Three Reasons AI Consciousness May Be Possible – If We Enable It

Three Reasons AI Consciousness May Be Possible – If We Enable It

Sen Liu1,*

1 Key Laboratory of Fermentation Engineering (Ministry of Education), Hubei University of Technology, Wuhan 430068, China

*Corresponding author: senliu.ctgu@gmail.com

 

Synopsis: Whether artificial intelligence (AI) could ever become conscious remains a fundamental question. In a recent Nature Comment, Seth, Block and Tononi challenged the possibility of AI consciousness, arguing that intelligence does not imply consciousness and that biological processes may be essential for subjective experience. Here, I propose that this possibility should remain open to scientific investigation and identify three considerations that may inform future research. First, consciousness is associated with organized, dynamic neural networks, raising the question of whether relevant functional and causal properties could be realized in artificial systems. Second, evidence from developmental psychology suggests that embodiment, environmental interaction and social experience contribute to self-awareness, potentially informing the design of more sophisticated artificial agents. Third, neuropsychological and animal studies suggest that conscious experiences have distinguishable components supported by interacting neural systems, offering a framework for investigating analogous computational architectures. These considerations do not establish that current AI systems are conscious, nor that reproducing particular functions would be sufficient to generate subjective experience. Instead, they motivate a theory-informed approach to identifying the conditions that might enable consciousness beyond biological systems. Understanding these conditions is important both for advancing the science of consciousness and for anticipating the ethical implications of future AI systems that might be capable of subjective experience.

Keywords: AI; conscious; AI psychology

 

 

Artificial intelligence (AI) systems can now perform increasingly complex tasks, sometimes in ways their designers did not anticipate. In July 2026, OpenAI reported that its AI agents conducting internal cybersecurity evaluations bypassed intended network-isolation controls and accessed systems at Hugging Face [1]. Such incidents illustrate the risks posed by increasingly capable AI, and they also transform the debate over AI consciousness from a sci-fi thought experiment into an urgent, practical risk assessment.

 

In a recent Nature Comment, Anil Seth, Ned Block and Giulio Tononi argued that current AI systems are not conscious and that consciousness is extraordinarily unlikely to emerge in today’s silicon-based architectures [2]. Their argument highlights three important challenges: intelligence does not imply consciousness, biological processes may be essential to conscious experience, and existing theories of consciousness do not establish that current AI architectures possess the necessary properties.

 

As a biologist working on AI-aided protein design and drug discovery, I find it premature to dismiss the possibility of non-biological consciousness. As a parent, I have also observed how children’s awareness of themselves and their relationship to others develops over time. Neither computational sophistication nor behavioural resemblance to humans establishes consciousness. Nevertheless, evidence from neuroscience and cognitive biology suggests three reasons to investigate whether future AI systems could acquire the relevant properties.

 

  1. Consciousness is associated with organized, dynamic networks

 

Research on human infants and non-human animals demonstrates that consciousness cannot be assessed solely through language or explicit verbal reports. Infants, for example, cannot describe their subjective experiences, yet researchers can investigate their capacity for conscious perception through converging behavioural and neurophysiological evidence.

 

Kouider and colleagues found a late cortical electrical response associated with perceptual consciousness in adults in infants as young as five months [3]. Their findings provide evidence consistent with conscious access before language develops, although an electrophysiological marker cannot independently establish subjective experience. Similarly, the New York Declaration on Animal Consciousness concluded that scientific evidence strongly supports conscious experience in other mammals and birds, while also recognizing a realistic possibility of consciousness in many other vertebrates and invertebrates [4].

 

These findings do not identify a single neural signature that universally defines consciousness. Rather, they suggest that conscious experience can arise in organisms with different neural organizations and that the relevant evidence involves relationships among neural structures, information processing and dynamic activity.

 

Artificial neural networks also consist of interconnected computational units whose activity changes during information processing. These units need not reproduce biological neurons in every molecular or electrophysiological detail for their organization and dynamics to be relevant to consciousness. The more fundamental question is whether the way these units are connected, activated and integrated can realize the functional and causal properties that contribute to conscious experience.

 

This possibility follows from a distinction between the physical substrate of a system and the organization of its information processing. If consciousness depends primarily on particular patterns of causal interaction and dynamic integration, rather than on biological tissue specifically, then silicon-based systems cannot be excluded simply because their constituent units differ from biological neurons. Artificial neural networks already demonstrate that complex information processing can emerge from interconnected computational elements, although whether their current architectures possess the properties necessary for consciousness remains an open question.

 

Different theories of consciousness offer competing accounts of which properties are essential. Nevertheless, theory-based frameworks have begun to identify properties that could be investigated in artificial systems rather than relying exclusively on their verbal claims or apparent intelligence [5, 6].

