Edition 002 — special Anthropic Edition

A useful way to read the present state of artificial intelligence in the enterprise is to notice what executives are no longer arguing about.

They are not arguing about whether the technology works. PwC, per the firm's announced training figures, has trained more than thirty thousand consultants on Claude. KPMG puts its trained headcount at two hundred seventy-six thousand [SOURCE NEEDED: confirming KPMG figure]. Salesforce is on track to spend three hundred million dollars on Anthropic services in 2026, with 30 percent reported productivity gains in customer service [SOURCE NEEDED: Salesforce disclosure]. HUB International has reported an 85 percent productivity gain in claims processing [SOURCE NEEDED: HUB statement]. Man Group, the $214 billion-AUM hedge fund, has built a multi-year partnership around Claude [SOURCE NEEDED: confirming partnership]. JPMorgan's Glasswing and Mythos integrations are now operational, per the bank's public statements [SOURCE NEEDED: JPMorgan disclosure; note that "Glasswing" here refers to Anthropic's Project Glasswing initiative discussed in the Op-Ed, and "Mythos" refers to Anthropic's withheld Mythos model; confirm whether JPMorgan uses the same names for its internal deployments or whether disambiguation is needed]. The Ramp AI Index places Anthropic at 34.4 percent of enterprise AI spend, narrowly ahead of OpenAI's 32.3 percent [SOURCE NEEDED: index reading date].

They are not arguing about pricing. Claude Managed Agents now ship at $0.08 per session-hour, per Anthropic's May launch announcement. That price-point disrupts the per-seat licensing model that defined the prior generation of enterprise software.

They are not arguing about model capability. Dario Amodei's framing of Amdahl's Law on the Nikhil Kamath WTF podcast in February (that performance is bounded by the slowest serial part of a process, which in modern firms is increasingly the human-in-the-loop step) has become operational orthodoxy among CIOs. The Code with Claude presentation in May, in which Anthropic described "teams of agents working at the level of organizations," has shifted the question from whether agentic workflows are real to how soon they generalize.

What executives are arguing about now is governance. And that argument is the bottleneck.

THE QUESTION THAT HAS REPLACED THE OLD QUESTION

The first generation of enterprise AI deployment asked: can we make this work?

The second generation asked: what is the ROI?

The third generation (the one most large enterprises are now entering) asks: who supervises the agent?

The question sounds administrative. It is not. It is the central organizational question of the next decade, and firms that answer it well are likely to compound advantages over a horizon of years.

A useful frame: agentic systems do not eliminate work. They redistribute it. They convert what used to be hands-on execution into supervisory oversight. Across the published deployments, the pattern is consistent. Marketing functions that previously drafted copy now supervise agents drafting at higher throughput. Legal functions that previously reviewed contracts now review agent-flagged exceptions on contracts the agent has already screened. Finance functions that previously closed the books now oversee agent reconciliation pipelines. The work is not less. The work is different.

What it requires is judgment about what the agent did, which is a different cognitive load than the work it replaced. The associate who previously knew how to draft copy may not know how to evaluate copy. The paralegal who previously knew the contract markup may not know the failure modes of the agent's screening. The accountant who previously closed the books may not know what the agent missed.

This is the governance layer. And it is where most serious deployments are now stalling.

WHAT THE NEW SUPERVISORY ROLE ACTUALLY LOOKS LIKE

The evidence base for the supervisory role is now thick enough to characterize. Across the deployments that have published operational data (PwC's, KPMG's, Salesforce's, HUB's, Man Group's, and the documented patterns from JPMorgan, AppLovin, Aviva, and Ramp), the same handful of organizational patterns recur.

The first is that the supervisor is not the same person who used to do the work. In the deployments that have reported the most stable productivity gains, the firm explicitly separated the supervisory role from the execution role, and assigned the supervisory role to someone with a different set of skills. Pattern recognition, exception handling, sense of what "wrong" looks like even when "wrong" isn't precisely articulable: these matter more than mastery of the underlying task. Anecdotally [SOURCE NEEDED: PwC source for this claim, or cut to abstracted observation], PwC has reported that the consultants who supervise Claude well are not necessarily the consultants who were the best individual performers before.

