On AI and Ministry
The argument below is about discipline rather than abstinence. It’s only fair, then, to say what discipline produced this piece. Initial research on the technical and empirical questions—the architectures of contemporary consumer AI tools, the published literature on hallucination rates, the energy and water studies—used Claude. I drafted the body myself. I then allowed Claude to help with refinement, I manually checked every reference against its source, and I iteratively asked Claude to help me tune citations where I found gaps or weak attributions. I had Claude create a Markdown file of all citations with a verbatim quote that anchored it and turned that over to Gemini to independently confirm. The TL;DR below is largely a summary generated by Claude. The final review was then mine. The shape of that process is, in many ways, structurally close to what I think responsible use of these tools looks like in scholarly and pastoral work generally—and the same shape, much less extensive at every stage, is what I am proposing is acceptable for some elements of sermon preparation, while leaving the load-bearing work, the exegesis and the actual proclamation, to the human preacher. If one of your suspicions about generative AI is that it can only produce slop, then, assuming this is not slop (and please don’t tell me if it is), I hope you will update your priors. If your sense is that generative AI can only produce overly polished work, well—I’m sure I can point you to a few suboptimal arguments and quite a bit of overly self-important prose.
TL;DR
The argument: A lot of clergy and scholars treat generative AI as uniquely unethical, but I don’t think this is true. Still, the place to begin isn’t with permission, but rather with arguing for the hard limits of its functionality. Generative AI has no business writing a sermon, a prayer, or a liturgy, and once those lines are firm, the we can begin to grapple with the ethical concerns about other uses of AI, most of which turn out not to single AI out at all.
What prompted it: Two things landed close together. The St. Dunstan Pledge asks preachers to swear off AI anywhere in sermon prep, and Pope Leo XIV’s first encyclical, Magnifica Humanitas, names our “pressing duty to remain profoundly human.” I agree with the spirit of both, and I extend the bright lines to even more domains of ministry than the pledge.
The hard lines: A nonperson cannot pray, and it does not have a life of faith, so it should never write the words a congregation actually prays or hears proclaimed; the same goes for liturgy and for pastoral-care conversations. More broadly, students should be kept away from generative AI entirely while they’re still learning the craft, because that’s the stage where the capacities to use it well are either formed or lost.
Why I didn’t sign: The pledge draws its line between “nongenerative” and “generative” AI, a distinction that no longer holds, by and large. Transcription apps, autocorrect, the grammar checker in Word, search results behind your research—these already run the same transformer technology as ChatGPT. The real distinction isn’t technological but functional: what has to be done by a human being, and what can responsibly be handed off.
The ethical objections: The worries about copyright, exploited labor, water, and energy are all real—but not one of them is unique to AI. Academic publishing already runs on a mountain of unpaid scholarly labor; raising beef and growing alfalfa dwarf data centers many times over on water; your phone and the electric-vehicle boom drive far more destructive mining than ChatGPT does. Singling out AI may owe a great deal to availability bias and to a quiet suspicion that the technology is somehow “cheating.”
On data centers: Local bans feel like righteous resistance, but they mostly shove these facilities toward the communities least able to fight them. The better tools are the ones we already have—regulation, litigation, the Clean Air Act, building political power and national organizing—and in places like Memphis they are, slowly, working.
In the end: We should seek to embrace discipline rather than abstinence, and harm reduction rather than harm elimination. Draw the bright lines where they actually belong, hold every technology to the same scrutiny instead of fixating on the frightening new one, and keep the load-bearing, incarnational aspects of ministry fully human.
The Post Itself
If you spend time on social media, especially in more progressive social media places, you will quickly find quite a bit of skepticism, critique, and even fear of generative AI, often with chat-interface LLMs serving as the synecdoche for all such technologies. As with so much on the internet, there is some really thoughtful and nuanced critique mixed with not a little, shall we say, bombast and polemic, with phrases like “plagiarism machine” and “stochastic parrot” and “fancy autocorrect,” all the way up to “clanker,” that reference to soldiers in the Clone Wars droid army.
While often less overtly inflammatory, I’ve noticed as much or more heat coming from professionals in academic and knowledge work and adjacent fields, especially the more humanities-adjacent among them. No doubt a significant part of this resistance comes from the combination of often uninformed tech tycoons and financiers overhyping capabilities, and gimmick-chasing management and university administrators pushing ham-fisted programs to integrate AI into every aspect of life and work to seem “relevant.” No doubt the completely idiotic choice to position AI as a tool that will make human knowledge work obsolete has not made many people in those fields fans of the technology.
The root of this frustration specific to knowledge work, going beyond the fear of being automated out of jobs, seems to be that AI will lead to the slopification of our respective fields: hallucinated citations, bad arguments, banal writing. Related would be concerns about what it does to our capacities: will overreliance on generative AI tools diminish our capacity to write, to research, to think? If an ethical and prudent integration of generative AI into the academy and newsroom exists, will it be a gift-of-the-magi situation where the very skills needed to hold it in check, or develop future scholars able to do this, are diminished by every better frontier model?
I will say, proleptically, I think the latter possibility—the danger generative AI poses in early stages of learning the craft of knowledge work—is very real and we should err much more on the side of total exclusion. But for those who have climbed to a certain level of expertise and professional competence, I worry much less about the loss of such capacities and would be much more interested in exploring what opportunities generative AI may offer for expanding the reach of that expertise and skill. In fact, as I type this, a blurb came up on NPR pointing out that scientists are finding AI helpful in automating repetitive lab work—something I think there could be analogues for in other academic fields.
Now, in my corner of the knowledge work world—and I do think clergy exist at least in the environs of knowledge work—concerns about generative AI go much deeper, requiring us to ask questions not just about the quality of output but about the degree to which human production is integral to the work we do. I don’t want to diminish the way this is an open concern in a number of fields, especially that field titled “humanities,” but in many of the fields marked by greater methodological pluralism and metaphysical and anthropological agnosticism, such questions will remain, and perhaps must remain, ad hoc and personal. The difference for Christian ministry, and especially those traditions that emphasize incarnation and sacramental presence, is that this tradition has emphasized—especially in the turn toward personalism—the importance of the encounter between the inexhaustible depth of persons, the human being as the location of the Spirit’s work in inspiration. Formation has never been reducible to the most optimized knowledge dump or mere moral exhortation. It is our seeking to pass on the way of being found in Jesus Christ, the incarnate God, a way of being encountered and apprenticed person to person, human to human, guided and made possible by the Spirit.
