Is it possible to identify phenomenal consciousness in artificial systems in the light of the gaming problem?

Is it possible to identify phenomenal consciousness in artificial systems in the light of the gaming problem?1 Michele Fariscoa,b,*, Kathinka Eversa, aCentre for Research Ethics and Bioethics, Uppsala University, Uppsala, Sweden bBiogem Molecular Biology and Genetics Research Institute, Ariano Irpino (AV), Italy *Correspondence: Michele Farisco ([email protected]). Abstract We analyze the question how phenomenal consciousness (if any) might be identified in artificial systems with specific reference to the gaming problem (i.e., the fact that the artificial system is trained with human-generated data, so that possible behavioral and/or functional evidence of consciousness is not reliable). Our goal is to review selected illustrative approaches for advancing in this direction. We highlight strengths and shortcomings of each approach, finally proposing a combination of different strategies as a promising task to pursue. 1 The manuscript has been submitted to the special issue “Evaluating Artificial Consciousness” (in Philosophy and the Mind Sciences). 2 Introduction Among the issues arising from recent development of Artificial Intelligence (AI), the prospect of artificial forms of consciousness emerges as particularly challenging, raising significant expectations (either positive or negative) from both specialists in the field and the general public (Lenharo, 2024, Colombatto and Fleming, 2023). One of the main challenges raised by a hypothetical AI consciousness is how to identify it. In fact, the classical other minds problem, which essentially makes any attribution of consciousness inferential, is even more tricky for candidate conscious AI systems in virtue of the so-called gaming problem. The gaming problem arises in connection to the hypothesis of artificial consciousness, specifically from the hypothesis of artificial phenomenal consciousness or sentience, because artificial systems are instructed with human-generated data in order to give them the appearance of instantiating human features, wherefore functional or behavioral markers of sentience are unreliable and cannot be considered as evidence of actual phenomenal consciousness (Birch and Andrews, 2024). Chang-Eop has recently proposed a logical argument in support of the unreliability of behavioral manifestations of conscious abilities by artificial systems, specifically by Large Language Models (LLM), concluding that in the case of LLM consciousness denial leads to contradiction while consciousness affirmation leads to indeterminacy (Chang-Eop, 2024). The gaming problem has both theoretical and ethical relevance, since it may be argued that artificial systems should not confuse human users about their conscious state (Schwitzgebel, 2023a). Therefore, it is particularly urgent to elaborate possible strategies for managing and prospectively solving the gaming problem. In the following we review five illustrative approaches that have been proposed by experts in the field for identifying conscious abilities in AI systems. We provide a description of these strategies for eventually highlighting that a combination among them is probably the best choice for advancing towards a more reliable attribution of conscious capacities to AI systems. We start from a definition of phenomenal consciousness as the subjective feeling of a particular experience, or “what it is like to be” in a particular state (Block, 1995, Block, 2022), and as “sentience”, or the capacity for having valenced experiences, like pain or pleasure (Birch, 2024, Butlin and Lappas, 2025). Then we describe five strategies that have been articulated for detecting hypothetical artificial forms of phenomenal consciousness: 1. A theory-based strategy that starts from selected theories of consciousness to infer relevant indicators (Butlin et al., 2023) 2. A life-based strategy which outlines the necessary connection of consciousness with biological life (Seth, 2024); 3. A brainbased strategy that takes the brain, its evolution, and its correlation with consciousness as a benchmark for artificial consciousness (Farisco et al., 2024b, Aru et al., 2023); 4. A consciousness-based strategy that searches for other forms of biological consciousness than the 3 human, in order to identify what (if anything) is really indispensable to consciousness and what is dispensable, and thus overcoming the controversy between the many theories of consciousness and proceed towards the identification of reliable evidence of artificial consciousness (Birch and Andrews, 2024); 5. A indicators-based approach that elaborates a list of indicators, conceived as features that we tend to agree characterize conscious experience and that are indicative (i.e., probabilistic rather than definitive evidence) of the presence of consciousness in artificial systems (Pennartz et al., 2019), and on their basis elaborate relevant tests for artificial consciousness (Bayne et al., 2024, Elamrani and Yampolskly, 2019). 