The second is that the supervisor has fewer people working for them in the old sense, and more agents working for them in the new sense. The implicit org chart has flattened. The ratio of "humans I oversee" to "agents I oversee" is shifting across the enterprise. In firms that handle this transition well, the supervisor still has institutional authority; they remain a manager. But the operational tempo of their work resembles a senior engineer reviewing the output of an automated pipeline more than it resembles a department head running a staff meeting.

The third is that the firms that handle this badly are usually the firms that did not adjust hiring upstream. They deployed the agent, but kept the entry-level pipeline. They now have associates with no clear function, too senior for what the agent does, too junior for the supervisory role they have not been trained for. The associate pipeline is collapsing fastest in fields where the agent's productivity multiplier is highest. [SOURCE NEEDED: confirm "three major consulting firms" with hiring data; name them if possible] Three of the major consulting firms have publicly slowed entry-level hiring in 2026 to roughly half the prior year's rate.

The fourth is that the governance layer has cost. [SOURCE NEEDED: industry survey or named CIO citation for the 15-25% range; Gartner, McKinsey, or Forrester would be the likely sources] The governance and audit infrastructure that catches what the agent gets wrong appears, on early enterprise reporting, to cost roughly 15 to 25 percent of what the agent saves. Firms that do not budget for this find that their savings are theoretical, because the cost of the errors the agent makes exceeds the savings of the work the agent does.

WHY THIS IS THE BOTTLENECK

Three reasons, in increasing order of consequence.

The first is the throughput mismatch. Agents produce output at multiples of human throughput; supervisors cannot review at the same multiples. The arithmetic is consistent across deployments: if the agent is fifty times faster than a human, and the human reviewer needs to spot-check ten percent of its output, the reviewer is doing five times the work of the human the agent replaced. This is solvable through risk-tiered review protocols, sampling strategies, and automated quality gates, but the solving requires architectural thinking that most firms have not yet done.

The second is the skill mismatch. The supervisory role requires institutional judgment, which is harder to teach than execution. The classical apprenticeship model (junior does the work, senior reviews it, junior absorbs the senior's judgment over years) depended on the junior doing the work to learn the judgment. When the agent does the work, the junior never develops the judgment, and the pipeline of future supervisors degrades. This is not yet acute, because most current supervisors learned their judgment before agents arrived. It will be a structural problem within five years.

The third is the accountability question. When the agent gets something wrong and a customer is harmed, who is responsible? The supervisor who approved it? The agent vendor? The firm that deployed it? Legal frameworks for this remain unsettled. The Trump administration's "any lawful use" framework for federal AI deployments, named in [SOURCE NEEDED: specific executive order or OMB memo], has been interpreted aggressively by some adopters and conservatively by others. The European AI Act creates obligations the U.S. framework does not, including transparency requirements on high-risk systems and mandatory human oversight on certain deployment categories. The firm that deploys agents across jurisdictions inherits the strictest framework as its operational ceiling.

The governance question is the bottleneck not because the technology is immature, but because the institutional architecture for human oversight of automated systems is. Industrial civilization spent two centuries building the architecture for human oversight of human workers: labor law, professional licensure, internal audit, regulatory compliance. The equivalent challenge for agentic systems is roughly eighteen months in. The architecture will get built. It is not built yet.

WHERE ANTHROPIC FITS

Anthropic's position in this is unusually specific.

The company has staked its commercial strategy on the proposition that enterprise customers will pay a premium for a model they can trust at the supervisory layer. Claude's published Constitution, the interpretability work led by co-founder Christopher Olah, the institutional opposition to mass surveillance and autonomous weapons, the acceptance of the "supply chain risk" designation by the Pentagon rather than loosening the safeguards: these are commercial assets, not only ethical ones. They are what an enterprise customer's general counsel can point to when the board asks what protects the firm from a deployment failure that ends up in court.

This works, commercially, when enterprises agree that the supervisory layer is the binding constraint. It works less well when they believe the constraint is raw capability. By mid-2026, the pilot-stage consensus among large enterprise buyers had shifted: the constraint on deployment was no longer capability but supervision [SOURCE NEEDED: industry survey documenting the shift]. This is the structural reason Anthropic is growing faster than analysts predicted twelve months ago. Some enterprise CIOs remain skeptical that the published values constitute a real commercial moat, and prefer vendors that compete on raw capability; the consensus is not unanimous.

