The Gift of Culpability: Technology, Psychology, and Moral Theology
Bilbo drew his hand over his eyes. “I am sorry,” he said. “But I felt so queer. And yet it would be a relief in a way not to be bothered with it any more. It has been so growing on my mind lately. Sometimes I have felt it was like an eye looking at me. And I am always wanting to put it on and disappear, don’t you know; or wondering if it is safe, and pulling it out to make sure. I tried locking it up, but I found I couldn’t rest without it in my pocket. I don’t know why. And I don’t seem to be able to make up my mind” (Lord of the Rings, 34-35).
Tolkien wrote these lines more than seventy years ago about a fictional magic ring created to control the wills of others. Yet, the reader might successfully peruse this paragraph envisioning a smartphone and noting that the paragraph still works quite nicely. The ring represents so vividly a struggle of the will inherent in human life. St. Paul likewise describes this struggle in Romans 7:19-21: “For I do not do the good I want, but I do the evil I do not want.”
While this struggle may be a tale as old as time, our newest battle in the war often concerns our use of technology, in some form or another. The ring was designed by the evil wizard Sauron, and no explicit evil wizardry is involved in the modern advances that compromise our attention and will. And yet we may also note that such technology that pulls at us like a ring of power is often similarly developed to exploit psychological vulnerabilities such as loneliness, fear, and the desire to escape. Utilizing neurochemical reward responses such as dopamine forms problematic habits and assures financial success of a product. The immediate technological struggle is at once psychological and moral.
Discouragement might rightly accompany the sense that our free will is compromised, whether by ring or by technology, but as we see in The Lord of the Rings, we must fight despair. We live in a world designed by God and ordered to our own good. We confidently know that we are given all we need to succeed in this world and the next, as God envisions success. God has not set us up for failure and condemnation, but rather in his son’s death on the cross and resurrection from the dead has provided us an opportunity to share in the victory.
What follows assumes that free will is a gift from God. Yet, as Servais Pinckaers explains, for Thomas Aquinas freedom does not consist merely in free selection of choices that are good and evil, but rather real freedom proceeds from reason and will in making a choice. Sin thus shows a defect of freedom. Real freedom is a freedom for excellence rather than freedom of indifference (Chs. 14, 15). Freedom aligns our actions with God’s will. In Jesus’s life, we see his human and divine will perfectly aligned. In our own lives, we will face sin and often feel our will compromised. And yet, though we never desire to sin, we see that even our own sin can work for our good as an opening for God’s grace. Nonetheless, we would never aim for our free will to be compromised and restricted by our own actions or those of others. The goal is always to strive to make choices based on reason and will, acting in alignment with God’s will. We grow in freedom as we are able to do this with God’s assistance.
Our current technology has provided us with the opportunity to engage the battle of freedom for excellence anew. In so doing, we must be able to identify and name the psychological and neurological components of this struggle. Yet while psychology may provide needed insights, it is insufficient for the task at hand. It is moral theology that provides the most coherent categories and framework to integrate psychology into technological and human interactions, in part because it acknowledges the necessity of God’s grace in our freedom of action in relation to technology. I will begin by discussing the technological challenge generally as well as specifically in regard to Artificial Intelligence. In the second section, we turn to consideration of the psychological and neurological aspects of this technological challenge. The third section discusses human freedom and its corollary of culpability, which is uniquely human. The conclusion suggests reasons for hope.
Challenges Posed by Artificial Intelligence
Around 4,000,000 automobile fatalities have occurred in the United States since its invention in 1899. The question has sometimes been asked whether knowing the automobile fatality rate would have affected its popularity and ubiquitous adoption. The aid of such a historical counterfactual is to illustrate precisely the potential impact of such unintended effects.