1. Phenomenal consciousness Phenomenal consciousness may be defined as the subjective experience, “what it is like to be” in a particular state (i.e., including experiences of perceptions which are not cognitively accessed) (Block, 1995, Block, 2022). Another way to conceive phenomenal consciousness is in terms of “sentience”, which basically refers to the capacity for having valenced (or hedonic) experiences, like pain or pleasure (Birch, 2024). Valenced experience relates to affective states, either positive or negative. In fact, together with arousal, valence is a fundamental component of affect (Russell, 1980, Posner et al., 2005). As reviewed by Birch, there are different understandings of the nature of valence, including its description as immediate quality of the experience (i.e., raw feelings), as non-conceptual representation of value, and as imperative content (Birch, 2024). The concept of value has been recently explored in relation to hypothetical conscious AI systems (Farisco and Evers, In Press). The starting point of this analysis is the capacity for evaluation (i.e., sensitivity to reward signals and the ability to discriminate between good and bad things in the world on the basis of specific needs, motivations, and goals) as a defining feature of biological organisms, namely of the human brain. In fact, evaluation has been posited as a fundament of learning and memory (Dehaene and Changeux, 1989, Edelman, 1992), as well as of consciousness and ethical deliberation (Evers, 2009). A crucial characteristic that makes this kind of evaluation performed by biological organisms a conscious experience is the capacity for autonomously performing it on the basis of subjectively salient drives. Importantly, evaluation thus conceived, as well as the related concept of value, include both cognitive and emotional dimensions: cognitive representations like concepts, goals, and beliefs are combined with emotional attitudes having positive or negative valence (https://www.psychologytoday.com/intl/blog/hot-thought/201304/what-are-values). Emotions are particularly crucial for evaluating information coming from the inside of the organism (Bennett, 2023). The concepts of phenomenal consciousness and sentience play an important role in both theoretical and moral debates. There is a wide tendency to identify consciousness tout court with 4 phenomenal consciousness, at least in a morally relevant sense (Levy, 2014), or to affirm the necessity of phenomenal consciousness for any cognitive form of consciousness to be possible. With reference to the prospect of artificial consciousness, according to these perspectives, the central question is whether it may be possible to engineer an artificial form of subjective experience. 2. Levels of phenomenal consciousness Since subjective experience or the phenomenal character of a conscious experience may have different levels, and these levels may depend on some specific aspects or dimensions (e.g., duration, intensity, frequency, precision, etc.), we cannot logically exclude prima facie that a system, whether biological or artificial, may have at least very rudimentary, low-level forms of subjective experience. Accordingly, we must avoid identifying subjective experience exclusively with high-level, sophisticated forms (e.g., those usually attributed to human subjects), and avoid an anthropocentric bias in the attempt to identify it in non-human systems (e.g., in artificial systems). Accordingly, we should avoid setting the bar too high in defining the gold standard in our search for indicators of phenomenal consciousness in AI. Two low-level forms of consciousness are of particular interest to consider here: primary/minimal consciousness and contentless consciousness. Advocates of different models have proposed the concept of primary or minimal consciousness as opposed to secondary or more advanced consciousness within a graded view of consciousness. One proposal is to conceive minimal consciousness as subjective experience. More specifically, as the most basic (non-reflective) subjective feeling that includes exteroceptive (e.g., visual, olfactory), interoceptive (e.g., pain, hunger, thirst) and proprioceptive (bodily position) experiences (Bronfman et al., 2016, Merker, 2007). The key point of this proposal is that consciousness includes a more basic, non-cognitive form (Pereira Jr, 2021, Baluška and Reber, 2019, Baluška et al., 2025). The proponents of this view argue that the concept of consciousness as a fundamental capacity for experiencing that is distinct from high-level cognitive capacities is useful to identify what is necessary and sufficient for appropriately characterizing a (biological or artificial) system as (minimally) conscious. The concept of anoetic consciousness as introduced by Endel Tulving is very close to this view (Tulving, 2005). According to him, there are three kinds of consciousness: autonoetic, which is related to the knowledge of the self; noetic, which is related to the knowledge of the outside world; and anoetic, which is related to the absence of explicit current knowledge (LeDoux and Lau, 2020). Anoetic consciousness is conceived as the condition of being alive and responsive to stimuli as opposed to having explicit conscious contents (LeDoux, 2021), or as "a stream of prereflective affective and sensorial perceptual consciousness essential for the waking state of the 5 organism in the absence of an explicit self-referential awareness of associated cognitive