The Theology of Robot Religiosity

From a theological perspective, the idea that other creatures—biological or synthetic—might become religious subjects is not completely absurd. The Judeo-Christian tradition describes a loving God who takes an interest in every element of creation (e.g., Matthew 5) and a created world that is called to praise God with each of its elements (Psalm 148), even though the exact way that should happen is shrouded in mystery. But this is not the kind of religious subjectivity that is most interesting when it comes to intelligent robots. As elements of this same world, robots would also be called to participate in creation’s praise. But if they did so in the same way as trees and rocks, one would hardly call that religion. What matters most is whether they would seek and worship God in a way that is closer to the complex way humans do it. Would they have a deep longing for the transcendent? Would they have beliefs? Could they receive divine grace and revelation?

On theological grounds alone, because God is radically free and unpredictable, such a possibility should not be excluded. It is not a big stretch of imagination to say that a loving God would be interested in communicating with any sufficiently developed creature that could withstand such communication. Human religions, at least the monotheistic ones, claim to have been initiated through a divine act of revelation. Theologically speaking, this means God is interested in communicating with certain kinds of beings because, as these religions strongly claim, there is at least one such being God reaches out to: humans.

In the Judeo-Christian narrative, there is a covenantal relationship between God and humanity that sets us apart from the other creatures. But God’s election of humanity was presumably not arbitrary. We cannot know for sure why God chose Homo sapiens, and the question of what exactly sets humans apart and makes them in God’s image and likeness (Genesis 1:24) is one of the big puzzles in theological anthropology (van Huyssteen 2006; Cortez 2010, 14–40). If the notion of imago Dei is indeed linked with the unique I-Thou relationship humans can have with God and with each other (Dorobantu 2022a, 2024), then it can be reasoned that God elected our species precisely because we were the only ones capable of such relationality. This theological tradition speaks of a personal God who is eager to share love, knowledge, and an appreciation of the good with other free personal selves, a God eager to reveal Godself and even incarnate for the redemption of these other personal selves. It is therefore not controversial to suppose that, as soon as such a creature has emerged in the universe through evolutionary processes, God would invite it into a special kind of relationship. Following this logic, it is consistent to suppose that, were other such creatures to emerge, God would want to do the same thing with them. So, intelligent robots might theoretically also one day become subject to divine revelation and thus authentically religious, were they to develop into the kinds of creatures that can relate personally with God. The only two questions left are: (1) what kind of abilities are required for a creature to become interested in, and interesting for, God in this special way, and (2) whether AI can aspire to develop them.

With respect to the first question, we are obviously in speculative territory, but sentience looks like a sine qua non condition for such I-Thou relationality. Not all sentient creatures could be authentically religious, but surely the latter is not possible without sentience. Only a creature endowed with consciousness, a first-person perspective, and feelings could potentially become a worthy partner for God, a Thou to whom God could speak and who could respond. The AI would need to develop some form of interiority; there would need to be “someone home” (Peters 2022). Searle distinguishes between strong AI—a conscious entity, an emulation of intelligence, that has its own mind and phenomenal experience—and weak AI, a mere simulation of intelligence, a mindless information processor that never truly knows, understands, or feels anything (Searle 1980). For AI to become religious, it would obviously need to be strong AI. Whether strong AI is possible is hotly debated without much consensus, mainly because we do not understand how consciousness works in biological creatures such as ourselves, and, more generally, we do not understand how consciousness is possible at all in a material universe, at least not according to our current understanding of the laws of nature. This is one of the biggest questions in philosophy, popularly known as the “hard problem of consciousness” (Chalmers 1995). What is important to note here is that theology would not in principle be opposed to the possibility of AI as a religious subject, were technology to develop in the direction of strong AI.

