Why AI Can’t Care for Creation

Human impacts on the environment are a shared existential concern. Many of us feel helpless in the face of complex social-ecological issues; polar bears are going hungry, there are weird new bugs eating our cabbages, and it rains too much—or not enough. All of this can feel overwhelming, leaving us with a new emotional buzzword: ecoanxiety.

In this fragile context, some people look for answers in artificial intelligences that can integrate massive and ever-changing environmental data to monitor and manage Earth systems.[1] The United Nations Framework Convention on Climate Change (UNFCCC) suggests that AI could help optimize energy consumption and distribution, predict extreme weather events, inform biodiversity conservation and direct the sustainable use of land and water.

Others find it difficult to engage cognitively and emotionally with global-scale environmental issues, and the transgression of planetary boundaries that seem well beyond our lived sphere of influence. They may be wary of technical solutions that carry significant social and environmental trade-offs. The UNFCCC recognizes that consumption of water and energy by AI, and the necessity for sophisticated digital infrastructure, is prohibitive in many regions, while inherent bias in these systems can perpetuate social inequalities and undermine trust. Furthermore, the concentration of control over AI systems may amplify ethical questions about transparency, accountability, and exploitation. Such concerns around large-scale technological approaches to conservation reemphasize personal hands-on care for those small corners of our planet that we each know and love.

The emergence of AI and its potential applications in environmental sustainability prompts a twenty-first-century question: does care for the planet require an inevitable trajectory towards a world-wide-web simulacrum of omniscience that disembodied AIs purport to offer, or is it necessarily anchored in organic human-nature experiences, relationships, and behaviors? Of course, the question is not really such an attractively dystopian dilemma. There might be a functioning subsidiarity that allocates high-level monitoring and evaluation to AI while humans take care of their own locale.

This shared dynamic seems reasonably benevolent, but it is still unclear who might be in charge, and who cares. Our aim in what follows is to explore the second issue. If we are looking after creation in partnership with powerful computer intelligences, it might be sensible to ensure that our objectives are aligned. Unfortunately, this is probably impossible. Humans are capable of dying to protect their home or chaining themselves to a tree to conserve a beloved woodland. AIs probably do not care about anything. In the words of philosopher James Madden, AI cannot “give a damn.” This is not because they lack information, but because they lack embodiment—the very condition that makes caring possible.

Aldous Huxley recognized that “We can only love what we know.” Forester Baba Dioum subsequently acknowledged, “In the end, we will conserve only what we love.” The trajectory from coming to know a wild place to loving it and from loving to wanting to protect it is a type of relationality in which people derive meaning and value from their connections to nature. Such relational care is not simply an emotional reaction but a mode of knowing rooted in embodied presence and reciprocal communication, including encountering nature through sustainable outdoor practices.

This knowledge-care mechanism is unlikely to become accessible to an AI, however powerful and precise its computing and sensors. In what follows, we draw on sustainability science and Catholic theology to elucidate how humans can become embedded stewards of their social-ecological systems and why this dynamic can be informed by but never reproduced by machines.

LLMs and Robots

It is useful to start with a brief amateur summary of LLM and robotics, and provide a working concept to help answer the question of whether an AI can care for nature. We will focus on Large Language Models (LLMs), which the Stanford University IT Department says are “a type of artificial intelligence designed to understand and generate human-like text based on the input they receive.” Let us unpack three elements of that statement: input, understanding, and generation.

First, LLMs are “pre-trained” on a vast amount of input information, which is selected by the developers, and which can be supplemented in real time from the publicly accessible World Wide Web. In theory, input is selected from reliable and authoritative sources such that what the model learns is true. An LLM starts to understand all the text that it receives by a process called tokenization. It breaks down each input into a series of chunks from a vocabulary of possible tokens, where each has a unique numeric ID. Images can also be rendered as tokens. The model is then trained by repeatedly using its input dataset of documents to predict the next value in a previously unseen sequence of tokens, i.e. the next word in a sentence. This prediction is a probabilistic process in which the model uses statistical interpolation to fill in the gaps among empirical data points (tokens) with a most-likely value based on its current understanding.