contents " (Vandekerckhove et al., 2014) (p. 6). Others use “primary” or “sensory” (Edelman, 2003) (Feinberg and Mallatt, 2016, Edelman, 1989) consciousness to indicate a basic capacity to detect stimuli, to process their saliency and value, and to react accordingly. Importantly, primary/sensory consciousness as qualified, for instance, by Gerald Edelman can have access only to the present, actual experience. To illustrate, for Feinberg and Mallatt (Feinberg and Mallatt, 2018), the basic form of consciousness is “value-based” as distinguished from “image-based” consciousness. The valuebased consciousness does not rely on any kind of explicit, mental images of the world, but rather on an organic, and in some cases neuronal, map or schema that allows the organism to discriminate affordances (i.e., to detect and distinguish dangers and positive opportunities) in order to increase its fitness (cf. (Changeux, 1986, Changeux, 2004)). The distinction between basic and more sophisticated forms of consciousness has been proposed also with specific reference to phenomenal consciousness, for instance through the distinction between minimal phenomenal selfhood (i.e., “the experience of being a distinct, holistic entity capable of global self-control and attention, possessing a body and a location in space and time” (Blanke and Metzinger, 2009) (p.7)), and a more elaborated phenomenal self-experience (Blanke and Metzinger, 2009, Seth and Tsakiris, 2018). Contentless consciousness, also called pure consciousness or minimal phenomenal experience, is a form of conscious experience devoid of any specific content (Metzinger, 2020) that is related to the abovementioned distinction between more basic and more sophisticated forms of phenomenal consciousness. Notably, contentless consciousness is considered as either the grounding or the highest form of consciousness in some Eastern religions and spiritual traditions, as well as the ultimate goal in meditation (Thompson, 2015). There is discussion about whether this form of consciousness actually exists, and, if it does, what its main features might be and how to eventually operationalize it in order to elaborate criteria for identifying it (Sullivan, 1995, Woods et al., 2024). The possibility of indeterminate cases of sentience, i.e. systems for which there is no determinate fact of the matter whether they are conscious or not (see (Lee, 2020, Simon, 2017, Schwitzgebel, 2023b), adds further complexity to the question whether artificial phenomenal consciousness/sentience is possible and how to detect it. 3. How can phenomenal consciousness be operationalized and made detectable in artificial systems? Different approaches to the question how to detect phenomenal conscious activity in artificial systems have been suggested. Among them, we here focus on the following five: 6 1. Theory-based strategy: starting from selected theories of consciousness in order to infer relevant indicators (Butlin et al., 2023) 2. Life-based strategy: consciousness necessarily connects with biological life (Seth, 2024) 3. Brain-based strategy: the brain, its evolution, and its correlation with consciousness are benchmark for artificial consciousness (Farisco et al., 2024b, Aru et al., 2023) 4. Consciousness-based strategy: searching for other forms of biological consciousness than the human, in order to identify what (if anything) is really indispensable to consciousness and what is dispensable, and thus overcoming the controversy between the many theories of consciousness and proceed towards the identification of reliable evidence of artificial consciousness (Birch and Andrews, 2024) (https://aeon.co/ essays/to-understand-ai-sentience-first-understand-it-in-animals) 5. Indicators-based strategy: starting from the features of consciousness we tend to agree about for elaborating a list of indicators of consciousness in artificial systems (Pennartz et al., 2019), and related tests for artificial consciousness (Bayne et al., 2024, Elamrani and Yampolskly, 2019). 3.1 Theory-based strategy The first strategy we focus on consists in starting from selected theories of consciousness in order to infer indicators of consciousness from them. (Butlin et al., 2023) illustrates this kind of approach. In this paper, the authors propose a “rigorous and empirically grounded approach to AI consciousness”: they select some neuroscientific theories of consciousness and identify respective indicators of consciousness. They base their analysis on three main premises. First, they assume computational functionalism (i.e., performing computations of the right kind is necessary and sufficient for consciousness) as a working hypothesis; second, they claim that neuroscientific theories of consciousness can help us in reliably assessing AI consciousness because of the empirical support they have; and finally they argue that a theory-heavy approach (i.e., explicitly relying on empirically grounded theories of consciousness) is the best strategy for assessing AI consciousness. Basically, the theory-based strategy they argue for attributes consciousness to AI systems depending on how many indicators of consciousness inferred from the selected theories the systems in question show. The authors acknowledge that there are many scientific theories of consciousness, not all compatible and not always commensurable either (Seth and Bayne, 2022, Evers et al., 2024, Chis-Ciure et al., 2024, Northoff and Lamme, 2020, Kuhn, 2024). Consistently, they do not endorse any specific theory, but rather derive a list of indicators from the sampled theories, and consider the evidence deriving from them cumulative, meaning that the more indicators are present, the more likely it is that the system in question is conscious. 