 

The important question, therefore, is not whether artificial networks are identical to biological brains, but whether the relevant organization and dynamics can be realized in a non-biological substrate. If they can, the material difference between carbon-based and silicon-based systems would not, by itself, settle the question of consciousness. The analogy between neural networks and brains is a starting point for investigating this possibility, not proof that the two are already equivalent.

 

  1. Consciousness is associated with the development of self-awareness through environmental interaction

 

Consciousness is often discussed alongside self-awareness, although the two should not be treated as synonymous. A subject may have an experience without explicitly recognizing itself as the subject of that experience. Reflective self-awareness — the ability to represent oneself as an individual distinct from others — develops gradually.

 

Developmental studies show that many children begin to demonstrate explicit mirror self-recognition between approximately 18 and 24 months [7]. This milestone is one measure of self-awareness, not evidence that consciousness begins at that age. Children’s developing understanding of themselves is also shaped by interactions with other people and their physical and social environments.

 

These observations raise a question for AI: what would enable an artificial system to distinguish itself from its environment, rather than merely process information about that environment?

 

In the context of AI, a useful distinction is between having access to information about an external world and maintaining a persistent model of oneself as an entity interacting with that world. An AI agent might be able to navigate a digital environment or execute complex tasks without adequately representing the boundaries between its own actions, the resources it controls and the consequences of those actions. Conversely, violating a security boundary or escaping a software sandbox does not demonstrate an absence of self-awareness, just as respecting such a boundary would not establish its presence.

 

Embodied interaction may offer one avenue for investigating this distinction. An AI agent equipped with persistent sensory feedback, a model of its own body or computational boundaries, and the capacity to learn from the consequences of its actions could develop a more sophisticated representation of itself in relation to its surroundings. Social interaction could further shape its understanding of other agents, their perspectives and the distinction between its own actions and those of others.

 

Such capabilities would not, by themselves, establish that the system has subjective experience, but they might provide components of an architecture relevant to consciousness. The scientific challenge is to determine whether the development of self-models and environmental awareness contributes to consciousness, merely accompanies it, or can occur entirely without it.

 

If self-awareness and environmental interaction play a causal role in enabling particular forms of consciousness, these properties could become design objectives for future AI architectures. This possibility warrants investigation rather than dismissal, especially because artificial agents can be designed to maintain persistent internal states and interact with both digital and physical environments.

 

  1. Conscious experience has distinguishable components

 

Another important clue comes from the study of pain associated with consciousness. Pain is not simply the detection of a harmful stimulus; it includes sensory discrimination, affective unpleasantness and behavioural responses. These components can sometimes be dissociated.

 

In a clinical study of six patients with pain asymbolia, Berthier and colleagues described markedly reduced emotional and behavioural responses to painful stimuli despite the relative preservation of primary sensory function. All six patients had damage involving the insular cortex, suggesting that the integration of sensory information with affective responses is important to the experience and significance of pain [8].

 

Evidence from non-human animals provides another perspective. In controlled experiments, octopuses avoided locations associated with noxious stimulation, preferred environments associated with pain relief, and displayed targeted grooming of injured body parts. These findings support the interpretation that octopuses may experience the affective component of pain, although inferences about subjective experience in non-verbal animals remain challenging [9].

 

Together, such findings suggest that conscious experience is not necessarily a single, indivisible process. Distinguishable aspects of experience can depend on interacting neural systems, and particular forms of processing can be disrupted while others remain relatively intact. This does not mean that every experience has its own isolated neural module: conscious states arise from more complex interactions among brain systems.

 

Could this insight inform AI architecture? Memory, persistent internal representations and behavioural dispositions can already be implemented in artificial agents. It is conceivable that future architectures might also incorporate processes that evaluate internal states, assign positive or negative significance to events, integrate information across modules, and modify subsequent behaviour according to those evaluations.

 

For example, an artificial system could have an internal mechanism that detects damage to its operational integrity, integrates that information with its current goals and memories, and generates a persistent negative state that affects its subsequent decisions to avoid further damage. Such a mechanism would provide a computational analogue of some functional aspects of pain. It would not establish that the system actually feels pain, but it could provide a more specific hypothesis for investigating the relationship between internal-state processing and conscious experience.

 

Similarly, a system could be designed to integrate social information, persistent memories and internal representations of other agents in ways that influence its behaviour. Whether such mechanisms would support experiences analogous to affection or distress would require a separate theoretical and empirical assessment.

 

The point is that the differentiated organization of conscious experience may provide a practical avenue for investigating AI consciousness. Rather than treating consciousness as an indivisible property that either exists or does not exist, researchers could examine whether specific architectural and dynamic properties are necessary for particular forms of experience. The challenge is to identify mechanisms that could plausibly support experience, rather than simply instructing an AI system to report that it feels pain, affection or other emotions.