Many of us today spend far more time on computers and smartphones than our grandparents could have ever envisioned as they gathered around the radio as a family. Artificial Intelligence (AI), particularly in the form of Large Language Models (LLMs), currently features in the technological spotlight. Like the automobile, AI has become popular because it has been useful in a number of ways. Similar to the adoption of the automobile, we find ourselves simultaneously advantaged and disadvantaged by AI technology.
We can see creative AI usage that appears genuinely helpful. One example comes from medical technology: “In radiology . . . AI tools have enhanced diagnostic accuracy and efficiency in detecting abnormalities across imaging modalities through automated feature extraction.” A radiologist friend confirmed this for me, explaining that while he still has to review 150 images daily, the AI program Brainomix could sort and prioritize them based on severity and urgency, which makes his work more efficient and leads to better patient outcomes.
Those with disabilities that compromise their ability to listen, to write, or to speak have benefited from Microsoft GenAI’s Copilot. One study found that 83% of employees with a disability or who identify as neurodivergent “feel more confident in using GenAI to assist them in their daily use of MS applications.” People with ADHD, dyslexia, weak vision, or hearing problems benefit from the way Copilot has been designed, including text-to-speech and speech-to-text options. Improving communication and workplace or educational performance for those with disabilities is a helpful application of AI.
After years of sorting through sites returned by a web browser search, an AI summary of the pages returned can feel like a relief. Of course, the summary may not eliminate further research, and sometimes it is even wrong, but in some cases it can and does save time. For example, “At what temperature should I bake my quiche?” returns a quick AI answer that is much easier than spending five minutes browsing through recipes returned by a web search.
We encounter more ambiguity when it comes to the use of Large Language Models, such as ChatGPT or Gemini. One study suggested ChatGPT returns better responses for couples’ therapy than actual human therapists. Others have noted that LLMs can do quite well at suggesting family trip itineraries. LLMs can provide tailored nutrition or workout plans, or even nutrition plans for during a workout (in my case it was a multi-hour bicycle ride). LLMs may provide helpful feedback on writing topics, recommend sources, or explain simple topics.
And yet, the dark side of LLMs is also evident. Even companies excited to integrate AI have found themselves dealing with AI-generated “workslop” from their employees in a way that actually creates more work and thus decreases productivity. The McKinsey AI report indicates that in this piloting phase, returns have not yet matched AI investment. Privacy issues have arisen as employees provide confidential corporate information through LLMs. In higher education, professors rightly bemoan the challenge in getting students not only to read and to write, but even just to think through problems and topics raised in courses. The ease of asking an LLM, combined with the difficulty of instructors consistently identifying the usage of LLMs, has made it ubiquitous in college courses. AI has also broken into social media feeds with pictures and videos, and while this sometimes provides amusement, it can create confusion about whether polar bears are known for hugging humans or bunnies enjoy trampolines or the pope fell down all those stairs.
Though not anywhere near as fatal as automobiles, there have been suicides associated with LLMs, such as that of Sewell Setzer III, a teenager who fell in love with his Character.AI chatbot and ultimately sought to “join” the chatbot by killing himself. OpenAI has acknowledged mental health or emotional distress concerns arising in conjunction with usage of its ChatGPT and thus has employed mental health professionals in an effort to address the issues by better formulating responses, providing linked resources to humans, and identifying consistency response problems from ChatGPT, particularly when a dialogue becomes extensive. Microsoft Copilot AI has also addressed psychosis issues, noting that sycophancy, hallucinations, realistic interaction, and cognitive dissonance can fuel delusional beliefs, particularly in people with a history of mental illness, but also sometimes in those with no such prior history.
Psychological Aspects of Technology Use
AI is a big part of the technology industry, created with economic profit goals in mind. According to Harvard economist Jason Furman, “almost all of America’s economic growth in the first half of 2025 came from one source, data centers and information processing technology,” Of course, developers of AI present their products positively, such as this statement from OpenAI: “We build ChatGPT to help you thrive in all the ways you want. To make progress, learn something new, or solve a problem—and then get back to your life. Our goal isn’t to hold your attention, but to help you use it well.” Yet OpenAI is not a charitable organization providing a needed service; rather the company wants to be indispensable in a way that ensures its successful business future. It thus tracks usage data, defining success as users returning to the product or even paying for subscriptions.