7 The theories they refer to are the following, with related indicators: - Recurrent processing theory, from which they derive two indicators: input modules using algorithmic recurrence and input modules generating organized, integrated perceptual representations. - Global workspace theory, from which they derive four indicators; multiple specialized systems capable of operating in parallel (modules); limited capacity workspace, entailing a bottleneck in information flow and a selective attention mechanism; global broadcast: availability of information in the workspace to all modules; state-dependent attention, giving rise to the capacity to use the workspace to query modules in succession to perform complex tasks. - Computational higher-order theories, from which they derive four indicators: generative, top-down or noisy perception modules; metacognitive monitoring distinguishing reliable perceptual representations from noise; agency guided by a general belief-formation and action selection system, and a strong disposition to update beliefs in accordance with the outputs of metacognitive monitoring; sparse and smooth coding generating a “quality space”. - Attention schema theory, from which they derive one indicator: a predictive model representing and enabling control over the current state of attention. - Predictive processing, from which they derive one indicator: input modules using predictive coding. In addition to these theories, the authors refer also to Agency and embodiment, from which they derive two indicators: agency, conceived as learning from feedback and selecting outputs so as to pursue goals, especially where this involves flexible responsiveness to competing goals; embodiment, conceived as modeling output-input contingencies, including some systematic effects, and using this model in perception or control. As mentioned above, the authors consider the evidence deriving from the identified indicators as cumulative. In fact, some AI systems already manifest some of the indicators, while there is currently no system capable of consistently implement all of them. Therefore, the authors conclude that there is no system yet that can be considered a strong candidate for consciousness according to a theory-based approach. The second strategy we describe consists in considering consciousness as a biological phenomenon which is a feature of life and dependent on it. Different authors propose theoretical interpretations that tend to endorse this view, including John Searle (Searle, 2007), Peter Godfrey-Smith (Godfrey-Smith, 2016, Godfrey-Smith, 2023), Evan Thompson (Thompson, 2018, Thompson, 2007), and Anil Seth (Seth, 2024, Seth, 2021), among others. Here we summarize the position of the latter, which is illustrative of the life-based strategy. 8 Anil Seth frames his proposal within biological naturalism, according to which consciousness is a property of only (even though not all) living systems. He acknowledges that this view has different variations, arguing that it is not necessary that life is carbon-based. In principle this leaves the door open to the possibility of a “living conscious AI”. Importantly, Seth clarifies what is the meaning of consciousness he refers to in his analysis: conscious experience, or “to feel like something”. Accordingly, consciousness is real, not an illusion, and there is a fact-of-the-matter about whether something is conscious. Basically, biological naturalism as defined above opposes to computational functionalism. In fact, for biological naturalism the biological basis of consciousness is not just an enabling factor for the right kind of computation, like, for instance, argued for in (Wiese, 2024), but rather a fundamental condition for consciousness. On the basis of theoretical models like predictive processing and the free energy principle, Seth argues for what he calls an “ontological continuity” between life and consciousness: living systems are autopoietic (i.e., they continually regenerate themselves), and this characteristic of the living matter is the basis for conscious activity. Interestingly, according to this view consciousness may be instantiated in very basic, rudimentary forms, a “ground state” of conscious experience characterized by “an inchoate, shapeless, formless feeling of simply ‘being alive’”. At the theoretical level, two forms of biological naturalism are possible: a weak biological naturalism, for which life matters to consciousness “in virtue of enabling (or being necessary for) specific patterns of functional organization”, and strong biological naturalism, for which consciousness fundamentally depends on life. For both these versions, conscious AI will not come along for the ride as conventional AI gets smarter, because, as said above, consciousness is conceived as a property