As stated, sentience is a mere minimum prerequisite, and more is required for a sentient creature to qualify for an authentic personal relationality with God and other selves. There is not enough room to explore this question in detail here, but a certain degree of intelligence is also evidently required for sustaining relationships over time, being able to take the other’s perspective, simulating social interactions in one’s mind, and reasoning about what is morally good and appropriate. This might be the reason humans can sustain such relationships while nonhuman animals cannot, at least not to the same degree of complexity. However, certain limitations might be just as important as these capabilities for rendering an intelligent sentient creature a personal self. Features like vulnerability, mortality, self-insufficiency, and even an upper limit to one’s rational abilities and knowledge might be instrumental in bringing about the kind of creature that can be a religious subject (Dorobantu 2021a). AI may or may not develop in that direction, and for the time being it seems that it does not, but the important point is that theological grounds do not in principle exclude such a possibility. It is the more naturalistic approach to religion explored in the following section, as opposed to the theological one, that seems to be more restrictive and cautious about the likelihood of robots ever becoming religious.

Religious Robots in Sci-Fi

In this section, I argue that the evolutionary account of religion should make humans rather skeptical about the possibility of artificial systems developing authentic religion. This is a big claim, which might surprise many, so it is worthwhile to first outline how counterintuitive this view is. If we are to interact with AI programs that claim to be religious, it might be helpful to be aware that our intuitive understanding of the nature of religion and our inherent tendency to anthropomorphize AI predispose us to believe such claims too easily.

The sci-fi genre is a good place to illustrate these biases. When it comes to AI, the popularity of sci-fi is indicative of the fact that these stories resonate deeply with our intuitions and anxieties about how such technologies might develop. I argue that the way we imagine religious robots in sci-fi betrays a deep, albeit wrong, intuition that religion has to do mostly with the intellect. Robots that turn religious in sci-fi do so mainly for intellectual reasons, or at least as a result of intellectual deliberation. As human intuition goes, if AI systems become intelligent enough, they will surely want to understand the world better and start asking questions about their genesis, their identity, hidden causality in the world, and ultimately about why there is something instead of nothing. This type of thinking is very common in sci-fi.

QT-1, also known as “Cutie,” from Isaac Asimov’s short story “Reason” (Asimov 1941) struggles to understand how it was possible to have been created by such inferior beings as humans. Looking for a more satisfying explanation for its own existence, it develops a religious-like belief system around the energy converter, the primary machine on the space station. It starts regarding it as a divine-like entity and refers to it as the “Master,” while regarding itself and the other robots on the station as the Master’s primary servants.

Klara, a solar-powered robot designed to be an artificial friend to children in Kazuo Ishiguro’s novel Klara and the Sun (Ishiguro 2021), develops quasi-religious beliefs and practices around the sun, which she regards as a benevolent life-giving power. Due to her dependence on solar energy, she starts worshipping the sun, asking it for guidance and attributing to it the ability to intervene in human affairs. Klara even performs a kind of ritual in front of the sun, asking it to cure her human friend.

Demerzel, a robot in the TV series Foundation (an adaptation of Asimov’s trilogy), is depicted as an adept of the Luminist religion, claiming to have been blessed with visions and epiphanies. She attributes her religiosity to the need to believe in something greater than herself, despite her instincts being entirely progrmamed as computer code (season 1, episode 6, “Death and the Maiden”).

In all these depictions, the intellectual dimension is crucial in the development of religion in AI. Robots turn to religion because of their need to make sense of events in the world and their own identity. Religious explanations are depicted as more plausible accounts of reality than what meets the eye or what the robots had been programmed to think. If the popularity of such sci-fi stories is a reliable indicator, it shows that intuitively, humans believe intelligent robots will eventually develop something like religion because they will be like us and thus need religion to make sense of the world and construct meaning, just like we do, as convincingly argued by William Clocksin (2024) and Robert Geraci (2024) in articles in this same issue. I aim to show that this intuition is wrong on two counts. First, making sense of the world is likely not the main reason humans developed religion. Second, intelligent robots will not be anything like humans with respect to the capacities and propensities relevant for becoming religious.

Religion in Human Evolution

To explore whether authentic religion could plausibly emerge in robots, one can start by asking how religion emerged in human evolutionary history and how it is enabled by our specific embodiment and cognitive architecture. Inevitably, this approach departs from the only example we have of a religious creature, which is ourselves. The roots of this approach are thus inevitably anthropocentric, but even such a limited approach might lead to valuable hints about potentially generalizable features of religion.