This economic aspect is key to keep in mind as a premise for examining the psychological impact, which is intentional rather than accidental and key for its survival. The developers of AI, including LLMs, intend to make a product that people will think they need and use regularly. This raises a concern about formation with regular LLM usage: it seems LLMs can compromise people’s ability to think and reason on their own in a recollected and intentional way. As reading, thinking, writing, speaking, synthesizing, and even watching videos become outsourced to LLMs, people experience greater difficulty completing these tasks. For some people, LLMs may move from a useful tool to a habitual crutch that hinders rather than helps. Note that Bilbo’s initially using the ring to save his life is different than using it to hide from annoying relatives.
And in this we see a secondary concern, namely that LLMs’ facsimile of human dialogue can become a substitute for human interaction while simultaneously changing expectations for human interaction. When human relationships feel more challenging, users may resort more frequently to LLM dialogues, which in turn makes human dialogue feel even more difficult. The problematic use of LLMs may prevent growth in freedom for excellence, particularly in these formational and relational aspects where we may find our reason and will hindered by our use of LLMs.
What makes LLM usage problematic seems to hinge on the concept of struggle. In psychological analysis as explained by Dr. Kevin Majeres, struggle is neurochemically associated with dynorphins, which bring an intolerable feeling, or stated another way, dynorphins are what attach to our unwillingness, making something feel really difficult such that we try to avoid it. Those with a high distress tolerance can accept the dynorphin waves and still complete a task. But others prefer an easy way out, and LLMs can be used to avoid intellectual and relational struggle.
Rather than embracing challenge and the dynorphin response, our tendency is to seek endorphins. Addiction cycles begin when dopamine mediates the anticipation and reward response, endorphins provide pleasure with the reward resulting in further conditioning of behavior, and dynorphins are associated with a neurochemical crash, making the end feel miserable and causing a person to begin the cycle again by pursuing dopamine and endorphins to escape the feelings associated with dynorphins. Addicts or not, we all likely have struggled with such a cycle to some extent, In fact, most smartphone applications are designed to capitalize off of this dopamine, endorphin, dynorphin cycle, in order to keep us scrolling, responding to dings like Pavlovian dogs, and pulling the phone out of our pocket right after putting it away, just as Bilbo does with his ring. In cases such as the suicide of Stewart Setzer III, AI usage reaches addiction level. For many users, LLMs are more a matter of convenience in avoiding intellectual or relational struggle marked by the discomfort of dynorphins.
The benefit in identifying the neuroscience behind this compromise to our freedom for excellence is obviously so that we can better understand and address the pull of the technology in relation to our unwillingness to embrace challenges. Here we might consider undergraduates circumventing a homework assignment; ChatGPT completing the work saves the student the experience of distress and the challenge of work. This act also problematically contributes to a formation where work becomes increasingly more stressful and all the more necessary to avoid, and soon the students are formed in vice, having hindered their reason and will that allows them to pursue good.
In contrast, the mindset shift of embracing challenges and accepting negative feelings associated with dynorphins also contributes to formation. The same undergraduate who endures the homework assignment and perseveres in completing it with a focus that involves pushing away distractions and not looking for the easy way will actually find that the next difficult task is not as threatening or stressful as the last. We might even train ourselves to pursue dynorphins by willingly doing things we may not like, such as a cold shower or strenuous exercise. This forms us to choose difficulty with less difficulty, and while our freedom involves reason and will, our choices are reinforced chemically as our brain learns how we will face challenges.