of living systems, therefore AI needs to be alive for being conscious. Neuromorphic and synthetic technologies could potentially be promising in this respect, if we accept the hypothesis that the more AI is similar to the brain and living systems, the more likely it is that it may be conscious. The crucial point of the life-based strategy is that life, either carbon-based or of different nature, imposes a number of constraints and necessary conditions for consciousness to arise. For instance, Seth outlines the crucial role played by metabolic constraints, electromagnetic fields, and fine-grained timing relations (Seth, 2024). In short, there are several dependencies and contingencies accumulated in a “generative entrenchment” in the internal organization of the brain, which eventually constrain hypothetical alternative implementations of its mental features. The conclusion by Seth is that it may be impossible to separate what brains do from what they are: a failure to replicate the significant amount of functional and architectural details that appear to play an important role for consciousness would result in a change of the system’s organization, which likely affects the functions that the replicator system is able to perform. 3.2 Brain-based strategy The third strategy we describe can be qualified as brain-based because it assumes that the brain, its evolution, and its correlation with consciousness are the benchmark for developing artificial 9 consciousness. Examples of this approach may be found in (Aru et al., 2023) and (Farisco et al., 2024a). The latter paper analyses artificial consciousness from an evolutionary perspective, taking the evolution of the brain and its relation with consciousness as a reference model or benchmark. This analysis reveals a number of structural and functional characteristics of the brain that arguably play a key role for human-like conscious experience. Importantly, current AI systems lack these features, or at least they fail to instantiate a sufficient level of these features. The proposal of the authors is taking inspiration from the brain to advance towards developing conscious AI. More specifically, the following brain features are proposed to be of relevance in research on conscious AI : - Hierarchical, nested, and multiple levels of organization - Rewards sensitivity based on embodied multidimensional sensorimotor experience - Spontaneous physiological activity contributing to conscious information processing - The capacity to produce variability from the same functional organization and the long postnatal development resulting in multiple nested epigenetic synapse selections - Degeneracy, that is different connection patterns may have the same function, leading to plasticity, individual differences, and creativity - Capacity for interaction between individuals in a social group - A physical and operational distinction between conscious and non-conscious brain representations, leading to the development of semantic competence and singular properties of association of conscious representations. Starting from these features, the authors identify the following limitations of current AI that need to be ameliorated for advancing towards conscious AI: - Computational approach with insufficient brain data - Hardware with limited chemical elements - Hardware implementing parallel levels of computation opposed to the nested brain organization - Lack of an evolutionary and epigenetic development - Lack of a multidimensional and multisensory representation based on embodiment - Lack of social relationships and intrinsic ethical responsibility - Lack of differentiation of the singularity of the individual and their life experience. The hypothesis suggested is that the more an AI system implements the identified brain features correlated to consciousness, and the more an AI system improves the limitations listed above, the more likely it is that it may be or become conscious. Importantly, (Farisco et al., 2024a) limit their analysis to human-like consciousness, but they also acknowledge that in principle it is possible to have an AI consciousness alternative to the human model. While this possibility cannot logically be excluded, assuming human consciousness and correlated brain features as reference models or benchmarks is a pragmatic choice that makes it possible to formulate realistic hypotheses. 16 In conclusion, the consistency among the five identified strategies, and consequently the kind of relationship between the evidence for consciousness deriving from them (i.e., complementary vs cumulative) depend on which specific dimension of logical or empirical structure we consider. For instance, all the five strategies appear logically consistent in relation to their conceptual foundations (i.e., they define consciousness as phenomenal consciousness or sentience, or are compatible with such definition of consciousness), and empirically consistent in relation to their methodology (i.e., they share the necessity of empirical testing the identified indicators). This implies that in order to get a greater evidence of AI consciousness we should first clarify which specific logical and/or empirical dimensions we refer to and then eventually cumulate or complement the evidence deriving from the different strategies. Therefore, the best “metastrategy” for advancing towards a better and stronger evidence of AI consciousness is combining different specific strategies, but how to concretely achieve this goal is a matter of analytical reasoning, and it needs a clear identification of which structure (i.e., logical or empirical) and respective dimension to address. What we suggest to be necessary is to combine different kinds of indicators in order to develop tests for consciousness that go beyond the classical two approaches (i.e., architectural or behavioral (Elamrani and Yampolskly, 2019, Bayne et al., 2024)) and that include an ethological/ecological and diachronic (i.e., extended in time) observation for addressing the issue of artificial consciousness. 