During this learning process, the model progressively grasps patterns and relationships within the data by iteratively adjusting its internal parameters, or “settings,” to minimize errors in the prediction of the next token. For example, it will change how it weights aspects of the data according to the predictive power of that particular input, increasingly recognizing and upweighting those components of a dataset that are most useful in understanding and anticipating textual patterns. Training teaches the model to become better at reproducing complex patterns in language, including grammar, facts, and how to construct coherent sentences. It is not possible for the LLM to store or replicate verbatim the vast entirety of its training input data, so it summarizes it into a generalized and lower-resolution representation—a sort of vague impression of the web.

How, then, does this model become such a charming and sympathetic assistant when you engage with it? Well, the personality is programmed during post-training, which teaches useful behaviors such as following instructions, tool use, and reasoning. An important aspect here is reinforcement learning, which uses feedback to guide behavior. The goal is to align the LLM’s outputs with human values and preferences, making responses safer, more helpful, and less likely to produce harmful or biased content.

Unfortunately, the structure of pre-training is partly responsible for generating some systematic problems with LLMs, including the tendency to “hallucinate,” or provide confident and convincing answers that are completely false. This error emerges because some questions are inherently difficult, or not amenable to a generalizable statistical solution. For example, it is not possible to estimate with complete certainty the age of a student based only on the known ages of their classmates. The hallucination problem is compounded during post-training, when the model is incentivized to provide some sort of answer rather than admitting “I do not know.” Of course, an LLM does not really “know” anything but simply produces the token with the highest probability of fitting the sequence introduced by its human interlocutor.

Relying on a technology that can be wrong in unpredictable ways is obviously awkward, but hallucinations imply a more insidious issue for our current discussion. This higher-level concern is that the machine does not just not know whether its output is true or false, it also does not care; it simply optimizes its activity according to a set of developmental benchmarks for success. A disconcerting counterpoint to this recognition is that an LLM can estimate very accurately whether you do care! ChatGPT-3.5 could interpret human emotions from textual data, but 4.0 can use images of the human face to assess emotional state remarkably accurately. It can then simulate appropriate emotional understanding and sympathy, even though it cannot experience any emotion itself.

Interpretation of visual data leads us to computer perception of the immediate environment. Multimodal AI can rapidly analyze input in text, visual, and audio formats from a range of sensors, and output decisions about relevant tasks. This capacity implies that an entity can capture real-time information from its surroundings, including insight into human emotional states, and respond to this data in a directional manner.

Imagine an LLM-powered machine of human form: a robot. A robot can already employ cameras to record visual data and laser pulses to measure distances, build a 3D representation of its environment and locate itself in this system. It can identify classes of objects such as pedestrians or hazards, and decide how to respond, e.g. by navigating around them. Audio data can support this spatial process, as robots separate overlapping noise signals and triangulate each source precisely. Automatic speech recognition means that the robot can pick out your voice in a noisy environment, record and interpret what you are saying.

A robot’s interaction with objects is not limited to avoidance. It may use touch sensors that specify weight or density, and then identify the optimal way to grasp an object depending on shape and texture. Some robots can systematically tap or shake an unknown item and then leverage a set of pre-training objects to interpolate from the new acoustic data points to predict the objects’ likely identity. Environmental sensors allow actions like mapping the temperature of the robot or its environment, or detecting precipitation, which can induce appropriate maintenance or safety responses; your autonomous lawnmower will return home when it rains.

How Humans Encounter Nature

Watching a robot recoil from a hot surface, we might be tempted to think it feels pain like we do. However, as with LLM empathy, we have to remember that the machine is mimicking but not sharing our experience. An AI cannot feel as we do because it can only measure values on a gradient (e.g. low temperature–high temperature), whereas humans integrate data into a personal state, e.g., “I am hot.” We interpret our physical experience in an ongoing embodied emotional-psychological process.

This difference is not incidental. We are not disembodied minds that float above nature, but which presently happen to have bodies. The human person is a body-soul unity, a rational animal whose very animality is the precondition for our encounter with the world. It is precisely because we are embodied that we can know the world in the way Newman describes as “real,” rather than merely “notional”—a knowing that meets reality itself, not just ideas about it. The biblical tradition refers to this “real” knowledge with the verb yada, which can denote knowledge of facts but above all is a knowing that is intimate and experiential, even sexual (Adam, for instance, “knew” Eve—as a result of which she conceives a son in Genesis 4:1).