In a recent book, Oxford anthropologist Robin Dunbar (2022) proposes a very provocative account of the evolution of religion in humans, which goes somewhat against the grain of a widespread view in the cognitive science of religion known as the “Standard Model” (Powell and Clarke 2012). The Standard Model regards religion as maladaptive or, at best, a neutral development, an accident of human evolution, because it does not provide any fitness benefit for the individual—quite the contrary (for example, through self-imposed pain, celibacy, or even self-sacrifice).⁵ As the argument goes, religion is best characterized as a strange side effect, a misfiring of cognitive capacities that evolved for very different purposes. Such maladaptive behaviors or traits are well documented in human evolution when capacities that evolved for very particular reasons, purposes, and contexts start misfiring in strange ways when taken out of the original context (Reading 2011; Frankenhuis and Del Giudice 2012). A familiar example is our so-called “sweet tooth,” a fondness for sweet food that was probably very useful in the resource-scarce environment where our species evolved. If you do not know when your next opportunity to eat will arise, it makes sense to eat as many of those fresh berries as you can. But that same propensity can be self-defeating in our current environment where food is always available, potentially nudging us toward obesity and cardiovascular disease.

According to the Standard Model, religion might be precisely such a maladaptation. Psychologist Justin Barrett (2000) makes a strong case that religion might be caused by our so-called hyperactive agency detection device. He hypothesizes that human perception is very suited to detecting even the slightest trace of agency, to the extent that it sometimes produces false positives, seeing agency where there is none. From an evolutionary perspective, it pays off to be a bit paranoid and avoid even the slightest hint of danger. As Barrett’s argument goes, it is only natural for this agency detection device that runs continually in the background of our minds to produce strange conclusions in a world where we do not fully understand the causality around us, especially with regard to the natural elements. The human mind is thus poised to start projecting personal agency behind storms, floods, or the movement of celestial bodies. According to Barrett’s proposal, this is how prehistoric people might have first gotten the idea of supernatural forces or gods, which then led to the development of religion. If this is in fact how religion evolved, then it is a by-product of this hyperactivity of our minds.

Against such views, Dunbar (2022, 49–76) argues that religion is actually evolutionarily useful, and that there are very good reasons humans developed it. One reason is that religion helps strengthen the bonds between members of human groups, especially through rituals. Activities such as dancing, eating, and singing together trigger massive endorphin releases. If such activities are performed synchronously, the endorphin release is even higher. This is important because endorphins are the natural painkillers of the brain. The endorphin releases triggered by religious activities thus provide positive feelings of relaxation and satisfaction, increasing the pain threshold of the individuals.

Another reason Dunbar regards religion as adaptive, instead of maladaptive, is the way in which religion aids in reducing social friction within larger human communities. In human history, this effect is particularly noticeable after the agricultural revolution, when human communities grew dramatically in numbers. In relatively small human groups, the heat generated by social friction can dissipate in various ways, but for larger groups, that is more difficult. Life in big communities can get significantly more stressful, which may increase infertility rates. Conflict resolution between members is also more difficult to achieve in large cities than in small hunter-gatherer communities. Here, too, religion can be helpful, providing opportunities to diffuse such tensions and reset social relationships during religious festivals.

A key argument in Dunbar’s account is that religion is crucial in smoothening human collaboration on large scales. To be able to collaborate, people need to trust each other, but that usually requires extended periods of time to build up, strengthen, and test this trust. In addition, trust can be lost quite abruptly. Our minds are always busy updating their models of the other humans who are relevant in our lives, because in order to trust them, we need to reliably predict how they will behave in various situations. This is extremely demanding from a cognitive point of view, which is why there seems to be an upper limit to how many individuals one can keep track of in detail at the same time. Dunbar’s previous research convincingly documented this number to be around 150 (Dunbar 1998), which is already about three times higher than what our closest primate cousins are capable of. This came to be known as Dunbar’s number, and it is extremely relevant because it also explains why for most of our (pre-)history, human groups were limited in size to around 150 individuals.