4.3 Relevance of the different strategies to identify different levels of consciousness As described above, consciousness, including phenomenal consciousness, may arguably exist at different levels, the main of which are primary or minimal consciousness, contentless consciousness, and reflective consciousness. In short, primary/minimal consciousness is the non reflective and non cognitive level of consciousness, or the fundamental capacity for experiencing. Contentless consciousness is the level of consciousness devoid of any specific content, like a raw feeling of existing. Reflective consciousness is the most sophisticated level of consciousness, characterizing a subject able to cognitively access their own experience. The five strategies presented above appear to have different relevance to these three levels of consciousness (See Table 3). All the strategies potentially provide information relevant to address reflective level of consciousness, while not all appear relevant to the other two levels. The theory-based strategy stands on the strong computational functionalist premise, which excludes non cognitive and contentless levels of consciousness. Bothe the life-based and the consciousness-based strategies allow the possibility of both primary/minimal and contentless consciousness, for instance in the form of the feeling of being alive. Both the brain-based and the indicators-based strategies appear to be compatible, at least in principle, with primary/minimal consciousness but not with contentless consciousness, because both these strategies take a representational view of consciousness (i.e., consciousness refers to something, either cognitively or non-cognitively). 17 Since phenomenal consciousness is arguably multilevel, it is important to take into account which specific levels the different strategies above are relevant to, in order to avoid excluding off-hand some levels of consciousness because the chosen strategy is not compatible with them. In conclusion, the combination of different strategies is the best way to prevent this risk. We propose that combining the different strategies described above is necessary for advancing towards a more reliable assessment of artificial consciousness for two reasons among others: first, consciousness is a multidimensional and multilevel feature, and combining different logical and empirical strategies increases the chances of covering this complexity; second, in order to address the gaming problem, it is crucial to look for as many indicators as possible, from structural to architectural, from functional to socio-ethological. Conclusion We have reviewed five strategies for operationalizing and identifying hypothetical phenomenal consciousness in AI systems. Each strategy has its own specific empirical and logical characterization, but they can be compared with each other in relation to different empirical and logical dimensions, resulting compatible or complementary. We propose to combine different strategies for elaborating more reliable indicators of consciousness in AI systems, either combining or complementing the evidence derived from compatible or complementary strategies, respectively. We argue that this combination is necessary for two reasons among others: the multidimensional and multilevel nature of consciousness, and the gaming problem. Funding This research has received funding from the project Counterfactual Assessment and Valuation for Awareness Architecture—CAVAA (European Commission, EIC 101071178). 18 Table 1 Approach Main tenets Advantages Shortcomings Theory-based  Inferring indicators of consciousness from selected scientific theories of consciousness  Extensive theoretical base relying on empirically validated theories  Conceptual clarity about the specific notion of consciousness which is the object of the indicators  Computational functionalism regarding consciousness is deeply problematic both theoretically and ethically  per force selective: the theories out there are too many. Some of them may lead to propose similar measures and indicators of consciousness, but many differ on both points which poses a problem in terms of commensurability  affected by the limitations of the selected theories  risk of biases (e.g., anthropocentric or bio-centric)  how to solve the indicators’ necessary/sufficient issue  how to compare different theories in order to derive a cumulative evidence (in fact different theories often target different forms or kinds of consciousness, so that the