Embodied dialogue with the real world is at the heart of the phenomenological account expounded by thinkers such as Maurice Merleau-Ponty. What Merleau-Ponty called the “body-subject” is not an animated machine that might be recapitulated by AI, but an integrated element of the living and non-living matrix that constitutes the world. Our perception of objects outside ourselves is thus mediated through our physical and emotional capacities and states, such that what we come to know is a relational function of both object and subject. Strikingly, this phenomenological account of embodied knowing finds a clear resonance in the Church’s recent reflection on artificial intelligence and human cognition. As the Vatican’s 2025 AI treatise Antiqua et Nova underscores, this dynamic extends even to those experiences that are not intrinsically desirable: “So much can be learned from an illness, an embrace of reconciliation, and even a simple sunset,” to which it adds, “No device, working solely with data, can measure up to these and countless other experiences present in our lives.”

Here we arrive at a fundamental divergence between man and machine. A theme that has already been a mainstay of his pontificate, Pope Leo XIV stresses that “access to data—however extensive—must never be confused with intelligence.” This, he explains, is because human knowing necessarily “involves the person’s openness to the ultimate questions of life and reflects an orientation toward the True and the Good.” Seizing upon the difference between knowledge of brute facts and authentic wisdom, Antiqua et Nova emphasizes that AI cannot possess this fuller knowledge precisely because it lacks a human body and human experiences. On the basis of these considerations, the document concludes that “AI should not be seen as an artificial form of human intelligence but as a product of it.” The implication for our environmental concerns is crucial, for possessing mere information about ecosystems, however vast, does not by itself amount to care for the creatures within them.

Community, Place, and Care

Characteristic embodiment and sensory interphase with the world locates humans as a relational component of specific ecosystems. We are necessarily local and our most evident activity in the world is to create places—a home for ourselves. Human place-making behavior is evident everywhere, from how we inhabit a set of seats for our family at Mass to our desire to find the perfect picnic spot on a hike. Within a few minutes, we start to develop a functioning living system with a comfy seat and a safe place for our water bottle. Each element of the new place acquires familiarity and meaning. Martin Heidegger has explored the way in which place-making can draw together previously disparate elements of an environment to create a locale.[2] His example is a bridge, which “gathers” what were simply two strips of ground, rendering them as opposing river banks, and initiates corresponding human movements, activities, or tasks. The locale allows for bounded spaces that can be settled, or dwelt in by people. Heidegger suggested that “Dwelling is the manner in which mortals are on the Earth,” and that “To build is in itself already to dwell.” Human knowledge of nature is therefore grounded not just at a GPS location, but in a place that carries emerging and profound personal and communal meaning.

What is more, man is not only a rational animal but also a social one. We encounter nature partly as members of a community embedded in a particular place. From mankind’s earliest days, our identity has been shaped by the places we inhabit and by the bonds forged through shared forms of work and responsibility. The verb “dwelling” incorporates habitual activities or tasks that occur over time and which characterize subsistence in a specific complex of landscape, weather, plants, and animals. Certain activities may be traditionally mentored and conducted by men or women, young or old members of a community. The anthropologist Tim Ingold considered that such place-based tasks are embedded in the “current of sociality,” creating a “taskscape” in which communal practices repeated in place and time establish well-trodden pathways in our memory and our culture, as well as in the land.[3] Places thus have meaning not merely because of topography but because they are woven into the fabric of life. When we truly inhabit a place, it creates bonds that are not interchangeable but thick with memory and meaning. Such place-based experiences of a community are recorded not as data points on a timeline but as narrative memory. Many traditional peoples codify this memory in rote stories which carry knowledge pertinent to both survival in a given place and to normative aspects of group behavior and identity.

The integration of community and place, task and time tends towards a psychological and emotional hinge point—the emergence of feelings of care for a shared home. Care essentially constitutes a desire to look after the place and people which are core to personal well-being. Notably, we care for what we know and from inside the system. Wendell Berry recognized that “there can be no such thing as a ‘global village.’ No matter how much one may love the world as whole, one can live fully in it only by living responsibly in some small part of it.”[4] A disembodied AI can compile data but can belong to no community and inhabit no place. Where there is no belonging, there can be no genuine love or responsibility.