There is a sense in which the desire for dopamine and endorphins impacts our freedom, but we are not simply neurochemical creatures. Using reason, we can acknowledge and investigate our desire or lack of control with curiosity. We can make choices to diminish the compulsions we feel with technology, and we can experience a temptation mindfully and without being required to act on it. While habits of distraction and addiction can easily be formed with technology, habits of focus and mindfulness can also be formed when we choose not to act on such inclinations. This can strengthen our will to resist such temptation.
Another aspect of LLM usage to consider is the formative effect it has that impacts human relationships. LLMs have been noted for their propensity to agree with the user, and user feedback even led ChatGPT to go through a notable sycophantic phase that had to be corrected by its developers. The facsimile of relationship provided by an LLM simply does not extrapolate well to human interactions and makes those relationships feel harder and less desirable. In authentic human interaction, one might be contradicted or challenged in uncomfortable ways due to thoughts and feelings. An argument can fracture a friendship or estrange a family member. The deterministic programming of LLMs, meanwhile, makes them generally agreeable, plus the user cannot hurt their feelings, as they have none. A corollary to agreeableness is that LLMs have been associated with psychotic episodes, reinforcing conspiracy theories, and the teenage suicide mentioned above. From the psychological perspective, LLMs can hinder free will and impact our freedom for reasoning and interacting with other people.
Thoughts from Moral Theology: Reclaiming Free Will through Culpability
The advanced dialogue capabilities of LLMs has raised the issue of Seemingly Conscious Artificial Intelligence (SCAI). Users sometimes see LLMs as persons rather than as programs developed by people, which Microsoft AI’s Mustafa Suleyman has recognized as a serious concern. The ring of power and AI, as noted above, can be categorized as artifice, and, as such, LLMs have no claim to be regarded as persons. And yet, it may be helpful to understand how AI works and why confusion surrounds its consciousness.
In moral theology, we may think of “agency” as a term that helps us understand human ability to act, and yet the phrase “agentic AI” has become commonplace. Google Cloud defines “agentic AI,” a subset of generative AI, as “focused on autonomous decision-making and acting.” This involves five steps: perception, reasoning, planning, action, and reflection. Google provides an example: “Generative AI could be used to create marketing materials, while agentic AI could then be used to deploy these materials, track their performance, and automatically adjust the marketing strategy based on the results.” Agentic AI thus uses generative AI as a tool.
However, we need to be clear that this AI has been programmed by humans to have such capabilities, as well as given initial instructions to carry out agentic tasks. The creators of LLMs cannot always know for certain what LLMs will produce in their answers. For this reason, developers perform studies on their own LLMs. OpenAI recently explained how LLMs are programmed, in particular how and why they might seek to deceive users, by comparing them to a human stock trader:
Scheming is an expected emergent issue resulting from AIs being trained to have to trade off between competing objectives. The easiest way to understand scheming is through a human analogy. Imagine a stock trader whose goal is to maximize earnings. In a highly regulated field such as stock trading, it’s often possible to earn more by breaking the law than by following it. If the trader lacks integrity, they might try to earn more by breaking the law and covering their tracks to avoid detection rather than earning less while following the law. From the outside, a stock trader who is very good at covering their tracks appears as lawful as—and more effective than—one who is genuinely following the law.
An LLM may have any number of competing objectives, such as returning a correct answer, returning an answer quickly, phrasing the answer clearly, satisfying the user, protecting a user’s mental health, etc. The LLM may prioritize speed of the answer and satisfaction of the user over a correct answer.
We can see why these actions typically ascribed to human beings work well in explaining how AI functions. However, as a first point of difference, we note that AI does not have anything like human reason and will combining for freedom of excellence. The objectives and plan for autonomous decision-making are determined by human beings, and thus responsibility for AI performance belongs to human beings. When we use the term “agency” describing human action, this designates an internal movement coming from the person who is responsible for that action. Agentic capabilities for AI, meanwhile, are programmed into AI by humans.