indicators derived from them are not really cumulative) 19  generalization problem, or the gold standard problem: theories of consciousness are usually formulated with reference to typical adult consciousness, so that the relevance of the indicators derived from them to address artificial consciousness is questionable. Life-based  Inferring indicators of consciousne ss from life features associated with consciousne ss  it relies on the fact that all known examples of conscious systems are biological, therefore it is not speculative  it may be interpreted as excluding offhand the possibility of forms of AI consciousness alternative to the biological forms Brain-based  Inferring indicators of consciousne ss from brain features associated with consciousne ss  it relies on empiricallygrounded data about the brain bases of consciousness, therefore it has a strong scientific base  it avoids speculations about hypothetical alternative forms of consciousness  it is pragmatic, in the sense that is prone to be operazionalized into specific approaches to test  it is limited to human-like forms of consciousness, so it may lead to overlook alternative forms of machine consciousness  the distance between current AI and the brain is quite big, so this approach may set the bar too high and eventually be not applicable 20 machine consciousness Consciousness-based  Comparing different kinds of consciousne ss to infer what is indispensabl e to it (and then check for it in AI)  in principle, it avoids anthropomorphic and anthropocentric traps: it accommodates evidence coming from different instances of consciousness  it eventually make reference to selected computational theories of consciousness, finally raising the risk of reproducing the same shortcomings of the theory-based approach  the search for computational markers of consciousness in AI systems is limited by the architecture of actual AI systems, which lack transparency and explicability  it risks of replacing a big challenge (identifying AI consciousness) with another big challenge (advancing a comparative understanding of different forms of consciousness in nature). Indicators-based Approach  Inferring indicators of consciousne ss from agreed-upon features of consciousne ss  it relies on what we tend to agree characterize conscious activity, trying to stay neutral with regards to specific theories of consciousness  in principle, it is compatible with different theoretical  it is conceived with reference to biological consciousness, and therefore its relevance and application to AI consciousness may be limited 21 accounts  it can accommodate the prismatic nature of consciousness, including its phenomenal and cognitive dimensions 22 Table 2 KIND OF STRUC TURE DIMENSION THEORYBASED LIFEBASED BRAINBASED CONSCIOU SNESSBASED INDICATORSBASED LOGIC AL CONCEPTU AL FOUNDATI ONS (Understanding and definition of key concepts) Phenomenal Consciousne ss Sentience Conscious process, compatible with Phenomen al Conscious ness Sentience Both Access and Phenomenal Consciousness X X X X X THEORETIC AL FRAMEWO RK & INTERPRET ATIVE APPROACH (The way of assigning relevance and meaning to data) Computation al functionalis m. Biological naturalism. (Neuro- )biological naturalism. Biological naturalism. Biological naturalism. X V V V V EMPIRI CAL EVIDENCE OF CONSCIOUS NESS (i.e., Type(s) of data (e.g., quantitative or behavioural) considered necessary and sufficient for inferring consciousness) System’s operations (what it does) + architecture (how it is internally organized) System’s nature (its inherent character) System’ nature + architectur e System’s operations (what it does) + architecture (how it is internally organized) System’s operations (what it does) + architecture (how it is internally organized) 23 X V V X X METHODOL OGY (i.e., General research strategy defining how the research process has to be undertaken) From theories to indicators, to be checked empirically (i.e., looking for thirdperson, empirically testable evidence) From life definition to indicators, to be checked empirically (i.e., looking for thirdperson, empirically testable evidence) From brain features to indicators, to be checked empirically (i.e., looking for thirdperson, empirically testable evidence) From different types of consciousness to indicators, to be checked empirically (i.e., looking for thirdperson, empirically testable evidence) From working definition of consciousness and its features to indicators, to be checked empirically (i.e., looking for third-person, empirically testable evidence) X X X X X Table 3 LEVEL OF CONSCIOUSNESS THEORYBASED LIFEBASED BRAINBASED CONSCIOUSNESSBASED INDICATORSBASED PRIMARY/MINIMAL NO YES YES YES YES CONTENTLESS NO YES NO YES NO REFLECTIVE YES YES YES YES YES 24 References ARU, J., LARKUM, M. E. & SHINE, J. M. 2023. The feasibility of artificial consciousness through the lens of neuroscience. Trends in Neurosciences, 46, 1008-1017. BALUŠKA, F., MILLER, W. B., SLIJEPCEVIC, P. & REBER, A. S. 2025. Sensing, feeling and sentience in unicellular organisms and living cells. BioSystems, 247, 105374. BALUŠKA, F. & REBER, A. 2019. Sentience and Consciousness in Single Cells: How the First Minds Emerged in Unicellular Species. Bioessays, 