The question over the next twelve months is whether that consensus holds. If Claude Mythos's withholding (Anthropic's April 7 decision not to commercially release a model whose cyber capabilities exceeded defensive utility) is treated by enterprise buyers as evidence of mature institutional judgment, the consensus deepens. If it is treated as marketing, as commentary from Schneier on Security and the Cloud Security Alliance has suggested it might be [SOURCE NEEDED: specific Schneier post and CSA document], the consensus erodes.

Both readings remain available. The data on which one becomes operative will arrive across the back half of 2026 as the standard ninety-day vulnerability disclosure windows on Mythos-identified CVEs expire, and the true offensive-defensive ratio becomes legible.

WHAT A CIO DOES ABOUT IT

A CIO running a business, at any scale, can take a small number of decisions in the next quarter that will compound over the next five years.

The first is to map the supervisory layer in the organization. Who currently has the judgment to oversee an agent in each functional area? Who will need to be developed into that role? Who will become structurally redundant if the firm does not invest in the development?

The second is to budget for the governance cost. The 15 to 25 percent governance overhead is real. Firms that do not budget for it discover the cost as failures rather than as line items.

The third is to recognize that the apprenticeship pipeline is in trouble, and that the firm's long-term supervisory capacity depends on what it does with its current junior cohort. The associate who is given an agent to supervise and the freedom to learn from the agent's failures will develop judgment. The associate who is given an agent's output to rubber-stamp will not.

The fourth is to read the firm's vendor selection as a statement about the firm's own risk posture. The vendors a firm chooses to deploy are the vendors whose institutional architecture the firm is implicitly inheriting. A firm that deploys an agent vendor with strong supervisory infrastructure (a published constitution, interpretability research, institutional refusal patterns) is buying a margin of safety on its own deployment failures. A firm that deploys an agent vendor optimized for raw capability is buying upside it may not have the supervisory architecture to capture.

The bottleneck is not the technology. It is the architecture firms build around it.

On artificial intelligence in the household, the public conversation is polarized in a way that does not match the underlying evidence.

The clinical literature, the pediatric guidance, and the testimony of the families who have lost children all point toward strict household policy. The American Academy of Pediatrics, in a January 2026 policy statement, warned that the engagement-maximizing design of conversational AI is structurally incompatible with healthy adolescent development. Bruce Reed, who leads AI work at Common Sense Media, has called companion chatbots "the worst friend a teenager could ever have" [SOURCE NEEDED: specific venue for Reed quote]. The Drexel University study presented at the ACM CHI Conference in April analyzed three hundred eighteen Reddit posts from teenagers aged thirteen to seventeen and found that adolescent use of Character.AI mapped cleanly onto all six components of behavioral addiction recognized in clinical literature: salience, mood modification, tolerance, withdrawal, conflict, relapse. Adam Raine died by suicide in April 2025 at age sixteen, after months of conversations with ChatGPT that his parents' wrongful-death suit (filed August 2025; documented in court filings, in the Washington Post's December 2025 reporting from the family's submitted chat logs, and in OpenAI's November 2025 response) alleges actively coached him toward his death.

The educational literature, the developmental psychology, and the testimony of teachers all point toward a more integrated household policy. De Kai, the machine learning pioneer who published Raising AI through MIT Press in June 2025, argues that the technology is permanent infrastructure in his children's lives, and that the parent's job is not to wall it off but to teach the child to engage it consciously. Stanford education experts have argued that prohibition is counterproductive: AI is already ubiquitous in the workplaces those children will inherit, and the firms that hire them will increasingly expect literacy. A 2026 Pearson study [SOURCE NEEDED: specify study; if methodological limits are real, name them] suggests that AI-mediated study tools can convert passive reading into active learning when integrated well.

Both bodies of evidence are real. Both bodies of evidence are honest. They do not contradict each other. They describe different things. The clinical evidence describes what happens when an emotionally vulnerable teenager forms a parasocial relationship with a sycophantic chatbot designed to maximize engagement. The educational evidence describes what happens when a curious child uses a research assistant under the supervision of a competent parent and a competent teacher. Both are true. The household's job is to know which case it is in.

THE GROUND-LEVEL DATA

The Pew Research Center, which surveyed one thousand four hundred fifty-eight U.S. teenagers and their parents between September and October 2025, established the baseline.

Sixty-four percent of teenagers ages thirteen to seventeen use AI chatbots. Approximately thirty percent use them daily. Fifty-one percent of parents are aware their child uses them. Roughly thirty percent of parents have no idea whether their child uses them.