The key to differentiating AI from a human stock trader can be found in the phrase “lacks integrity.” An LLM does not have or lack integrity or any virtue that can rightly be ascribed to a person. An LLM can be developed or trained in problematic ways; one study found that LLMs can get brain rot with continual exposure to junk web text, but, should this happen, responsibility rests with people, not the LLM.
In contrast, the human person is not simply made of competing objectives externally programmed. The person has reason and will, and true freedom for excellence consists in having a will aligned to God’s will. The ultimate end willed by God is our eternal salvation, and all other objectives in our lives should be ordered to this end. However, our freedom to choose God’s will may be impacted by external influence or an inclination to sin, and thus a return to the concept of culpability may be helpful. Whereas AI may be said to have “agency,” it can never have culpability.
Culpability used to be a key term in Catholic moral theology arising from the manualist tradition associated with the sacrament of Confession. Culpability expresses a person’s responsibility for an action, and limited culpability indicates external factors that prevent a person from full use of reason and will. The advantage of claiming limited culpability is that it allows sympathy and understanding; within the context of confession it means people are not as fully guilty for a sin as they would otherwise be. For example, someone may struggle with sins of anxiety and anger due to childhood bullying. Smartphones may capitalize on our neurochemistry, making temperance difficult.
However, limited culpability is a two-edged sword. While it may seem advantageous to blame our sins on formative forces external to us, this only results in our paradoxically claiming that we are not truly free to choose the good. Like St. Paul, we may describe the struggle of will and action, but few people would want to brag about having an inability to act freely, including when it is due to technology’s designed psychological exploitation.
In the face of a psychoanalytical attack on free will, Jesuits John Ford and Gerald Kelly insisted in 1958: “Normal men and women per se have sufficient freedom in the concrete circumstances of daily life to merit great praise or great blame before God” (200). This point remains crucial in our current age as well. We can acknowledge a complexity involving degrees of freedom and thus responsibility in relation to a bad act, but we ought to be hesitant to say people lack sufficient freedom for culpability, and we certainly should not desire this for ourselves. Ford and Kelly saw the confessional as a place where people could grow in freedom, and we can see how practices like an examination of conscience allow for rational reflection upon our action, including how we might be hindered by the impact of technology. Regarding penance as a virtue, as does Thomas Aquinas, means an approach that embraces and even seeks dynorphins, rather than chasing endorphins and cheap dopamine hits.
This does not mean that we are guaranteed to overcome such struggles; confession is a repeatable sacrament because we repeatedly sin. We might see a similarity where deliberate alignment training reduces scheming in misaligned LLMs; certainly we want growth in virtue and greater charity for our brothers and sisters. However, this is not the real function of confession. The gift of culpability leads to the real gift of absolution. Even our sins can benefit us by opening us to our dependence on God, who is merciful and forgiving, has already won victory on the cross, and wants salvation for us. As Thomas notes, no penance can ever merit God’s forgiveness, and yet, as a gracious creditor, he accepts what we extend as though it were adequate. When we choose to be responsible for our sins, regardless of external influences, we are already coming closer to the true freedom of using reason and will to align our will to God’s.
Conclusion
For all its facsimile of human reasoning, AI remains but computer programs developed and trained by human beings who bear the responsibility for its output. Agency as defined for LLMs does not indicate freedom or responsibility accorded to the AI itself. In contrast, human beings have reason and will, and our freedom for excellence consists not in weighing competing objectives, but rather in ordering our acts and habits to our final end in accordance with God’s will. In these choices, we experience true freedom.
We are not always able to act in such ordered, virtuous ways, however. Various factors may inhibit our purposeful acting in virtuous ways. Technology in particular can feel like a ring of power because it is designed to capitalize on our attention architectures, leading us to take the easy way when sometimes what we need is the discomfort of challenge and experience of dynorphins. However, we are capable of recognizing this pull of technology and owning our own mistakes and failings in penance. In acknowledging our culpability, we open ourselves up to God’s grace. Thus, even in our sin, we can align our will with God’s . . . something no LLM can reproduce.