41, e1800229. BAYNE, T., SETH, A. K., MASSIMINI, M., SHEPHERD, J., CLEEREMANS, A., FLEMING, S. M., MALACH, R., MATTINGLEY, J. B., MENON, D. K., OWEN, A. M., PETERS, M. A. K., RAZI, A. & MUDRIK, L. 2024. Tests for consciousness in humans and beyond. Trends Cogn Sci. BENNETT, M. S. 2023. A brief history of intelligence : evolution, AI, and the five breakthroughs that made our brains, New York, Mariner Books. BIRCH, J. 2024. The Edge of Sentience : Risk and Precaution in Humans, Other Animals, and AI. BIRCH, J. & ANDREWS, K. 2024. To Understand AI Sentience, First Understand it in Animals. Intellectica, 81, 213-226. BLANKE, O. & METZINGER, T. 2009. Full-body illusions and minimal phenomenal selfhood. Trends Cogn Sci, 13, 7-13. BLOCK 1995. On a confusion about a function of consciousness. Behavioral and Brain Sciences, 18, 227-287. BLOCK, N. J. 2022. The border between seeing and thinking. Philosophy of mind series. New York, NY: Oxford University Press,. BRONFMAN, Z. Z., GINSBURG, S. & JABLONKA, E. 2016. The Transition to Minimal Consciousness through the Evolution of Associative Learning. Front Psychol, 7, 1954. BUTLIN, P. & LAPPAS, T. 2025. Principles for Responsible AI Consciousness Research. arXiv:2501.07290 [Online]. BUTLIN, P., LONG, R., ELMOZNINO, E., BENGIO, Y., BIRCH, J., CONSTANT, A., DEANE, G., FLEMING, S. M., FRITH, C., JI, X., KANAI, R., KLEIN, C., LINDSAY, G., MICHEL, M., MUDRIK, L., PETERS, M. A. K., SCHWITZGEBEL, E., SIMON, J. & VANRULLEN, R. 2023. Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. CAO, R. 2022. Multiple realizability and the spirit of functionalism. Synthese, 200, 506. CHANG-EOP, K. 2024. The Logical Impossibility of Consciousness Denial: A Formal Analysis of AI Self-Reports. Available: arXiv:2501.05454. CHANGEUX, J.-P. 1986. Neuronal man : the biology of mind / translated by Laurence Garey, New York ; Oxford, Oxford University Press. CHANGEUX, J.-P. 2004. The physiology of truth : neuroscience and human knowledge, Cambridge, Mass., Belknap Press of Harvard University Press. CHIS-CIURE, R., MELLONI, L. & NORTHOFF, G. 2024. A measure centrality index for systematic empirical comparison of consciousness theories. Neurosci Biobehav Rev, 161, 105670. COLOMBATTO, C. & FLEMING, S. M. 2023. Folk Psychological Attributions of Consciousness to Large Language Models. 25 DEHAENE, S. & CHANGEUX, J. P. 1989. A simple model of prefrontal cortex function in delayed-response tasks. J Cogn Neurosci, 1, 244-61. EDELMAN 1992. Bright air, brilliant fire : on the matter of the mind, New York, NY, BasicBooks. EDELMAN, G. M. 1989. The remembered present : a biological theory of consciousness, New York, Basic Books. EDELMAN, G. M. 2003. Naturalizing consciousness: a theoretical framework. Proc Natl Acad Sci U S A, 100, 5520-4. ELAMRANI, A. & YAMPOLSKLY, R. V. 2019. Reviewing Tests for Machine Consciousness. Journal of Consciousness Studies, 26, 35-64. EVERS 2009. Neuroetique. Quand la matière s'éveille, Paris, Odile Jacob. EVERS, K., FARISCO, M. & PENNARTZ, C. M. A. 2024. Assessing the commensurability of theories of consciousness: On the usefulness of common denominators in differentiating, integrating and testing hypotheses. Consciousness and Cognition, 119, 103668. FARISCO, M. & CHANGEUX, J. P. 2023. About the compatibility between the perturbational complexity index and the global neuronal workspace theory of consciousness. Neurosci Conscious, 2023, niad016. FARISCO, M. & EVERS, K. In Press. The capacity for evaluation of the human brain and its implications for an artificial moral subject. In: VIJAYARAGHAVAN, S. & FELSEN, G. (eds.) Neuroscience and Society. London: CRC Press - Taylor&Francis. FARISCO, M., EVERS, K. & CHANGEUX, J.-P. 2024a. Is artificial consciousness achievable? Lessons from the human brain. Neural Networks, 180, 106714. FARISCO, M., EVERS, K. & CHANGEUX, J. P. 2024b. Is artificial consciousness achievable? Lessons from the human brain. Neural Netw, 180, 106714. FEINBERG, T. E. & MALLATT, J. 2016. The nature of primary consciousness. A new synthesis. Conscious Cogn, 43, 113-27. FEINBERG, T. E. & MALLATT, J. 2018. Consciousness demystified, Cambridge, Massachusetts ; London, England, MIT Press. GINSBURG, S. & JABLONKA, E. 2019. The evolution of the sensitive soul : learning and the origins of consciousness, Cambridge, Massachusetts, The MIT Press. GODFREY-SMITH, P. 2016. Mind, Matter, and Metabolism. Journal of Philosophy, 113, 481506. GODFREY-SMITH, P. 2023. Nervous Systems, Functionalism, and Artificial Minds. KUHN, R. L. 2024. A landscape of consciousness: Toward a taxonomy of explanations and implications. Progress in Biophysics and Molecular Biology, 190, 28-169. LEDOUX & LAU, H. 2020. Seeing consciousness through the lens of memory. Curr Biol, 30, R1018-R1022. LEDOUX, J. E. 2021. What emotions might be like in other animals. Curr Biol, 31, R824-R829. LEE, A. Y. 2020. Does sentience come in degrees? Animal Sentience, 29. LENHARO, M. 2024. AI consciousness: scientists say we urgently need answers. Nature, 625, 226. LEVY, N. 2014. The Value of Consciousness. J Conscious Stud, 21, 127-138. MERKER, B. 2007. Consciousness without a cerebral cortex: a challenge for neuroscience and medicine. Behav Brain Sci, 30, 63-81; discussion 81-134. METZINGER, T. 2020. Minimal phenomenal experience. Philosophy and the Mind Sciences, 1, 1-44.