This explains why the emergence of religion was so instrumental in our species’ history. According to Dunbar, religion provided people with shared narratives in which an unlimited number of individuals could believe. People could suddenly trust and collaborate with others outside the social network of 150 individuals their brains could viably keep track of, not by getting to know them very well, but on account of shared religious beliefs. If that is so, it implies that religion was critical in enabling the emergence of civilization as we know it. Without it, our species might have remained stuck within the confines of Dunbar’s number, never really tapping into the latent potential of collective intelligence and large-scale collaboration.

Crucially, shared beliefs and narratives do not account for how religion first emerged. In Dunbar’s account, the evolution of religion can be understood in two phases: shamanic and doctrinal (see also Watts and Dorobantu 2024). The shamanic phase is characteristic of hunter-gatherer societies. It is eminently embodied, very immersive, and centered around ritual, especially trance-dancing, sometimes combined with the use of psychedelics. The doctrinal phase—which entails specific beliefs about certain gods and thus something that might be called theology—began much more recently in human history. It is largely connected to humans’ transition from small hunter-gatherer tribes to big cities.

The main takeaways for the question tackled in this article are that the doctrinal phase only occurred subsequent to the shamanic one and, even more importantly, that doctrinal religion never completely replaced shamanic religion. Instead, it merely built on top of it, creating an additional layer that appeases our discursive-logical type of intelligence. The shamanic dimension is still active and powerful in all modern religions, which are thus both shamanic and doctrinal. The latter emphasizes adherence to dogmas and principles, while the former is more about the experiential dimension of religion, having to do with the more ancient parts of our brains and ways of knowing. The shamanic layer is thus arguably the driving force behind religion, with the doctrinal dimension being left to do the explanatory job.

This distinction is at the heart of Dunbar’s subtle criticism of the Standard Model, with its focus on beliefs and the intellectual dimension of religion at the expense of the human body, ritual, and community. If Dunbar is right, then I believe this implies a significant shift in framing the discussion of potential authentic religiosity in robots. The main implication is not to rule out the possibility of robot religiosity but to shift the discussion about how and why it could emerge.

Dunbar’s account links the emergence of religion to the specific vulnerabilities of early humans, who needed to boost their immune systems, increase their pain tolerance, mitigate the social friction accumulating in their communities, and scale up their cooperation beyond just their closest acquaintances. The human body and embodied cognition are much more central to the emergence of religion in this account than in the Standard Model. This might be relevant to the discussion of hypothetical religious robots because their embodiment would be very different, and so would their needs, limitations, and vulnerabilities. Just because humans developed religion as a solution to the specific challenges encountered during the evolution of our species, there is no reason to think AI should do the same.

Human religion is rooted more in the unconscious, embodied, affective dimension of our cognition than in the logical, rational one. Religion may indeed have something to do with how we understand the world, but this intellectual need is not its primary driver. This is one of the two core fallacies in the sci-fi stories mentioned earlier (the second one is unpacked in the next section). This fallacy is rooted in an outdated view of how religion emerged and what functions it primarily serves. The idea that religion emerged as a set of theories about how the world works can be traced back all the way to late Victorian scholars of religion like Edward Tylor and James Frazer, who believed religion began as a theory about causality in the world and was put together by “savage philosophers” in hunter-gatherer communities (Tylor [1871] 1903, 429). Barrett’s hyper-agency detection device proposal follows the same principle, linking religion to our imperfect way of inferring causality and agency. If religion were indeed about that, then perhaps strong AI might also one day develop something similar to make sense of its own existence. However, what is increasingly seen in the science of religion, as typified in Dunbar’s proposal, is a realization of how much religion is rooted in the body and the unconscious, affective, less rational part of our cognition. Religion is intimately linked with our particular embodiment, the specific needs of prehistoric human communities, and the idiosyncrasies of our cognition. It might thus more likely be a typically human phenomenon than a universal state that all possible intelligent agents go through.