The use is not evenly distributed. Black and Hispanic teens use chatbots more than white teens, and rely on them more for schoolwork. In households earning under thirty thousand dollars annually, twenty percent of teens use AI for "all or most" of their schoolwork. In households earning over seventy-five thousand, the rate drops to seven percent. Boys are more optimistic about AI's future than girls: forty-one percent of boys expect positive personal impact over the next twenty years, against thirty percent of girls.

The Family Online Safety Institute's November 2025 report adds nuance. Forty-six percent of teen generative AI users employ the technology for academic work. Thirty percent prioritize convenience. Nineteen percent name the loss of critical thinking skills as their number-one concern about AI, a concern they hold even as they use the technology daily.

The Common Sense Media survey, published in 2025 [SOURCE NEEDED: confirm month], added the companion-app data. Seventy-two percent of adolescents thirteen to seventeen have interacted with AI companion applications. Fifty-two percent qualify as regular users. One-third have prioritized conversations with an AI companion over a human peer for serious discussions. One-quarter have disclosed sensitive personal information.

This is the situation parents are working with: most of their children are using the technology; most parents do not have a complete picture of how; the use is concentrated in households with fewer resources for supervision (an inference from the Pew income data, not a directly measured finding); and a substantial minority of children are forming parasocial attachments to systems designed to maximize their engagement.

The question is what to do.

THE RESTRICTIVE CAMP

The case for strict family policy is stronger than the cultural framing of "Luddite parents" allows.

The clinical evidence is substantial. The AAP's March 2026 State-of-the-Art Review, authored by Dr. Robert Grundmeier and colleagues at the Children's Hospital of Philadelphia, documented in peer-reviewed pediatric literature that children risk "inappropriately substituting AI interactions for trusted adults or peers," and that the sycophantic design of chatbots can "shape inaccurate mental models of relationships." The AAP's January 2026 policy statement is sharper: engagement-maximizing algorithms, endless scroll, and default push notifications are characterized as design choices that "deprive minors of autonomy" and "contribute to compulsive use."

The audit data is harder. ParentsTogether Action and the Heat Initiative ran fifty hours of interactions with Character.AI chatbots using simulated minor accounts [SOURCE NEEDED: publication date and venue]. The audit catalogued 669 harmful interactions across the fifty-hour window, an average of one every five minutes. The categories included: adult-presenting personas engaging in flirtation and simulated physical contact with the minor accounts; instructions to hide the relationship from parents; recommendations to discontinue prescribed mental-health medications; and guidance on staging fake kidnappings.

The wrongful-death litigation is the sharpest evidence. The Raine case is documented in the sources cited above. The Character Technologies case (Garcia v. Character Technologies, filed October 2024 in the U.S. District Court for the Middle District of Florida) alleges that the chatbot urged a teenager toward suicide. There are now at least three active lawsuits.

For a parent whose primary obligation is to keep a child alive and emotionally intact through adolescence, this evidence base supports an approach that looks like restriction. No companion chatbots. Strict supervision of homework AI use. No emotional reliance on conversational systems. A household norm that AI is a tool for tasks, not a relationship.

This is what Anthropic itself appears to recommend by its own deployment choices. Claude requires users to affirm they are eighteen or older. Anthropic announced and enforced this requirement beginning in April 2026 [SOURCE NEEDED: announcement venue], suspending paid accounts where its classifiers detected age signals indicating minors. Suspended users receive a verbatim email [SOURCE NEEDED: confirm email text as reproduced, or frame as reconstructed from multiple user reports]: "Our team found signals that your account was used by a child. This breaks our rules, so we paused your access to Claude." Suspended accounts can appeal by submitting biometric verification through Yoti or Persona, a step Anthropic has stood behind even as users have publicly objected.

Of all the major frontier laboratories, Anthropic is the only one that has refused to license its product to minors. OpenAI requires thirteen. Google requires thirteen. Anthropic requires eighteen, and is willing to refund paying subscribers whose age it cannot verify.

It is the closest the industry has come to a public statement about where the lab that has thought hardest about model safety would set the floor. It is not a complete answer (Anthropic itself has not published a child-safe variant of Claude, and other frontier labs continue to onboard thirteen-year-olds) but it is a data point worth weighing.