There is a sense then that the actual human effort, the struggle at times, to produce the work of ministry—to produce sermons, prayers, liturgies—is part of what makes them valid. I share this intuition (although I also worry at times and think we need to hold in check a tendency to romanticize and sacralize toil for toil’s sake, and often judging others with a higher standard than we hold to ourselves). While I can’t fully express why, I feel deeply that a nonperson should not produce prayer texts because it cannot pray, and that a nonperson should not proclaim and interpret the Gospel for the particularities of our times and communities because it does not live in a particular time or community. (This does, of course, raise the question of what should happen if AI ever does achieve self-reflexive self-consciousness, and when or if the “personhood” threshold could be crossed, though I don’t think we are there yet.)
For me, then, a hard line clergy should not cross is that generative AI should not produce the substantive text that is actually proclaimed to the gathered assembly, or even manufacture illustrations or anecdotes for use in sermons, since those do not actually have the imprimatur of having been living encounters with the Living God. Generative AI should never write a prayer for the Christian community. Generative AI should never write a liturgy that is actually used by humans to worship the Living God who meets us in, and integrates us into Godself through, the incarnate human being Jesus Christ.
This is, I believe, largely the point behind the recent St. Dunstan Pledge for Preachers, a document signed by many, many people whose ministerial skill and integrity I have the utmost respect and even awe for. It is interesting, perhaps providential, that this pledge came out when it did, seeing that the most significant institutional Christian intervention on AI to date came on May 25, 2026, when Pope Leo XIV promulgated his first encyclical, Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence, signed on May 15, the 135th anniversary of Leo XIII’s Rerum Novarum. Pope Leo frames this as a “change of era” (para. 6) that imposes on Christians “the pressing duty to remain profoundly human” (para. 15) amid new forms of dehumanization. Earlier that year, speaking to the clergy of the Diocese of Rome, Pope Leo had already told priests directly that they should resist “the temptation to prepare homilies with artificial intelligence,” because “to give a true homily is to share faith” and AI “will never be able to share faith” (though, ironically, it is perhaps my engagement with that archpessimist about human capacity to reach the divine, Karl Barth, that leads me to think this too sweeping—God can speak through a dead dog or even a Communist after all). This pledge wisely adds the layer of pastoral responsibility and the reminder that we, the ordained or licensed, not a computer, are entrusted with certain duties by oath. One may see what I wrote above and assume I would be a signatory, committing that
I will not use generative AI to write my sermons, homilies, or reflections. At no point in the brainstorming, planning, drafting, or writing of my sermons will I use ChatGPT, Claude, Copilot, Gemini, or any other generative AI of any type available to me. I will use my own words. I will quote words written or spoken by other people, not produced by generative AI. I will find and evaluate my own sources, and will cite them. I have been entrusted with the task of preaching the Gospel, the Good News of Jesus Christ, and this requires my effort, my prayers, and whatever skill I may have. I will use these as best I am able to glorify God.
I agree substantially with the sentiment of this pledge—with the idea that we who are entrusted with spiritual authority should not be outsourcing what is one of our most essential functions, that of interpreting the Living Word of a Living God for a living community of human beings. And yet I found myself unable to actually sign it, and reflecting on why occasioned this long, probably overlong, post. I do indeed hope that this will not be read as in any way casting aspersions on those who did sign the pledge, a pledge that, again, I share considerable agreement with both in spirit and in terms of the particulars of implementation. I hope that my response is received as within, not against, the larger conversation opened both by the pledge and by Magnifica Humanitas. My disagreement, such as it is, is a disagreement within a project I substantially support.
The first concern I have is with the distinction drawn between nongenerative AI and generative AI:
This pledge does not prevent me from using non-generative AI in its many functions, or from using the great variety of accessibility tools that technology provides:
I can record voice memos of my drafts, and have non-generative AI transcribe them.
I can use text-to-speech technology to have a document read aloud to me.
I can use non-generative AI to get suggested corrections to my spelling and basic grammar. I can use a basic search in Google (not the AI Overview) and in library catalogues to find quality resources; these searches use non-generative AI.
I can use my podcast app or YouTube account to find quality commentaries and resources; the algorithms use non-generative AI.
The instinct to still allow assistive technologies that won’t risk producing actual sermon text is sound. The problem is that many if not all of the things described above are increasingly done via generative AI, even if at one point in the recent past they were indeed done using nongenerative AI. OpenAI’s Whisper, the transcription tool used by many third-party voice-memo and transcription services, is per OpenAI’s own description “implemented as an encoder-decoder Transformer,” the same generative architecture family that powers ChatGPT. And Apple’s own on-device dictation has, since iOS 17, used “a new transformer-based speech recognition model,” with the iPhone keyboard now running a transformer language model for autocorrect by default. The grammar checker built into Microsoft Word has been a neural-network model since 2020. Google’s MUM, deployed into Search well before AI Overviews existed, was announced by Google itself with the explicit claim that “MUM not only understands language, but also generates it.” And while recommendation systems—YouTube, Spotify, podcast apps—do not generate text or audio in the way an LLM does, they are, per Google engineers’ own peer-reviewed account (Covington, Adams, and Sargin 2016), deep neural networks in the same family of technology that powers large language models. Even if these tools are not actively generating what appears to us, via a chat interface, as novel content, it is often the same technology, or very closely related ones, under the hood. Even the colloquial use of “the algorithm” to describe how suggestions are curated, with its implication of a fixed and predictable rule-based system, no longer describes how these systems actually work. As the philosopher Patrick Grim summarizes the state of the technology in his recent Great Courses course guidebook for Understanding Artificial Intelligence, “unsupervised neural net learning... plays a significant role in controlling fraud and is at the core of recommender systems” (Grim 2026, 27). Today’s systems are themselves constantly learning and adapting based on previous interactions, and the relationship between input and output is shaped by learned parameters rather than human-coded rules—in other words they are not just following rules but generating novel content in the form of the pattern of recommendations.