This conclusion is somewhat surprising, given that most current discussions about robots and religion (McBride 2019; Sampath 2018), including those typified by the example of the Google engineer and LaMDA, revolve around notions of doctrine and what might be plausible for robots to believe in. In contrast, Dunbar’s account reveals that doctrine represents just the proverbial tip of the iceberg in human religion. Whether that tip alone could count as authentic religiosity is debatable, but a strong case can be made that robo-shamanism might need to precede and underpin robo-doctrine. However, it is far from clear whether intelligent robots will have the kind of needs and historical contingencies to develop anything like shamanism.

Intelligent Robots Will Not Be Humanlike

The second core fallacy of human intuition about intelligent robots, as captured in sci-fi, is the too quick assumption that they will develop very humanlike concerns, fears, and desires. This is why we readily see them becoming existentially curious and ultimately religious. But this assumption is predicated upon a rather naïve anthropomorphizing of AI. Because AI is so competent at simulating human intelligence, we are too quick to assume it sees the world and thinks like us. The very notion of human-level AI emerging sometime in the future is often interpreted to imply that AI is somehow inching toward becoming very similar to us. In reality, AI is a very different type of entity, and in the future, it might become even less humanlike.

It is by now common knowledge that current AI programs have very distinct ways of learning and solving problems (Song et al. 2024), even though their output is often similar to what a human person would produce. Image-recognition software does not learn about the possible objects in pictures and their characteristics. Instead, it is being fed hundreds of thousands of human-labeled examples that enable it to learn strange but effective patterns and correlations present in the training data. Humans can achieve similar results with far less examples, and sometimes only one example will suffice. When AlphaGo learned to play the strategy game Go, it did not start with the abstract principles of the game (e.g., that territory is more easily gained in the corner than in the center of the board). Instead, it was just fed thousands of games played by human experts and left to learn by itself the correlation between various positions and the likelihood of victory (Silver et al. 2016). When large language models like ChatGPT learn to generate human language, they do not learn concepts, objects, or the relationships between them. Nor do they learn the syntactic rules of language. Instead, they consume colossal amounts of human text, mainly from the internet, and learn to reliably predict what is most likely to be the next word based on the context. This is why when AI programs do make mistakes, they usually look very strange, and not like the kind of mistake any human would make (e.g., Eykholt et al. 2018). Even the name of the architecture on which these programs are run—artificial neural networks—is deceiving, because they have little in common with biological neural networks; they are crude, simulated simplifications. Not much about these AI programs, if anything at all, is humanlike.

Robots of the future could become even more alien. Even if they somehow became sentient, thus strong AI, as opposed to today’s weak AI, they would likely develop very non-humanlike kinds of minds and think very non-humanlike kinds of thoughts (Shanahan 2016; Dorobantu 2021b). To illustrate this, let us speculate on just three aspects of their cognition: their sensorial perception and the kind of mental world they might construct based on it; their experience of the passage of time; and their introspective abilities.

First, a conscious robot’s world of perception would be profoundly different from our own. Because of this, the AI would likely inhabit a completely different world, be driven by completely different needs, and have completely different thoughts, feelings, and aspirations. It might develop different senses far stranger than, for example, the ability to see in infrared or ultraviolet. One such alien sense could be an ability to intuitively grasp someone’s identity and whereabouts, not through face or voice recognition but by “smelling” their digital footprint from their connected devices, a sort of digital sixth sense. The AI might detect slight changes in the radio field of Wi-Fi and Bluetooth and thus infer movement, something akin to echolocation (Brooks 2017). With such an ability, the AI might immediately perceive changes in people’s breathing and heart rates and thus infer their emotional states. This would amount to a emotional intelligence of a very strange kind.

Second, a conscious robot might think ridiculously fast and experience the passage of time as much slower than we do. Speed of thought is believed to be proportional to the speed at which electrical signal can travel between neurons, and computer hardware can support much higher speeds than biological tissue. A conservative estimation is that AI could think around ten thousand times faster than humans, thus experiencing the passage of time ten thousand times more slowly. As futurist James Lovelock put it, “the experience of watching your garden grow gives you some idea of how future AI systems will feel when observing human life” (Lovelock 2019, 81–82). This quantitative difference might also amount to a qualitative shift: at such speeds, the AI might even end up living in the weird quantum world. How could we have anything in common with a creature inhabiting such an alien world of perception?