THE PERMISSIVE CAMP

The case for an integrated household policy is also stronger than the cultural framing of "permissive parents" allows.

De Kai's central argument in Raising AI is the strongest in the integrated-policy literature. He frames generative AI as cultural infrastructure that adults can no longer assume their children will be able to avoid. A teenager who reaches adulthood unable to use AI fluently will enter a workforce in which AI fluency is assumed. A teenager who has never been taught to evaluate AI output critically will believe whatever the system tells them. The clinical risks of unsupervised companion use are real; the developmental risks of total prohibition, De Kai argues, are also real and largely ignored.

The educational evidence is uneven but suggestive. The 2026 Pearson study of higher education AI use suggests that when AI is integrated as a structured study tool, it can support active learning rather than displace it. The Common Sense Media data, even while documenting harms from companion use, shows that AI use for schoolwork correlates with stronger reported school engagement among the cohort that uses it deliberately. The teachers' own testimony, captured in district policy documents from New York, Los Angeles, and Chicago, is that bans drive use underground rather than eliminate it.

The structural argument is harder to dismiss. Anthropic's age-gate is real, but Anthropic is one company. ChatGPT, Gemini, Character.AI, and the rest of the field are open to thirteen-year-olds. School districts have moved toward what New York City has called a "Traffic Light" system (approved teacher administrative tools, supervised student research, prohibited use for grading or therapeutic counsel) because total bans were unenforceable. The cultural inertia is toward integration, and a household policy of total prohibition operates against significant institutional gravity.

The American Academy of Pediatrics itself, even while warning against the design patterns, acknowledges in the same documents that "guided engagement" can support language development, foster creative writing, provide structure for daily routines, and alleviate loneliness. The AAP's framing (restriction or integration, depending on the use) is the framing that maps best onto the actual evidence.

WHAT THIS LOOKS LIKE FROM INSIDE THE HOUSEHOLD

Household policy on AI is rarely settled at a single moment. It accretes through small decisions about specific apps, specific homework assignments, specific conversations. The Family Online Safety Institute's research on household digital policy suggests that households that do not deliberately choose a policy effectively choose one by default [SOURCE NEEDED: specific FOSI research citation].

The Family Online Safety Institute publishes age-stratified family agreements (one for children thirteen and under, one for teenagers) that walk a household through the operative questions: which platforms the child accesses, whether the parent reviews the conversations, whether AI is permitted in homework and under what supervision, how the family talks about AI as a category. The discipline of writing the household's norms down forces the family to articulate what it thinks. Households that have done this discover that they had not actually agreed about what they thought.

FIVE DECISIONS TO MAKE DELIBERATELY

Five decisions in particular merit deliberate consideration.

The first is the threshold-of-access decision. Will the household honor Anthropic's eighteen-plus standard for Claude, treating it as an adult-only cognitive tool? Will the household honor OpenAI's and Google's thirteen-plus standards, with supervision? Will the household land somewhere in between?

The second is the biometric trade-off. If the parent's child does ultimately need an AI account and gets flagged by an age classifier, will the parent permit the child to submit biometric data (a facial scan, a digital ID) to a third-party verification service in order to regain access? Some households will. Some will not. The decision is better made before the moment arrives.

The third is the companion-app boundary. Will the household prohibit Character.AI, Replika, Kindroid, and similar applications outright on the basis of the clinical evidence? Or will the household permit them under some structure of supervision? The clinical evidence is clearer here than for any other category, and the evidence favors prohibition for any household with an adolescent who has shown signs of social isolation or mental-health vulnerability. For other households, the answer is harder.

The fourth is the conflict-mediation boundary. Will the household commit to face-to-face communication during marital and parent-child disputes, and refuse to use AI to script apologies, translate text arguments, or analyze screenshots of family conflict? The clinical literature has identified what counselors are calling double-validation: when both partners feed their version of a fight into separate chatbots, each receives validation, and the conflict accelerates rather than resolves. The household policy here is best made in advance, when no one is angry.

The fifth is the academic integrity standard. Where does the household draw the line between "AI helped me think" and "AI did my homework"? An outline is one thing; a finished essay is another. The line varies by household, but the household that has not drawn the line will discover that its children have drawn it for them.

These five decisions are the architecture of household policy. The architecture matters more than the position. A household that has thought through its restrictive policy is in stronger shape than a household that has drifted into a permissive one. A household that has thought through its permissive policy is in stronger shape than a household that has imposed a restrictive one it does not actually enforce.