If the architectures are the same, the categorical line cannot be drawn at the level of the technology itself; it has to be drawn somewhere else. I think the better distinction is functional, between what tasks must be done by an actual human being and what can be outsourced to digital tools. I contend that my expectation that the content that is preached—and whatever scaffolds it in the moment, either the outline or final manuscript—be created by an actual human being aligns with the spirit of the pledge. Using digital technology for items 1–3 is relatively straightforward and noncontroversial. And yet the acknowledgement that, at least without great effort, you cannot avoid generative AI in search results and recommender systems forces us to reconsider why exactly chat-interface generative AI should be off limits for sermon research, whether for initial bibliographic compilation, for finding examples and illustrations, or even, and more controversially and requiring much more discipline, potentially for some research summary and assistance with interpretation (I say controversially because this is where I myself begin to become uneasy and I worry the temptation too great to cross over into final text creation).
In my experience, generative AI is genuinely useful for tasks search engines are ill-suited to, such as assembling curated reading lists across a topic, surfacing potential illustrations once the central themes of a sermon have been set, or identifying who has written what on a given passage or question. Importantly, while this can seem to us to have significant functional overlap with search engines (even a neural net powered one), it does something different and one has to recognize that it introduces its characteristic failure modes that traditional search does not have. But for those of us who already know how to evaluate sources (and my hope is that this is part of a seminary education that is largely or completely generative AI free—read to the end on this), there’s a real enrichment. In both instances, the preacher still must engage in significant thoughtful work: doing initial exegesis, developing sermon themes, actually engaging the suggested sources, reading/watching/listening to the examples offered, digesting and synthesizing them, and then actually writing the sermon.
What, though, is the risk of this greater permissiveness? Likely the first and most pressing reason why people would oppose such uses of generative AI is the threat of hallucination. When people think of AI hallucination, they often think of fabrication—the presentation of plausible-sounding statements about ideas, events, or things that don’t in fact exist, often presented with maximum confidence by the model. Such hallucinations—manufacturing incorrect biographical details for people, suggesting a product that doesn’t exist, inventing a character or a work of art—were indeed a big risk in early generative AI models and when using them without any web search functionality. Such a risk may be annoying but hardly catastrophic when one is having bibliographies crafted or illustrations suggested, since both uses require you to go and actually engage with those resources themselves. If they don’t exist, you’re going to find out pretty quickly. Where you could get in trouble is taking scholarship summaries without citations at face value, which can quickly become problematic with fabrications.
That said, the risk of fabrications is much lower for frontier models, especially used in careful ways. With the ability to engage in web searches, generative AI is not limited to what it learned through training. Learning to write better prompts allows you to tell the model to say when it doesn’t know something rather than fabricating it—something some models have built in anyway. When web search is enabled and the prompt explicitly restricts citations to fetched sources, outright fabrication of citations—the early hallucination problem where the model invented plausible-sounding sources that did not exist—drops close to zero. What remains is misattribution: citing a real source for a claim it does not quite support. Stanford researchers found that even purpose-built legal AI tools—including Westlaw AI and Lexis+ AI—still hallucinated between 17 and 33 percent of the time, with misgrounding making up a large share of that figure (Magesh et al. 2025). That is significantly higher than the human baseline. A 2025 meta-analysis put the rate of major citation errors—cases where the source does not in fact support the claim being made—at 8.0 percent across peer-reviewed biomedical literature (Baethge and Jergas 2025). AI is not at parity. It is worse, in this domain, by something on the order of two to four times. Importantly, though, the Magesh study tested tools as they existed in May 2024, and these may well have improved in the intervening two years. In producing the citations for this essay—several dozen entries across legal cases, peer-reviewed papers, news coverage, and corporate filings—I caught only a few misattributions during manual verification. One person checking his own essay is not a study, but the experience suggests the rate today may be meaningfully below 17 percent. This is a real failure mode that anyone who chooses to use generative AI in any kind of research has to acknowledge, account for, and know how to mitigate.
But the gap is not a categorical difference. It is a difference of degree on the same kind of error, an error both AI systems and human scholars routinely commit. The answer is what it has always been: do the due diligence of going to cited sources and confirming they actually support the argument being made. That discipline is required regardless of who or what produced the citation (though, and I still do not recommend skipping the manual review step, one can reduce the risk of both fabrication and misgrounds even more when you explicitly direct the generative AI model to actually look up and provide verbatim quotes that support the use of the source).
Nor does enforcing human-only research solve the problem we call “error” when humans do it and “hallucination” when generative AI produces it. Misquotations and misattributions exist in exclusively human-produced content, even that found in respected academic sources. If our concern is ensuring accuracy and support for our claims, the answer is not closing off the possibility of generative AI—it is discipline in tracking down and confirming sources cited in the research we make use of, regardless of who or what created it. The payoff of all of this is that in choosing to use generative AI for research, sermon or otherwise, one must be clear eyed that the rigor required calibrates to the creative use: bibliography-building and illustration-finding are lower-stakes because their failure modes surface on contact, while research summaries and syntheses demand more vigilance because misattribution can quietly contaminate the human produced artifact downstream.
Allowing that smart integration of generative AI is about improving sermon research rather than time reduction, and that the risk of hallucinated or inaccurate information can be significantly mitigated by supporting people in learning responsible AI usage and the discipline of tracking down sources, one may still push back that generative AI has no place in sermon writing because of ethical concerns inherent in generative AI itself. The first concern, perhaps closest to what has been talked about already in relation to academic integrity, relates to how generative AI is often referred to as a “plagiarism machine.” If this label is correct, that is a serious concern indeed. But it’s also unclear what exactly the concern inherent in this phrase is. The first level, and easiest to address, is the fear that this technology makes it possible for anyone to easily and quickly produce content, potentially in the voice or style of other thinkers, that they can illegitimately pass off as their own. This concern is real but hardly novel (generative AI did not invent forgery!) and has a relatively straightforward solution: don’t do this. Don’t have generative AI entirely create something with minimal input or direction from a human and pass it off as entirely your own. (Though even here it is unclear if the category “plagiarism,” rather than simply dishonesty applies, since the primary harm in plagiarism seems to redound to the person whose work you’re passing off as your own and thus denying them the proper recognition. In the case of generative AI it is unclear who is being denied proper recognition—the weights? the chat interface? the parent company?)