Third, a robot’s introspective capabilities might also be profoundly dissimilar to our own. According to John McCarthy (2007, 1178–79), one of the founding fathers of AI, it may have complete accessibility to its internal states and operational algorithms. In contrast, humans can be aware of only a fraction of our underlying cognitive processes. Everything about human nature depends on this depth of our unconsciousness: our relationships, our aspirations, our virtues, and our sins. This inherent lack of self-knowledge is what compels us to create art and forge relationships as means of exploration. We do not completely know ourselves, which is why we need to explore, but even a lifetime of exploration cannot completely dissolve the mystery. To the contrary, it often enhances it. AI might entirely know itself. With a perfect recollection of everything it ever experienced or learned, and an ability to analyze the exact causes of its actions and accurately predict its future behavior, AI would have very little in common with our way of being. If religion and human existential restlessness are deeply rooted in an imperfect knowledge of ourselves, there are few grounds to believe AI would have similar traits and interests.

Conclusion

Discussions about robots becoming religious might therefore hinge not so much on them being able to produce humanlike outputs but on how humanlike they would be “on the inside.” Are current AIs becoming more humanlike? Perhaps in terms of their outputs, but surely not in terms of their ontology, internal structure, world-modeling, and problem-solving. Large language models such as ChatGPT and LaMDA might produce outputs that seem very humanlike, to the extent that they might even convince people of their personhood, sentience, and humanness, but their inner workings are radically different from our own. Human-level competency is not the same as humanlike cognition, and when it comes to developing religion, the idiosyncrasies of the latter seem to be what matters. If the limitations and particularities of human nature are as important as they seem for the emergence of religion, then it is safe to predict that AI will not undergo similar developments without having the same kind of embodiment, needs, or evolutionary constraints.

This conclusion is merely tentative, and it only refers to a type of religiosity very similar to our own, which is the only we can so far imagine. What this article argues, especially in its second half, is that human religion seems to have emerged in a “bottom-up” fashion, and that the “bottom” part still largely underpins, albeit imperceptibly, even the most intellectual forms of religiosity seen today. This important bit often seems to be overlooked when imagining future artificial systems that are also religious. However, it cannot be completely ruled out that other forms of, or paths toward, religion exist that are equally authentic but simply impossible to imagine with our current knowledge. We only know of our own long and idiosyncratic path to becoming Homo religiosus, but perhaps techniques like predictive coding, probability estimation, or evolutionary computation could lead to the emergence of something sufficiently close to what we call “religion” in artificial systems. One question that would then arise is whether such developments would lead to the proliferation of as rich and diverse forms of religious expression as with human religiosity. Another question would concern the fascinating content of AI religious beliefs, which would probably go far beyond simply reinterpreting the sacred texts of human religion with a robotic twist, as has been suggested (Sampath 2018; McBride 2019). The theologian can only feel a thrill about the possibility of ever witnessing such developments and learning from what they might have to offer (Dorobantu 2022b). However, the religious scholar needs to emphasize the idiosyncratic dimension of religion as we currently best understand it, and thus the unlikelihood of something as profoundly different as AI ever becoming religious.

In a 2014 homily, Pope Francis provocatively said that, were a martian ever to ask to be baptized, we should probably comply, so as not to risk pushing back against the very work of the Holy Spirit (Ohleiser 2014). If the pope were to adapt that sermon today, he might say the same about an intelligent robot. As shown in this article, on theological grounds, such an attitude might be warranted. However, from a cognitive perspective, it might be wise to try to probe the martian or robot’s sincerity a bit more deeply before celebrating its conversion.

Acknowledgments

This work was supported by the Templeton World Charity Foundation under Grant TWCF0542. The opinions expressed in this publication are those of the author and do not necessarily reflect the views of the Templeton World Charity Foundation. An earlier version of this paper was presented at a 2022 symposium at Vrije Universiteit Amsterdam entitled “Could Robots Be Religious? Philosophical, Theological, and Ethical Perspectives on the Future of Artificial Intelligence.” The symposium was supported with a grant from CLUE+, the Vrije Universiteit Interfaculty Research Institute for Culture, Cognition, History and Heritage.