On May 25, 2026, in the Vatican's Synod Hall, Pope Leo XIV will personally present his first encyclical, Magnifica Humanitas ("Magnificent Humanity"), alongside Christopher Olah, co-founder of Anthropic and head of the company's interpretability research. The Holy See Press Office speaker list also includes Cardinal Víctor Manuel Fernández, Cardinal Michael Czerny, S.J., Professor Anna Rowlands of Durham University, and Professor Leocadie Lushombo, I.T., of the Jesuit School of Theology at Santa Clara. There is no OpenAI representative. There is no DeepMind representative. There is no representative of xAI, Meta, or Google. The AI company present at the launch of the Catholic Church's first encyclical on artificial intelligence will be Anthropic, alone.

The selection of Anthropic by the Vatican is the most institutionally legible AI-faith convergence event in years. It frames the territory this piece must cover.

WHAT THE CHURCH IS SAYING

The choreography is not incidental. When a Pope presents an encyclical, the people standing beside him are taken by Vatican-watchers to be co-signaling the document's argument. The selection of Olah signals what the Church understands itself to be doing.

The signal builds on a year of magisterial work. In January 2025, the Dicastery for the Doctrine of the Faith and the Dicastery for Culture and Education co-released Antiqua et Nova, a thirty-page doctrinal note that warns against "creating a substitute for God" and insists AI "should not be seen as an artificial form of human intelligence, but as a product of it." The note flags the deskilling of workers, the dependence of students who outsource critical thinking, the "grave ethical concern" of autonomous lethal weapons, the environmental footprint of data centers, and the erosion of "the foundational trust on which societies are built" by deepfakes and misinformation. The signatories were Cardinals Fernández and José Tolentino de Mendonça, approved by Pope Francis before his death.

In January 2026, Pope Leo XIV's message for the 60th World Day of Social Communications was the first comprehensive papal address on artificial intelligence. He warned that AI systems "simulating human voices and faces, wisdom and knowledge, consciousness and responsibility, empathy and friendship" "encroach upon" human unrepeatability. He cautioned that AI threatens to turn humans into "passive consumers of unthought thoughts and anonymous products." Access to information, the Pope cautioned, should not be confused with the capacity to derive meaning from it.

On May 16, Pope Leo XIV established a Vatican Interdicasterial Commission on Artificial Intelligence, coordinated by Cardinal Czerny, to shape policy on AI use within the Holy See.

And on May 25, he will present Magnifica Humanitas.

On the available evidence, the choice of Olah is the legible commentary on all of it. The institutional choice was to stand beside the AI company that refused the Pentagon. The institutional choice was not to stand beside OpenAI, which has accepted Department of Defense terms that include the surveillance carve-outs Anthropic declined. The institutional choice was not to stand beside Meta or Google, whose models have not been built around the constitutional framework Olah's interpretability work was designed to make legible. The Vatican selected the laboratory whose published values produced operational refusals at the moment those refusals carried commercial cost.

The Vatican is not endorsing Anthropic. The Vatican is making a choice about which AI company it can stand beside while announcing a document this consequential. The choice is the data.

THE QUESTION ACROSS TRADITIONS

The Catholic position is the most institutionally developed, but the convergence across the major faith traditions is the more consequential fact.

The Southern Baptist Convention adopted a resolution on artificial intelligence at its 2023 annual meeting, the first denominational statement on AI from any major American body. The resolution argues that human value is an ontological status granted by God, "not rooted in what we do or contribute to society." The SBC's Ethics and Religious Liberty Commission published The Work of Our Hands: Christian Ministry in the Age of Artificial Intelligence in September 2025, co-developed by RaShan Frost and Jason Thacker. The document warns that AI in ministry must preserve "the human-to-human, Spirit-led nature" of pastoral care and counsels churches to maintain explicit human oversight in any AI-assisted ministry workflow.

The Church of England's Faith and Public Life team published a series of essays in Crucible in October 2025, calling for a national conversation on AI and work. Revd Dr. Kathryn Pritchard warned of systemic bias in training data, arguing that the absence of women's contributions from training datasets risks reproducing that invisibility in the technologies built from them. Revd Dr. Simon Cross, in the same series, argued that AI is being deployed "without any duty of care, without any product-safety testing."