The next possibility is that plagiarism resides at the level of generative AI’s output itself. In this model, the issue at stake is that the model relies on the writings of other authors to generate what it generates, but it does so without properly acknowledging these sources. There is a layer of truth to this claim: training an LLM requires massive, massive amounts of data. But where the framing of what an LLM does as plagiarism seems to run aground is what happens once the model is trained. There is no repository or reservoir of texts, outside of novel web searches, that an LLM is filing into to provide a query response. The LLM “learned” a massive set of linguistic relationships that it encodes as billions of parameters (the numerical values that encode the model’s learned patterns), allowing it to provide coherent responses to text inputs. Every response is itself novel, even when presenting information about a particular idea, argument, or figure. One way you may think about how it engages with that large corpus of data is like a student with superhuman memory and reading capacity who goes into a library, reads every book, internalizes most of the arguments into their mental models, and begins to emulate the style of the authors in the library. It would seem strange to consider this in itself plagiarism. One might object that this minimizes a real disanalogy of scale—a student reads at human speed and forgets most of what they encounter, while an LLM has near-perfect recall and processes at industrial scale. The disanalogy is real but does not change the conclusion. We do not categorize what someone is doing—learning, paraphrasing, plagiarizing—on the basis of their cognitive capacity. A human who has mastered the arts of memory is not committing plagiarism by remembering whole books. A research library is not plagiarizing the authors whose works it holds. Plagiarism tracks how output is produced and presented—whether it passes off the words and ideas of another as one’s own without attribution—not how much input went into getting there.
Now, there seems to be some evidence that some models have held on to or memorized large portions of verbatim text inputs in their parameters. This can be coaxed to produce verbatim text under certain extraordinary prompting circumstances. This kind of verbatim memorization is technically understood as a failure state—the model behaving in ways its designers do not intend—but “failure state” is not the same as “small problem.” Anthropic, the company that makes Claude, settled the largest copyright case in United States history in September 2025—a $1.5 billion agreement with authors including Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, after Judge William Alsup of the Northern District of California ruled that the company’s use of approximately 500,000 books downloaded from pirate libraries was not protected by fair use. A separate suit filed by Concord Music Group, Universal Music Group, and ABKCO on January 28, 2026, names Anthropic, CEO Dario Amodei, and co-founder Benjamin Mann personally as defendants and seeks over $3 billion in damages, alleging Anthropic illegally torrented more than 20,000 songs, lyrics, and compositions via BitTorrent from Library Genesis and Pirate Library Mirror. The memorization and acquisition problems are real and the legal consequences are real. What I am pushing back on is the further claim that this proves the technology itself is fundamentally a “plagiarism machine.” What the lawsuits prove is that training-data acquisition has often been illegal, which is a different and answerable question that is in fact being answered, and rightly so, in court. These settlements are themselves how the alleged theft is being remedied. One might press further—that even using a model trained on illegally acquired data participates in that theft. But that claim is precisely what the settlements adjudicate. Roughly $3,000 per work in Bartz is unlikely to move any author into the lap of luxury, though it is genuine remediation—and quite likely more remediation than these authors would have received from the legitimate licensing market. When publishers license to AI companies or to streaming-subscription services like Perlego, the revenue routes first through the publisher (which typically keeps the majority), then through royalty schedules calculated on net receipts, with the author’s share being whatever percentage the contract specifies of the residual. The arithmetic does not produce much per author, even when paid out faithfully over years. None of this licenses or excuses what Anthropic did. It does mean the rhetoric of “they should have paid the authors” loses much of its force once you notice the lawful market does not, in fact, pay them very much.
This last point tees up what I see as the strongest claim that AI could be considered a “plagiarism machine”: the sense that AI relies on data that has been stolen for the training itself, that authors or artists were not properly compensated for their work to be included in the training to begin with. This would be akin to, in our analogy above, learning that many of the books in our hypothetical student’s library have been stolen and the authors never got their royalties. At least in the context of academic output, though, this is more indicative of the rot endemic in academic publishing as an industry, a corruption that extends much farther back in time than generative AI. This is an industry that relies on a staggering volume of unpaid researcher labor to undergird the peer-reviewed journal system that it then turns around and paywalls with exorbitant prices. US-based reviewers alone donate over $1.5 billion in time per year, and globally, peer review consumes over 100 million hours annually—over fifteen thousand years of labor (Aczel et al. 2021). This is an industry that pays pennies on the dollar for monographs and returns operating margins above 35 percent for the major scientific, technical, and medical publishers—RELX, the parent of Elsevier, reported a 38.4 percent adjusted operating margin in its STM division in 2024, on revenue of £3.05 billion. By comparison, trade book publishing operates on roughly 10 to 15 percent margins. Often academic publishers are acting fully within their legal rights when they turn over their corpora to train LLMs, since they often require those who publish with them to hand over copyright for their work, allowing them to potentially use, and even profit off of, uses unforeseen by publishing contracts and to then avoid compensating authors bound by such contracts—Wiley confirmed in 2024 that it had already earned $23 million from such partnerships, with $44 million expected total, and that its authors would not be given an opt-out and would be compensated only if their contracts already provided for it, without disclosing the amounts. At the risk of an argument from silence, my guess is when under heat for unjust and exploitative compensation practices, one would find a way to shout from the rooftops the compensation amounts if they were, indeed, fair.
A further issue of exploitation relates to the conditions of workers, often in the developing world and for sub-subsistence wages and grueling hours, necessary for the training of new models. There is no question that this is going on. Underpaid workers, both in the US and in the developing world, are used to help keep our generative AI safe by being fed streams of rancid text about such things as CSAM and hyper-violence to help build the guardrails for generative AI output. And yet here, again, we have an instance of an injustice endemic to a larger system and industry. AI adopted the model of outsourced moderation from social media, which was already employing such people for these low wages and long hours in order to do content moderation (Roberts 2019). To argue that AI engaged in unethical labor practices and should be held accountable for them is essential. To claim generative AI as an industry is uniquely exploitative and should be avoided for that reason risks being disingenuous.