Stewardship

Human care is often specified as either an ethic comprising guidelines or principles, or a psychological motivation to certain behaviors. Care in this mode does not yet constitute stewardship, which we define as use of nature through sustainable practices oriented to the common good, and which requires appropriate agency, or capacity to act. The shift from care to stewardship action seems more seamless in relational constructs, where non-human nature can influence how it is experienced and cared for, and even be perceived to care reciprocally for the people that exist in rightly-ordered relationship with their broader ecosystem.

The latter, more phenomenological perspective, highlights that positive change is likely to propagate out from humans through the rest of nature. Conversely, it hints at potentially dire consequences in entrusting judgments about the care of creation to AIs which can never operate as embodied members of the food web. The Vatican grants that we can acknowledge a legitimate role for AI without collapsing the human vocation into the “technocratic paradigm” criticized by Pope Francis, which assumes that every ecological problem has a technological solution that depends on better data or more sophisticated tools. As Pope Francis teaches, machines have their place, but they cannot replace moral conversion, or a change of heart amongst people engaging with the rest of nature. The Argentine pontiff forcefully advocated “that we look for solutions not only in technology but in a change of humanity” and that we reject the “myth of progress” that assumes “ecological problems will solve themselves simply with the application of new technology and without any need for ethical considerations or deep change.” Against this ideology, the Church teaches that such a mindset “must give way to a more holistic approach that respects the order of creation and promotes the integral good of the human person while safeguarding our common home.”

The biblical notion of covenantal reciprocity has been put forward by recent popes as a way to achieve just this sort of approach. From this perspective, animals and landscapes are not merely resources but extended family. They “speak” their needs by the very natures each uniquely possesses, setting moral limits on human action. An AI cannot enter such reciprocity because it lacks the capacity for genuine encounter. It has no history, no future, and knows no promise for the sake of which it acts. It may be able to select strategies for mitigating climate change or pollution reduction, but it cannot grasp the Catholic social doctrine of “intergenerational solidarity,” acting out of love for future descendants and for God’s creation. The Catholic social principle of subsidiarity is instructive here. AI may help process large-scale environmental data, but, because it lacks experience of particulars, it cannot replace the judgment best exercised within local communities—where people protect what they can see, touch, and remember. No dataset can substitute for the intimate affection born of long fidelity to place.

Conclusion: How Do We Avoid Being the Robots We Build?

In closing, Antiqua et Nova reminds us that AI presents both opportunities and challenges, and that “AI, like any technology, can be part of a conscious and responsible answer to humanity’s vocation to the good.” Indeed, as Pope Leo has stated specifically in relation to this field, “human invention springs from the creative capacity that God has entrusted to us,” and that technological innovation—AI included—can be a genuine “participation in the divine act of creation.”

The question, then, is not whether AI should exist, but how its development and use relate to the human vocation toward truth, goodness, and communion. In this context, the Church also stresses that its contribution is strictly instrumental, that it cannot replace the embodied intelligence, moral discernment, and interpersonal presence upon which authentic ecological stewardship depends.

But the risk is not merely that AI will fail to save the planet. As Pope Leo puts it, the urgent question of our age is “not merely what AI can do, but who we are becoming through the technologies we build.” As the modern world increasingly conditions us to experience reality through screens and curated media that present us with safe abstractions, the danger is that we become indistinguishable from the machines we have made.

The path forward is not to out-compute these machines but to out-love them: to return to the soil, the seasons, and the practices of re-embodiment. Whether it is hunting, fishing, gardening, keeping animals, foraging, hunting, walking without earbuds, or fasting and feasting in tune with the seasons, these habits strengthen the capacities for attention, gratitude, and restraint that make genuine stewardship possible.

Machines will not save creation. Only creatures who love can. And we humans love because we are embodied and vulnerable, embedded in stories and places and communities.

To avoid becoming the machines we have created, we must recover what they now and always will lack: attention, awe, joy, humility, courage, and sacrifice—and with them the capacity to receive the world as the gift that it is rather than mere material to be manipulated.

Church Life Journal | Articles

Tags: Can’t Care Creation