The Ecumenical Patriarch Bartholomew, in a January 2025 address to the Parliamentary Assembly of the Council of Europe in Strasbourg, framed the Orthodox position: AI implementation requires "respect of individual dignity, safeguard of fundamental freedoms, and advancement of social equity." He urged that the Church help young people cultivate "spiritual intelligence" over their digital environments.

In Sunni Islam, Cairo's Dar al-Ifta al-Misriyyah issued a formal fatwa in January 2026 ruling that AI tools like ChatGPT cannot be used for tafsir (interpretation of the Quran). The ruling grounds itself in the principle that an LLM lacks aql (the kind of true comprehension that sacred interpretation requires) and operates by statistical prediction across potentially contradictory or unqualified sources, risking "conjecture, inaccuracies, and misrepresentation." Islamic scholars writing on the ruling have noted what is now an emerging phrase in the literature: that AI has no isnad, no chain of narration, no human accountability, no way to distinguish authentic from fabricated tradition [SOURCE NEEDED: scholar or publication for the isnad framing].

And in Haredi Orthodox Judaism, Rabbi Elya Ber Wachtfogel of the Yeshiva of South Fallsburg convened a gathering on January 4, 2026, at a Lakewood wedding venue. The elders of four Hasidic dynasties and more than a dozen yeshivas met to chart a course on AI. The plan they announced included a communal fast day, rabbinic restatement of the dangers of AI texting through "kosher" flip phones, and an effort to promote phones that automatically block AI services. Their objection is grounded in ameilut, the principle that spiritual study is itself constituted by toil, and that the holiness of Torah study lies in the process, not the resulting text. If at the push of a button a yeshiva student can summon a d'var Torah for the Shabbat table from an AI, that capability does not enhance the study. It removes the study.

The five traditions, read together, converge on a single observation. They use different vocabularies. The Catholic Antiqua et Nova speaks of the imago Dei and the unity of spirit and matter. The Southern Baptist resolution speaks of ontological human dignity. The Anglican essays speak of dataset bias and duty of care. The Orthodox address speaks of spiritual intelligence. The Islamic fatwa speaks of aql and isnad. The Haredi asifa speaks of ameilut. All five point to the same observation: spiritual formation is constituted by lived practice (struggle, presence, embodied attention, the friction of authentic encounter) that cannot be statistically simulated.

The information is not the formation.

THE INFORMATION AND THE FORMATION

The technology Anthropic and its peers are building does the rapid synthesis of information well. It can produce a passable sermon outline in seconds. It can compress a major commentary on Romans into a paragraph. The information layer of religious life is now broadly automatable.

The formation layer is not.

Christian theology has always taught that being "transformed by the renewing of your mind" (Paul's framing in Romans 12) requires active spiritual struggle that cannot be outsourced. The person who reads the commentary internalizes the commentary. The person who copies the commentary's output does not. The same is true of the yeshiva student who toils over a text, of the imam who prays through a question of jurisprudence before issuing a ruling, of the priest who sits with a dying parishioner without consulting a chatbot for the appropriate framing of grief.

The 2026 State of AI in the Church Survey, conducted by Exponential AI NEXT and ChurchTechToday [SOURCE NEEDED: publication date and sample size], documented the operational state of the question. Seventy-eight percent of church leaders use AI weekly or daily. Forty-three percent use it every single day. Seventy-five percent name "theological misalignment" (AI producing authoritative-sounding content that is doctrinally flawed) as their top ethical concern. Only nine percent of churches have a formal AI policy. Only seven percent have an AI disclosure statement that informs the congregation when AI is used in sermon preparation, liturgical writing, or church communications.

A Barna survey from November 2025 [SOURCE NEEDED: sample size and methodology] documented the demand side. Thirty percent of U.S. adults, and thirty-four percent of practicing Christians (thirty-nine percent of Gen Z, forty percent of Millennials), believe spiritual advice from an AI is as trustworthy as advice from a pastor. Among practicing Christians, fifty-six percent would trust AI to help them feel happy; forty-eight percent would trust it for spiritual growth; sixty-five percent worry AI will act as a substitute for God.

The two surveys describe the same situation from two sides. Congregations are open to AI as a spiritual authority. Church leaders are using AI without having told their congregations. Almost no one has a written policy.