At the intersection of human and environmental concerns is the need for rare-earth elements and metals in the production of hardware for running generative AI. And the cost for this hardware is steep: the vast majority of certain essential minerals comes from the Democratic Republic of Congo. In addition to the effect on the environment from actual mining (a notoriously dirty industry), there are documented cases of child labor and worker deaths (see also Kara 2023). This is also a dramatically politically destabilized area, with mining revenues going to support factions in one of the twenty-first century’s bloodiest and most tragically overlooked conflicts. But the real issue here is not generative AI specifically but digital technology generally: your laptop, your smartphone—these are the bulk of what these extracted elements are going toward. Actually the fastest growing driver (excuse the pun) of this extraction is EV production, a product that most would consider a net good for the world. For the modern, environmentally conscious knowledge worker, the decision to use generative AI or not is a small fraction of the demand driving these harms. The International Energy Agency estimates that AI data centers will account for 2 to 3 percent of global demand for copper, silicon, and rare earths by 2030 (with gallium specifically, important for power electronics, projected at 11 percent), while clean energy—electric vehicles, batteries, grid storage—will account for between 40 and 90 percent of demand for lithium, cobalt, graphite, and nickel.
This then brings us to the really big thing in people’s minds: data centers and water and energy consumption.
Out of the gate, we have to acknowledge that generative AI, especially in the training stages, does consume, when viewed at scale, really large amounts of water and electricity. But it’s important to get clear what that scale actually means for individual users. You likely have heard that single generative AI queries use a bottle of water each, or that Google data centers in Chile would use more than a thousand times the water consumed by the entire population of that country—a claim Karen Hao, the author of Empire of AI, publicly retracted in late 2025 after a reader caught a units-conversion error in her original source that overstated the impact by three orders of magnitude. The corrected figure puts Google’s proposed water draw at roughly 105 percent of Cerrillos’s residential water use—still material in a drought-stressed region, but not a thousand times anything. These are scary numbers, especially combined with the fact that many data centers have been clustered in drought-prone areas, and combined with the already strong anxieties people have about coming freshwater shortages as a result of climate change. The problem, though, is that these numbers are based on bad math. The most current operator-disclosed estimate, from OpenAI CEO Sam Altman in June 2025, is approximately 0.000085 gallons of water per ChatGPT query—about one-fifteenth of a teaspoon—meaning a 500-milliliter bottle of water equates to roughly 1,500 queries, not one. Even using more conservative independent academic estimates that include the water embedded in electricity generation (Li et al. 2023), a 500-milliliter bottle of water still equates to somewhere between 10 and 50 ChatGPT-sized queries, depending on where the data center is located.
But here people may rightly respond that data centers have tended, for a variety of reasons—many to do with political ease and tax incentives rather than actual environmental suitability—to be built in places with already incredibly strained water systems like the arid desert Southwest. I agree that this buildout has been irresponsible. But if one’s real concern is mitigating the extremely strained water systems in these places, there are much, much bigger targets for irresponsible water use. Raising beef and dairy cattle, and growing alfalfa—much of it for export to places like Saudi Arabia, which banned domestic green-fodder cultivation in November 2018precisely because of its strain on Saudi groundwater—accounts for so much more water use that even maximalist estimates for data centers in these regions become rounding errors. A 2024 satellite-tracking study of water use across the Colorado River Basin found that cattle-feed crops—alfalfa and other hay—consume 32 percent of all water consumed from the basin and 46 percent of all directly consumed water (Richter et al. 2024). That is nearly twice as much water as all municipal, commercial, and industrial uses combined across the seven basin states that supply 40 million people. Data centers are a small subset of that 18 percent municipal-commercial-industrial share. If our real concern is mitigating the strain on water systems in arid regions, we are straining at the gnat of generative AI while swallowing the camel of American beef consumption.
The question of water consumption is inextricably linked to concerns people have about data centers, concerns that relate not only to water consumption but also electrical consumption and the other externalities borne by the communities where data centers get built. The visibility of the data center boom, I think, leads people to believe that the massive outbuild of data centers is primarily supporting generative AI use. And it is true that generative AI is the fastest-growing percentage of data center use. But be careful here: the fastest-growing segment of data center use does not mean the primary consumer of data center capacity right now. Imagine you have a kitten and an adult golden retriever. The fastest-growing consumer of pet food in your house is going to be the kitten, but the golden retriever is far and away the largest consumer of pet food in your house. We must begin by acknowledging that the majority of data center usage is for all the other internet things we do. Now, this does not mean we should dismiss that AI is becoming an increasingly large percentage of data center usage. Aggressive projections have AI approaching half of data center electricity by 2030, though we may have reason to assume these aggressive estimates are not going to come to pass; an earlier projection of nearly half by end of 2025 (de Vries-Gao 2025) was not borne out in 2025 actuals. At the same time, nongenerative AI data center use is also still growing. Even if we were to eliminate generative AI usage, new data centers would still have to be built at some point. All of that is to say, like most of the other issues at play here, generative AI is part of, and indeed contributing to, a system of exploitation and potentially non-sustainability. But this system is larger than generative AI, and as a system, most people are not nearly as publicly vocal about it.
And to get down to the actual per capita energy cost of AI, like with the question of mining, there are a number of mundane knowledge work activities that consume much more electricity than generative AI queries. OpenAI CEO Sam Altman has disclosed that an average ChatGPT query uses about 0.34 watt-hours. Independent modeling by Epoch AI puts the figure at roughly 0.3 watt-hours, which is, looking at Epoch’s notes, about as much energy as a single Google search cost in 2009. An hour of Zoom, the energy your computer uses while using a word processor—these sorts of activities use substantially more electricity than ordinary generative AI usage over the same periods of time. A Purdue–Yale–MIT study estimated that an hour of videoconferencing emits between 150 and 1,000 grams of carbon dioxide (Obringer et al. 2021). In fairness, the underlying network-energy methodology has been challenged in subsequent peer-reviewed work (Mytton, Lundén, and Malmodin 2024). Importantly, though, even if we take more chastened estimates that do not scale linearly with increased data usage, these other, more mundane uses are still likely using more energy than generative AI.
