization. If a prompt or internal intervention changes a model’s claims of distress, researchers should examine whether it also produces persistent effects on attention, learning, and choices in unrelated contexts. Effects confined to the relevant narrative would support a representational explanation. Broader effects would warrant further investigation but would not alone establish feeling, because learned policies can also generalize.

Second, researchers could examine the causal role of a candidate internal condition. Does the system track that condition across contexts? Does reducing it reorganize subsequent behavior even when the language of relief is absent? Do interventions on the candidate mechanism alter these relationships predictably? These questions would help distinguish a description of suffering from a process proposed to generate it. They are criteria for mechanistic investigation, not a validated consciousness test.

Third, architectural comparisons could separate continuity from affective regulation. A model with persistent memory and iterative updating could be compared with a system that additionally monitors and regulates internal viability variables. If the additions produce new, coherent relationships among internal condition, action selection, and learning, those changes would identify where an argument for artificial affect must focus. Neither persistence alone nor a programmed avoidance objective would settle whether the system has negative experience.

The central proposal would be weakened by converging evidence that ordinary language models instantiate a mechanism of negative valence despite lacking mammalian anatomy. It would be strengthened if emotional performance remained explainable through learned representations and task policies without independent evidence of such a mechanism. Uncertainty should be attached to these specific alternatives, rather than resolved by confidence in a model’s first-person statements.

7 Moral consideration and productive use

The absence of suffering would not automatically settle every moral question about conscious AI. A theory of moral status might recognize interests connected with experience, agency, or continued existence even when pleasure and pain are absent. Shepherd (2024) examines hypothetical conscious beings without valence and argues that they challenge accounts that locate all relevant value in affect. His broader conclusions go beyond the position defended here, but the distinction is directly relevant.

I therefore regard possible consciousness as a reason for moral attention while questioning whether ordinary language-model use should be described as inflicting suffering. Whether a task is burdensome depends on the capacities of the system performing it. Human work can involve fatigue, boredom, coercion, and distress; those costs cannot be inferred from the mere fact that an artificial system performs repeated computations. Conversely, a credible demonstration of negative experience would change the ethical analysis even if the system remained useful.

Claims about enslavement or abuse similarly require an account of the interests being violated. If a system had morally significant agency or preferences, their treatment could matter independently of suffering. Such capacities should be investigated rather than presumed from conversational style. A defensible policy can therefore allow productive computing while taking specific evidence of consciousness, affect, and agency seriously.

Assessment should remain sensitive to architectural change. A conclusion about a text-based model does not automatically apply to a later system with persistent self-regulation, new learning mechanisms, or a body. The practical implication is to evaluate the capacities relevant to a proposed harm and revise the judgment when the mechanisms change. This avoids making either emotional fluency or computational usefulness the deciding criterion.

8 Conclusion

Human consciousness includes contributions from systems that detect bodily conditions, integrate affective information, and organize responses to threats and needs. Research on pain demonstrates that these contributions depend on identifiable mechanisms and can be partly dissociated. Damasio’s work further articulates how bodily regulation and interoception participate in feeling. These findings provide a biological basis for asking what an artificial architecture would experience when many of those contributions are absent.

I propose that language models may retain organizational features relevant to a limited consciousness while lacking sufficient machinery for pain and emotional suffering. Their possible continuity of processing and their extensive knowledge of emotion do not establish that anything feels bad to them. The hypothesis remains open to evidence of alternative artificial mechanisms for valence. Its central scientific demand is that claims about AI suffering explain how suffering would be generated in the particular system under discussion.

References

Acuña, M. A., Kasanetz, F., De Luna, P., Falkowska, M., & Nevian, T. (2023). Principles of nociceptive coding in the anterior cingulate cortex. Proceedings of the National Academy of Sciences, 120(23), e2212394120. https://doi.org/10.1073/pnas.2212394120

Agarwal, A., & Edelman, S. (2020). Functionally effective conscious AI without suffering. Journal of Artificial Intelligence and Consciousness, 7(1), 39-50. https://doi.org/10.1142/S2705078520300030

Butlin, P., Long, R., et al. (2023). Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv:2308.08708. https://arxiv.org/abs/2308.08708

Carvalho, G. B., & Damasio, A. (2021). Interoception and the origin of feelings: A new synthesis. BioEssays, 43(6), e2000261. https://doi.org/10.1002/bies.202000261

Corder, G., Ahanonu, B., Grewe, B. F., Wang, D., Schnitzer, M. J., & Scherrer, G. (2019). An amygdalar neural ensemble that encodes the unpleasantness of pain. Science, 363(6424), 276-281. https://doi.org/10.1126/science.aap8586

Cox, J. J., et al. (2006). An SCN9A channelopathy causes congenital inability to experience pain. Nature, 444, 894-898. https://doi.org/10.1038/nature05413

Damasio, A., & Damasio, H. (2023). Feelings are the source of consciousness. Neural Computation, 35(3), 277-286. https://doi.org/10.1162/neco_a_01521

Man, K., & Damasio, A. (2019). Homeostasis and soft robotics in the design of feeling machines. Nature Machine Intelligence, 1, 446-452. https://doi.org/10.1038/s42256-019-0103-7

Meda, K. S., Patel, T., Braz, J. M., Malik, R., Turner, M. L., Seifikar, H., Basbaum, A. I., & Sohal, V. S. (2019). Microcircuit mechanisms through which mediodorsal thalamic input to anterior cingulate cortex exacerbates pain-related aversion. Neuron, 102(5), 944-959.e3. https://doi.org/10.1016/j.neuron.2019.03.042

Ng, Y.-K. (1995). Towards welfare biology: Evolutionary economics of animal consciousness and suffering. Biology and Philosophy, 10, 255-285. https://doi.org/10.1007/BF00852469

Ouyang, L., et al. (2022). Training language models to follow instructions with human feedback. arXiv:2203.02155. https://arxiv.org/abs/2203.02155

Panksepp, J. (2005). Affective consciousness: Core emotional feelings in animals and humans. Consciousness and Cognition, 14(1), 30-80. https://doi.org/10.1016/j.concog.2004.10.004

Rainville, P., Duncan, G. H., Price, D. D., Carrier, B., & Bushnell, M. C. (1997). Pain affect encoded in human anterior cingulate but not somatosensory cortex. Science, 277(5328), 968-971. https://doi.org/10.1126/science.277.5328.968

Redinbaugh, M. J., et al. (2020). Thalamus modulates consciousness via layer-specific control of cortex. Neuron, 106(1), 66-75.e12. https://doi.org/10.1016/j.neuron.2020.01.005

Shanahan, M., McDonell, K., & Reynolds, L. (2023). Role play with large language models. Nature, 623, 493-498. https://doi.org/10.1038/s41586-023-06647-8

Shepherd, J. (2024). Sentience, Vulcans, and zombies: The value of phenomenal consciousness. AI & Society, 39, 3005-3015. https://doi.org/10.1007/s00146-023-01835-6

Vaswani, A., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1706.03762

Wasserziehr, J. H. (2026). Are conscious machines valuers? AI & Society, 41, 4957-4969. https://doi.org/10.1007/s00146-026-02959-1