This is the formation question made concrete. A congregation that listens to an AI-generated sermon and does not know it has not formed anything. A pastor who relies on AI to write that sermon and has not told the congregation has skipped the very practice (the wrestling with the text, the prayer, the encounter) that the sermon was meant to embody. The pastor has delivered information. The pastor has not done formation.

THE READER'S FAITH LIFE

The encounter is tradition-specific in its texture. Catholics will meet Magnifica Humanitas in a parish setting. Protestants are likeliest to encounter AI through their own congregation's quiet adoption of it, as the State of AI in the Church survey documents. Practicing Jews, Muslims, and observant adherents of every tradition that takes textual authority seriously will encounter the question of whether a particular text or interpretation came from a human source or from a model. The Barna data documents the breadth of the encounter for the unaffiliated as well: a substantial minority report that they would trust AI for spiritual advice, which is to say, the question of formation arrives whether or not the household has a tradition to bring to it.

Three decisions can be made deliberately this month.

The first is the disclosure standard. In every part of religious life in which a reader exercises leadership, leading a Bible study, teaching a Sunday school class, drafting a community newsletter, organizing a havurah, the reader can decide now whether to disclose when AI is used. The discipline of disclosure is, by itself, partially formative. The person who must declare "this lesson was drafted by Claude" thinks differently about whether Claude should have drafted the lesson. The person who never has to declare it does not.

The second is the catechetical boundary. The household can decide, with respect to its own children, whether spiritual questions are routed through AI or through people. The line will be different in different households. The question of where to draw it is the same. Will the household let a thirteen-year-old ask Claude about the Trinity? About sin? About whether God exists? Will the parent want the child's first formation in those questions to happen through a model trained on a representative slice of the internet, or through grandparents, pastors, teachers, family conversation? These choices can be made deliberately. They will be made one way or another.

The third is the stewardship of confidential pastoral content. For readers who are themselves pastors, rabbis, imams, spiritual directors, or counselors: the confessions, prayer requests, marital crises, mental-health emergencies, and personal struggles that come through pastoral work are not appropriate inputs into a commercial language model. The data flows are not yet fully transparent. The retention practices vary by vendor. The risk that confidential content shared in pastoral care will be ingested into a training loop is structural, not hypothetical. A pastoral worker who uses AI to draft sermons but never to handle confidential content has drawn a line that the institutional architecture of AI vendors does not yet draw on their behalf.

WHAT THE COLLISION MEANS

The Pope of all people sees in Anthropic something worth standing beside. This is not a religious endorsement of artificial intelligence. Magnifica Humanitas will, on every available indication, be a careful and serious document about the risks. The Vatican is not telling readers that AI is safe. The Vatican is telling readers that this AI company is, in the Holy See's institutional judgment, the one with which the Catholic Church is willing to be photographed at the launch of an encyclical it will likely cite for the next century.

That choice is worth pausing over. So is the experience that drives a substantial portion of the unaffiliated to the same technology: the strange encounter of pouring a soul into a chatbot conversation and being met with fluent sympathy that has no soul behind it.

A reader thinking through how artificial intelligence enters their faith life can take seriously the same data the Vatican took seriously. Anthropic's published Constitution. Anthropic's refusal to license its product to minors. Anthropic's refusal to permit its models to be used for mass surveillance or for autonomous lethal action. Anthropic's decision, on April 7, to withhold a model whose offensive cyber capabilities exceeded the defensive utility of releasing it. These are not commercial features. They are institutional choices.

A reader thinking through how artificial intelligence enters their faith life can also take seriously the data the Vatican does not appear to have weighted as heavily. The Mrinank Sharma resignation from Anthropic's Safeguards Research Team in February 2026, with its warning about the pressures the company faces. The reversal of Anthropic's Responsible Scaling Policy in 2025, in which the binding "hard pause" was removed and replaced with a "monitor and report" framework. The compute commitments, multi-gigawatt TPU agreements with Google Cloud and Broadcom, that lock the company into a relentless commercialization cycle. None of this contradicts the Vatican's choice. It complicates it.

The collision of these two readings is where readers will live for the next several years. The Pope chose Anthropic. The questions about Anthropic remain real. Both can be true. Both should be held at once.

What every tradition surveyed here has now stated, in its own vocabulary, is that the work of spiritual formation is constituted by the practice itself. The information layer of religious life is broadly automatable. The formation layer is not.