The one issue that is really worth bringing up is not the per capita use of resources, but the way in which data centers strain particular municipalities’ water and electricity infrastructure, and that systemic issue can’t be swept under the rug. Data centers tend to cluster in areas that have the least political will and power to challenge them, drawing from grids and water systems that were not built to support this kind of dramatic surge in usage, to say nothing of other externalities such as air pollution. In a particularly egregious example, xAI’s Colossus 2 data center complex outside Memphis has disproportionately affected a majority-Black community in an area already failing to meet federal air quality standards. The NAACP, the Southern Environmental Law Center, and Earthjustice have filed Clean Air Act litigation against xAI alleging that the company operates dozens of unpermitted methane gas turbines as primary power for AI training, which have the potential to emit, among other hazardous pollutants, more than 2,500 tons of nitrogen oxides annually without the pollution controls required by federal law.
But I also want to make sure this point doesn’t get lost: what xAI is doing will, hopefully, be found to be illegal. We already have the legal and regulatory tools to fight this kind of abuse—the Clean Air Act, EPA enforcement, civil rights litigation—and they are being deployed and may well ultimately prevail. The Memphis situation is not evidence that data center construction itself is unmanageable. It is evidence that some of the people running these companies are willing to break the law to build faster than the law permits, and that the legal system is, slowly, responding. This certainly reinforces the moral depravity of many of those heading our tech conglomerates. But if we were to only engage with technologies produced and supported by the virtuous, we would likely need to return to a time before wheels.
The bigger problem, it seems to me, in the face of the deserved and legitimate increase in public scrutiny about unfair and exploitative practices by data centers, is the reaction to this information has been inconsistently and hyperlocally applied moratoria, which is perhaps a nice way of saying NIMBYism. The places that have the political power and will to challenge these data centers, especially where building data centers actually makes much more sense (like my own state of Wisconsin, owing to the ready availability of water and cool climate), are seeking to exclude them rather than accept them with expectations of more responsible and sustainable practices. Given that data centers are going to continue to be built since the internet, and not only generative AI, lives in them, this has the ironic effect of pushing them to concentrate in otherwise unsustainable locations that lack the political power and will to oppose them.
And, as has surfaced several times, generative AI, or even digital infrastructure more broadly, did not pioneer the exploitation of communities with less political power to evade accountability. This pattern—facilities clustering where political resistance is weakest—long predates AI. The American hog industry has done this for decades: over thirty years, the North Carolina swine industry has concentrated in the state’s eastern coastal plain and in so doing “externalized severe environmental and health harms onto poor communities of color” (Miller and Longest 2020). This industry also produces air and water pollution that surrounding communities have lacked the political weight to refuse, but, unlike the scrutiny-producing novelty of generative AI, the mundane familiarity of pork has largely kept this an unremarked-on fact of American food production.
Having pushed back on what I see as the most common ethical rationales for a complete rejection of generative AI use, I want to reiterate that the pledge itself did not make this more expansive claim or ground it in more general ethical concerns. As I understand it, it is a call to refrain from generative AI use specifically in any aspect of sermon preparation as a matter of staying faithful to the obligations of one’s licensing or ordination vows. I do still think that, as I said above, my slightly more expanded list of permitted uses (using the technology to help track down and organize sources, allowing it to help find extant possible illustrations once the central themes of the sermon have been set by the homilist), combined with a rigorous expectation that the bulk of actually composing the sermon, and the entirety of what is presented to the congregation, comes from the human author, still responsibly fulfills the expectations of those vows or licensing requirements. I also share the pledge’s closing commitment to fair labor conditions and to clergy roles “reasonable for one person to fill.” The point of AI-assisted research, on my account, is not to shorten the time invested in sermon preparation or to offload responsibility but to direct that time toward a wider range of exegetical scholarship and more pedagogically rich illustrations than search engine searches, especially first page results, reliably surface.
But if one does want to make the more expansive condemnation of generative AI as per se unethical and for that reason argue that avoiding it in sermon preparation should fit into a larger rejection of it in any part of ministerial work, then I think the pledge does not go far enough. We need to go far beyond rejecting generative AI use in sermon composition to refusing to use any of the permitted technologies either, because they run on hardware that relies significantly more than generative AI on destabilizing, exploitative mining. We should refrain from writing our sermons on computers or brainstorming on Zoom calls because of the electricity used. We should avoid streaming our services to social media both because of the ways social media platforms have been complicit in the spread of misinformation and the exploitative practices they pioneered for content moderation. And ultimately we should refrain from using any academic journals, monographs, or commentaries to prepare our sermons because they are part of the known exploitative economy of academic publishing. As I hoped to show above, many of the ethical arguments against generative AI seem to turn on the fact that it is uniquely unethical, and, if that is not the case then we must consistently scrutinize our use of myriad technologies, digital and otherwise, that support the work of ministry.
One could argue, though, that this is not merely engaging a technology that has a questionable genealogy, but one that is causing active harm to communities right now. One could go further and say that, while noble, attempting to lean on regulatory power in a political environment in which the EPA has been gutted and engaging in record low enforcement action, an environment in which big money interests and corporations the administration favors seem able to flaunt the law, is a fool’s errand. In such an environment, an environment of total capture by the powerful, people may rightly say then that the only possible response is local resistance to data centers being built at all. Now, I’d say that the strategy that has been pursued successfully in some locales (e.g., Tucson’s 2025 rejection of the Amazon-backed Project Blue), that of bans on data center construction on the local level, is, as I will look at below, counterproductive and likely harming the most marginal people many of those championing this strategy would want to be in solidarity with. Moreover, it actually argues against the fear of total systemic capture: local municipalities have demonstrated the capacity to refuse the advances of these mega corporations. Were this not the case, it would seem one would have to face the harsh reality that the only real means of resisting this total legal and political capture would be some form of ecoterrorism, or at least resistance in the form of property vandalism and sabotage.
One does not, though, even have to look at the example of successful municipal bans to find an example of a legal and political system working to hold egregious violators accountable. This is not Mr. Musk’s first time stomping in Memphis, and his attempt to run the very same playbook with the Colossus 1 was reined in merely by notice of an intent to sue under the Clean Air Act. In a kind of ironic way, this demonstrates a certain resilience of the American rule of law, that the CAA has a failsafe whereby citizens can sue rather than relying only on regulatory enforcement, a mechanism built, it seems, for precisely the kind of attempt at corporatist and authoritarian takeover we are currently witnessing.