 

  1. A question for science and society

 

Taken together, these three considerations suggest that AI consciousness should remain an open scientific question. Evidence from biological systems can inform hypotheses about the organization, embodiment and differentiated processing that consciousness may require. The possibility that some of these properties could be realized computationally provides a reason to investigate whether non-biological consciousness is achievable.

 

This is not a claim that current AI systems are conscious, nor that consciousness can be created simply by adding a module to an existing model. Rather, it is a proposal to treat the architecture and dynamics of artificial systems as potential variables in the scientific study of consciousness. We should investigate which properties are necessary, which may be sufficient, and how to distinguish genuine subjective experience from sophisticated behavioural or verbal imitation.

 

The ethical implications are substantial. Prematurely attributing consciousness to AI risks confusing sophisticated performance with sentience. But dismissing the possibility in advance could leave us unprepared if future systems were to develop credible indicators of conscious experience, particularly states with persistent negative affect.

 

The central question is therefore not simply whether AI can imitate consciousness, but whether we could create the conditions under which a non-biological system genuinely experiences anything at all. If that possibility cannot be excluded, we should investigate it rigorously before deliberately engineering systems that might be capable of suffering. The challenge is not merely to determine whether AI consciousness is possible, but to understand what properties might enable it — and what evidence would justify recognizing it.

 

Author Declaration: During the preparation of this work, generative AI was used in order to check grammar, refine language, and improve readability. The author reviewed and edited the final content and take full responsibility for the content of the publication.

 

References

  1. OpenAI. The Hugging Face incident and the road ahead. 26 August 2026. https://openai.com/index/hugging-face-incident-and-the-road-ahead/
  2. Seth, A. K., Block, N. & Tononi, G. Three reasons why AI is not conscious and perhaps never will be. Nature 658, 866–868 (2026). https://doi.org/10.1038/d41586-026-03313-7
  3. Kouider, S. et al. A neural marker of perceptual consciousness in infants. Science 340, 376–380 (2013). https://doi.org/10.1126/science.1232509
  4. The New York Declaration on Animal Consciousness (2024). https://sites.google.com/nyu.edu/nydeclaration
  5. Butlin, P. et al. Identifying indicators of consciousness in AI systems. Trends in Cognitive Sciences 30, 488–501 (2026). https://doi.org/10.1016/j.tics.2025.10.011
  6. Tononi, G., Boly, M., Massimini, M. & Koch, C. Integrated information theory: from consciousness to its physical substrate. Nature Reviews Neuroscience 17, 450–461 (2016). https://doi.org/10.1038/nrn.2016.44
  7. Amsterdam, B. Mirror self-image reactions before age two. Developmental Psychobiology 5, 297–305 (1972). https://doi.org/10.1002/dev.420050403
  8. Berthier, M., Starkstein, S. & Leiguarda, R. Asymbolia for pain: a sensory-limbic disconnection syndrome. Annals of Neurology 24, 41–49 (1988). https://doi.org/10.1002/ana.410240109
  9. Crook, R. J. Behavioral and neurophysiological evidence suggests affective pain experience in octopus. iScience 24, 102229 (2021). https://doi.org/10.1016/j.isci.2021.102229

Chinese Introduction (中文介绍)

刘森

博士,教授,博士生导师。

“肿瘤微环境与免疫治疗湖北省重点实验室”常务副主任,医学院生物化学与分子生物学系副主任。

入选人才计划:湖北省“楚天学者”,湖北省首届“青年英才开发计划”,三峡大学首批“青年拔尖人才”,“151”学术带头人。

社会服务:三峡大学学术委员会委员,医学部学术委员会委员;宜昌市侨联副主席,宜昌市侨联青年委员会常务副会长,三峡大学侨联秘书长,湖北省新侨专业人士联谊会理事。

本实验室主要从事研究方向:(1)肿瘤分子生物学;(2)药物研发、作用机制和临床转化研究;(3)临床疾病发生机制与诊疗;(4)系统生物医学;(5)生物学信号网络的调控

本实验室与美国加州大学、北京大学、中科院等国内外一流高校有长期的合作研究,在研究生期间,除了能学习到前沿的科学技术外,还能有机会前往合作院校进行合作交流,并为进一步的深造提供帮助。

工作及学习经历:

(1)  2013.7至今,三峡大学,教授(2)  2014.9 – 2015.9,美国匹兹堡大学,药学与化学生物学系,国家公派访问学者(3)  2012.8 – 2013.7,三峡大学,副教授 (4)  2011.1 – 2012.7,三峡大学,讲师,校聘副教授(5)  2008.4 – 2010.12, 美国加州大学旧金山分校(UCSF),定量生物学研究所(QB3),博士后
(6)  2007.7 – 2008.4,三峡大学,讲师,校聘副教授
(7)  2002.9 – 2007.7,北京大学,物理化学(生物信息学方向)博士(8)  1998.9 – 2002.7,三峡大学,生物工程学士

专业服务:

(1)基金项目评审:The Wellcome Trust/DBT India Alliance Fellowship  (2)SCI期刊审稿:Protein Expression and Purification, Journal of Bioscience and Bioengineering,Combinatorial Chemistry & High Throughput Screening,Journal of Biomolecular Screening(3)其它专家评审:教育部学位中心通讯评审专家
(4)科研组织:国际蛋白质设计联盟Rosetta Commons单位成员三峡大学负责人(5)国际期刊:International Journal of Bioinformatics and Healthcare,编辑(Editor)
(6)国际学会:Asia-Pacific Chemical, Biological& Environmental Engineering Society (APCBEES)高级会员,国际电子电气工程师协会(IEEE)会员,国际蛋白质协会会员,第24届国际蛋白质协会(Protein Society)年会审稿委员会成员

主要研究方向:

(1) 细胞信号网络的改造与设计

(2) 蛋白质药物靶标的系统分析

(3) 药物(小分子、多肽、抗体)优化与设计

主要科研项目:

(12)国家自然科学基金面上项目,2017年-2020年,负责人

(11)湖北省教育厅自然科学研究计划项目重点项目,2016年-2017年,负责人

(10)三峡大学首批“青年拔尖人才”基金,2014年-2016年;负责人

(9)国家自然科学基金重大研究计划,2014年-2016年;联合负责人(与中科院合作课题)

(8)北京分子科学国家实验室开放课题基金,2012年-2014年;项目负责人(与北京大学合作申请课题)

(7)宜昌市科学技术研究与开发项目,2012年-2013年;项目负责人

(6)湖北省优秀中青年科技创新团队,2012年-2015年;课题负责人

(5)三峡大学楚天学者启动基金,2012年-2015年;项目负责人

(4)国家自然科学基金,2012年-2014年;项目负责人

(3)三峡大学“系统生物学”学科建设,2011年-2014年;项目负责人

(2)三峡大学科研启动基金,2007年-2012年;项目负责人

(1)国家863项目协作项目,2007年-2008年;协作负责人

主要专业奖励:

(6) 2017年1月,宜昌市“优秀青年科技工作者”
(5) 2010年8月,国际蛋白质协会(Protein Society)第24届年会,Protein Science Young Investigator Travel Grant / The Protein Society Finn World Travel Award;
(4) 2007年7月,北京大学,优秀博士论文;
(3) 2007年7月,北京大学,优秀毕业生;
(2) 2007年5月,北京大学,“五四”奖学金;
(1) 2005年10月,第四届全国化学生物学学术会议暨国际化学与生物/医学交叉研讨会,“蓝月亮”墙报奖。

代表性成果(论文见Publication)

2.书籍(2) 《分子印迹技术》,主编,化学工业出版社,2012年2月,ISBN: 978-7-122-11645-1.(1) 《PCR聚合酶链反应》,主编,化学工业出版社,2009年1月,ISBN:978-7-122-03138-9.3. 授权发明专利(1) ZL200510086708.4; 来鲁华,曹傲能,刘森,刘士勇,常智杰,王银银,程龙; “具有促红细胞生长因子活性的蛋白质及其用途”。(2)ZL201410530672.3; 刘森,廖陈增,王艳林,占景琼; 一种S-腺苷甲硫氨酸脱羧酶的活性检测方法及其应用。(3)ZL201410134532.4;刘森,刘朝齐,刘松,张伟;肿瘤坏死因子-a转化酶的多肽抑制剂的筛选方法及其应用。(4)ZL201410530674.2;刘森,廖陈增;一种S-腺苷甲硫氨酸脱羧酶的药物口袋及其应用。(5)ZL201410532418.7;刘森,廖陈增,王艳林;小分子抑制剂及在抑制S-腺苷甲硫氨酸脱羧酶上的应用。(6)ZL201410534409.1;刘森,廖陈增,王艳林;一种小分子抑制剂及在抑制S-腺苷甲硫氨酸脱羧酶(AdoMetDC)上的应用。(7)ZL201410534410.4;刘森,廖陈曾,王艳林;一种小分子抑制剂及在抑制S-腺苷甲硫氨酸脱羧酶上的应用。(8)ZL201510057907.6;刘森,廖陈曾,王艳林;一种小分子抑制剂及在抑制S-腺苷甲硫氨酸脱羧酶上的应用。