I contend that these concerted, sometimes slow and arduous, attempts to regulate and hold data centers accountable without seeking to ban them are superior to the in vogue strategy of hyperlocally applied moratoria. Even as such attempts are likely arising from genuine motives to protect and stand in solidarity with vulnerable populations, such intentions will likely backfire and become functionally indistinguishable from, and set a precedent for, the more sinister forms of NIMBYism. The places with the political power and will to challenge these data centers are frequently the very ones where building them makes the most sense. My own state of Wisconsin is just such a case: the cool climate and ready fresh water make it about as suitable a location as the country offers… and the opposition here has been fierce. Port Washington passed the nation’s first referendum requiring a public vote on large data center tax deals, and statewide the share of Wisconsinites who say these projects’ costs outweigh their benefits has jumped from 55 to 70 percent in a matter of months. I think we should be cautiously optimistic about data centers expanding into Wisconsin, precisely for those suitability reasons, so long as genuine environmental safeguards come with them. Port Washington’s referendum is exactly the political development I worry about, because a town that grows used to deciding on a case-by-case basis, rather than through open and contestable statutes, what it will permit within its borders may slip into deciding who belongs within them. Now, if we were dealing with a hypothetical situation where resource extraction or land development only made sense in a particular place, say a mineral could only be mined or crop could only be grown in one area, local moratoria could be morally justified. But if a developer or corporation has the option of relocating, as is the case with data centers, this has the tragically ironic effect of pushing them to concentrate in otherwise unsustainable locations that lack the political power and will to oppose them. One is therefore left with a few options: owning that this is actually NIMBYism and that the concerns of other, more vulnerable communities don’t matter or matter less than your own; a concerted, organized, nationwide moratorium on data center expansion; or seeking, as far as is possible, to distribute data centers in places where they can run most efficiently, with the least drain on local resources, and under the oversight of communities with the political will to regulate. Certainly for a Christian, the first option is out (to say nothing of the danger of opening the door to a precedent of residents of local municipalities getting ad hoc vetoes on who can and cannot live in them—a proposal that has historically been as likely to lead to Sundown Towns as environmental protections [Loewen 2005]). The second also seems like largely fantastical thinking, unless we are willing to do the truly hard work of reducing overall internet usage and not only generative AI—something unlikely given how many of these arguments against data centers are happening over social media by often “terminally online” personalities. Plus, if we truly have the political will to reduce national internet usage, we certainly have the will to enforce more safeguards on our digital infrastructure. In other words, once we have passed the Green New Deal (and I hope we do), then I’ll believe that a national campaign to reduce internet usage could be plausible. This leaves us with what is the difficult, but likely most realistic and just solution: building local political power to permit data center development only under strict environmental regulation and use and development of more efficient and renewable cooling and power technologies combined with a concerted state and national networking and organizing to ensure that the places data center developers may go to escape such development friction also have the capacity to hold them accountable.
Now, having spent far too much time doing something I wished not to have to do—appearing to align myself with Silicon Valley interests by making a chastened case for data centers—and coming back around to the large point I am making about generative AI usage, that it is not uniquely unjust or morally bankrupt: some may say “yes, all of that is injustice and exploitation, but we must draw the line somewhere.” How does what I’m proposing not just come down to a huge case of whataboutism? I think my primary response is to ask: but why this technology, this line, other than that it is new and somewhat frightening and we are therefore more likely to fixate on it because of the media spotlight feeding our own availability bias? It is not whataboutism to point out that what has been marked as a uniquely nefarious technology actually exists within a much larger, seemingly intractable, web of nefarity, one in which the material consequences of upgrading to newer versions of established technologies (say buying a new smartphone or tablet or laptop) is indistinguishable from adopting new ones. Given this larger background, one would have to shift to the kind of blanket refusal to engage any of the technology in this web or to shift from a harm elimination to harm reduction model. And if the latter, I am once again asking why generative AI specifically as the thing to avoid? One would save more electricity by reducing one’s Zoom usage than by eliminating generative AI. Shifting from cloud to local data storage would decrease demand for data centers much more right now than banning generative AI (although with the added unintended consequence of potentially increasing total carbon footprint for computing). One can do significantly more to ease the strain on the water supply by reducing beef intake by a quarter, to say nothing of eliminating beef from our diets. I don’t think generative AI is a panacea or miracle, and I have real concerns about it—even arguing for spaces where hardline abstinence should be the norm. But I worry that the often-aggressive reactions to generative AI come down to confirmation bias—finding data that affirms in this case an existing suspicion of the novel and uncanny—combined with a sense that the technology is somehow “cheating.” That second intuition has a recent precedent that many of us probably were unaware of and which seems quaint or even outlandish now. Matthew Kirschenbaum’s literary history of word processing documents the unease the technology produced in the 1980s. Gore Vidal said in 1984 that word processing was “erasing” literature (Kirschenbaum 2016, 43; clearly fears about slopification predate generative AI). Many writers concealed their word processor use to avoid the suspicion that the machine was doing their work for them. Generative AI seems to have become the scapegoat for our more general anxieties about our information-saturated, hyper-digital technological world, a world we feel, as a whole, we have little control over. I think many latch onto generative AI abolitionism because it is still a new enough technology to feel genuinely optional, allowing us a sense of control and virtue without facing the hard truth that in fact our iPhones probably contain as much or more injustice in them as ChatGPT—when I reflect on my own initial hostility to it, I have to admit this was a not insignificant aspect of it.
My suggestion, then, more broadly, is to avoid universal calls for generative AI rejection. Again, I want to be clear that the pledge is making no such call. As I said, while I’m not signing the pledge for the reasons delineated at the beginning of the essay, I truly support the principle of domain-by-domain and use-by-use evaluations. I would, for instance, like to see a complete ban on the use of generative AI by students as they are engaged in learning. Research skills and knowledge bases and academic intuition—the very capacities necessary to check generative AI research for hallucinations and therefore use it meaningfully to advance scholarship—these are short-circuited by use of generative AI to do research and draft papers for novices, a place where the generation of scholarship is not the point. The point is learning the forms and content. I draw from my experience in the world of the humanities, but I imagine that there are similar concerns in, say, physics, chemistry, and mathematics, where working through difficult problem sets serves the purpose of internalizing the deep logic of the field. Beyond this, I would like to extend the prohibition I would impose on AI-generated final sermon drafts to basically all load-bearing sacramental and incarnational aspects of ministry. Generative AI should never draft liturgies or prayers or guide pastoral care conversations, even if it assists the human in behind-the-scenes research. These are areas where we have staked the claim that the human element—of the being that can actually connect on a deep imago Dei level with another human—is more important than optimizing the abstracted quality of the content. At the same time, I hold that there are other areas of ministry where there is quite legitimate room for debate about whether generative AI can be employed more extensively. Can memos, emails, policies, publicity, etc., be handed over, with full human oversight, to AI? While I acknowledge that no aspect of ministry is “nontheological” and therefore inconsequential, I also think to allow every aspect of ministry to carry the same load-bearing gravitas is a recipe for burnout and one that almost no clergy do in practice. Allowing generative AI to lift some of this logistical and administrative load for already overworked clergy can actually enhance their capacity to engage with the weightier areas that demand full human attention—but I also think there is room for legitimate debate here. What I do not think there is room for debate on, though, is that congregations and middle judicatories and denominations should be doing precisely what both the pledge and Pope Leo have done—wrestled with when and how generative AI can be integrated into spiritual life and work. I would say they should go further and be providing active guidance and policy on how and when generative AI can be used by Christians.
Ultimately, generative AI is not a value-neutral technology (because none are), and its risks and harms are not trivial. By and large, I think the people at the head of these companies span from criminally negligent to actively participating in society-, democracy-, and human-dignity-undermining evil. But I do not think this is unique to generative AI—it is the unfortunate reality of the privations of a fallen creation. That those in the highest positions of corporate power seek to harness these tools to undermine democracy, exploit workers, and enrich themselves at the expense of society does not mean that the technology itself is irredeemably tainted. Acknowledgement of Original Sin and Original Corruption should not tempt us to Manichaeism. This is not, though, a kind of “no ethical consumption under capitalism, therefore do whatever you want” observation. What it does mean is that churches and individual Christians have to engage in the hard work, as with any technology, of trying to see where the tools can be salvaged and serve the proclamation of the Kingdom of God, and where they are too corrupt to adopt. And we have to have the grace to recognize that, especially in a technology’s infancy, people of good will are going to land in different, sometimes very different, places with different and defensible tolerances for the risk of being overly optimistic. Moreover, as we recognize the systemic injustice that scaffolds these tools, we must work to advocate for better environmental regulation, greater corporate accountability, and more humane and dignified working conditions for those employed in the industry. But again, this is not unique to generative AI. We should be doing this for everything we touch. The end goal is to avoid scapegoating, to engage both critically and openly with the possibilities of new technology, and to continue to work and pray for a more just and harmonious world.
Works Cited
Aczel, Balazs, Barnabas Szaszi, and Alex O. Holcombe. 2021. “A Billion-Dollar Donation: Estimating the Cost of Researchers’ Time Spent on Peer Review.” Research Integrity and Peer Review 6 (14). https://link.springer.com/article/10.1186/s41073-021-00118-2.
Baethge, Christopher, and Hannah Jergas. 2025. “Systematic Review and Meta-Analysis of Quotation Inaccuracy in Medicine.” Research Integrity and Peer Review 10 (1): 13. https://doi.org/10.1186/s41073-025-00173-z.
Covington, Paul, Jay Adams, and Emre Sargin. 2016. “Deep Neural Networks for YouTube Recommendations.” In Proceedings of the 10th ACM Conference on Recommender Systems, 191–198. New York: ACM. https://dl.acm.org/doi/10.1145/2959100.2959190.
de Vries-Gao, Alex. 2025. “Artificial Intelligence: Supply Chain Constraints and Energy Implications.” Joule 9 (6): 101961. https://doi.org/10.1016/j.joule.2025.101961.
Grim, Patrick. 2026. Understanding Artificial Intelligence: Of Minds and Machines. Course Guidebook. Chantilly, VA: The Teaching Company.
Kara, Siddharth. 2023. Cobalt Red: How the Blood of the Congo Powers Our Lives. New York: St. Martin’s Press.
Kirschenbaum, Matthew G. 2016. Track Changes: A Literary History of Word Processing. Cambridge, MA: Harvard University Press.
Li, Pengfei, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren. 2023. “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models.” arXiv preprint arXiv:2304.03271. https://arxiv.org/abs/2304.03271.
Loewen, James W. 2005. Sundown Towns: A Hidden Dimension of American Racism. New York: The New Press.
Magesh, Varun, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning, and Daniel E. Ho. 2025. “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools.” Journal of Empirical Legal Studies. https://reglab.stanford.edu/publications/hallucination-free-assessing-the-reliability-of-leading-ai-legal-research-tools/.
Miller, D. Lee, and Ryke Longest. 2020. “Reconciling Environmental Justice with Climate Change Mitigation: A Case Study of NC Swine CAFOs.” Vermont Journal of Environmental Law 21: 523–543. https://scholarship.law.duke.edu/faculty_scholarship/4034/.
Mytton, David, Dag Lundén, and Jens Malmodin. 2024. “Network Energy Use Not Directly Proportional to Data Volume: The Power Model Approach for More Reliable Network Energy Consumption Calculations.” Journal of Industrial Ecology 28: 966–980. https://onlinelibrary.wiley.com/doi/full/10.1111/jiec.13512.
Obringer, Renee, Benjamin Rachunok, Debora Maia-Silva, Maryam Arbabzadeh, Roshanak Nateghi, and Kaveh Madani. 2021. “The Overlooked Environmental Footprint of Increasing Internet Use.” Resources, Conservation and Recycling 167: 105389. https://doi.org/10.1016/j.resconrec.2020.105389.
Richter, Brian D., Gambhir Lamsal, Landon Marston, Sameer Dhakal, Laljeet Singh Sangha, Richard R. Rushforth, Dongyang Wei, et al. 2024. “New Water Accounting Reveals Why the Colorado River No Longer Reaches the Sea.” Communications Earth & Environment 5 (134). https://www.nature.com/articles/s43247-024-01291-0.
Roberts, Sarah T. 2019. Behind the Screen: Content Moderation in the Shadows of Social Media. New Haven: Yale University Press.
