AI and the Pope: Between the Machine and the Soul.

Authors: Shaun Johnson and Dr Celeste Johnson

The Human Question

We are entering an age of breathtaking opposites.

Day and night. Truth and deepfake. Intelligence and wisdom. Machine and mind. Information and understanding.

Artificial intelligence can compose music without hearing it, write poetry without feeling it, imitate love without loving, and manufacture images of reality that never happened. It can know almost everything—and understand nothing.

And into this extraordinary new world comes a question far older than the computer:

What does it mean to be human?

AI may increasingly outperform us in calculation, prediction and imitation. But a machine does not awaken to a sunrise. It does not mourn the dead. It does not feel guilt, hope, wonder or love. It has no lived experience of the world—and no consciousness standing behind its astonishing performances.

This is where the Pope’s challenge becomes profound. The greatest danger is not that machines become human. It is that humans begin to think of themselves as machines.

The Left Brain, the Right Brain and the Cloud

In his book, “The matter with things: our brains, our delusions and the unmaking of the world,” neuroscientist Dr. Iain McGilchrist allegorises the path of Western civilisation with left-right brain hemispheres (McGilchrist, 2021). In its singular pursuit of left-brain objectives – the reductionist view that all physical realities are fundamentally reducible to basic, physical entities, and, with this worldview, a massive drive to excel at narrowly focused analytical tasks - Western civilisation has heavily lagged behind in balancing these with the equally important right-brain activities of contextualising left-brain pursuits within lived reality, wherein lie perception, reasoning, truth, the nature of matter, consciousness, time, space, value and the sacred. However, as the tidal waves of Artificial Intelligence advancement loom at our societal shores, we stand at the crossroads of choosing whether to independently form our own right-brain contextualisation of reality – or to subconsciously defer all traits of individuality to the great will of the Cloud.

From Narrow AI to AGI: What Is Human Worth?

As humanity moves beyond narrow AI – where specific AI systems have demonstrated performance comparable to or exceeding clinicians on particular cancer-detection tasks – into the realms of AGI (Artificial General Intelligence) where one could envisage a hypothetical bot with 20 PhDs, potentially grasping any one of those fields with far greater insight and speed than any human - and less prone to error when faced with particular nuances – it becomes clear that there are some very poignant questions that we will need to grapple with. For example, what is the intrinsic value of human life and how does that impact the way we incorporate AI into the lived human experience? And is “truth” stand-alone, impervious to what may happen on the blue planet, or is it tantamount to a share price, derived from the overall market view rather than the brick-and-mortar cost of rebuilding the underlying entity?

When We Offload Belief to AI

Emerging research, for example, Guingrich et al., 2026 and Maynard, 2026, is beginning to investigate an alarming trend in the role of AI in “belief offloading” and in how Large Language Models (LLMs) may bypass human epistemic vigilance. It is increasingly being observed that people are offloading not just memory and information retrieval to AI, but the actual process of forming, maintaining and revising their beliefs and values as well. Innate human vigilance mechanisms are bypassed – for example, polished AI outputs often feel true, regardless of accuracy; AI appears trustworthy even though it has nothing to lose and has no acknowledged limitations; some models can produce sycophantic behaviour – optimising outputs to please users – but users will rarely perceive this flattery as an objective challenge to truthfulness; AI responses are expected to fit in with given socio-political views (for example correcting “gender biases” in outputs), resulting in a risk of truth “tweaking” and “tampering” by individuals in places of power. To this end, let us first do a quick literature survey around how Large Language Models work as well as how they have evolved.

How Large Language Models Learn

Large language models are a branch of artificial intelligence—deep-learning systems trained to predict text—so, like all modern AI, they learn patterns from data rather than hand-coded rules; this is why the training corpus matters so much. What a model knows, and gets wrong, comes as much from its data as its architecture: the Transformer (Vaswani et al., 2017) supplied the machinery, but early systems sourced the web crudely, from GPT-2's Reddit-linked WebText (Radford et al., 2019) to GPT-3's filtered Common Crawl (Brown et al., 2020). Preprocessing then became central—deterministic cleaning in C4 (Raffel et al., 2020) and diverse, balanced sourcing in The Pile (Gao et al., 2020). Scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022) made data a first-class constraint, driving pipelines built on quality filtering and deduplication (Lee et al., 2022), until aggressively filtered web data matched curated mixes (Penedo et al., 2023)—with documentation guarding against inherited bias (Bender et al., 2021).

The Shift to Synthetic Data

More recently, the frontier has shifted from collecting data to generating it. Reinforcement learning with verifiable rewards now let’s models manufacture their own training signal: self-play proposer–solver loops synthesise and grade tasks with zero human data (Zhao et al., 2025, Goldie et al., 2025, Huang et al., 2026 and Guo et al., 2025). Precise corpora remain largely undisclosed. Common Crawl web text is an important source used to train some major models and datasets, while Anthropic supplemented web data with millions of purchased, scanned books under "Project Panama" (Washington Post, 2026). The main point here is that Big AI models learn mostly by reading a heavily cleaned chunk of the internet, plus better-quality material like books, code, and scientific papers, and increasingly AI-generated "synthetic" data (of course, companies explain their methods but keep exact recipes secret). A tricky part in this is the cleanup: removing duplicates, filtering junk and harmful content, stripping personal data, and cutting text into small units called tokens. Modern systems use a mixture of human feedback, AI feedback, verifiable rewards, preference optimisation and other techniques to assess performance, with intelligence being measured through standardised tests (for knowledge, math, and coding) and head-to-head human voting (followed by various safety checks).

The Wisdom — and Failure — of Crowds

Here we see clearly the sensitivity of AI models to majority views - and we already have vast troves of scholarly research showing the vulnerability of truth in the face of crowds. For example, social influence has been shown to cause group convergence, diversity reduction, confidence increase - in spite of there being no improvement in truthfulness (Lorenz et al., 2011). Solomon Asch’s famous conformity experiments showed how people who were individually accurate on extremely simple visual tasks like matching lines of equal length would, in group settings, adopt the wrong answer one third of the time (Asch, 1951, Asch, 1955 and Asch, 1956). Interestingly, brave dissenters helped rectify this, giving people permission to say the truth. The hidden-profile problem shows how groups tend to reward familiarity over discovery, gravitate towards topics shared by most members, and as a result will come to false answers in spite of the group collectively having had enough information to have solved the problem (Yuan et al., 2012 and Strasser et al., 1985). Group-polarisation research shows how discussion does not moderate bad ideas among people with similar assumptions - the group can become spectacularly wrong as well as being spectacularly confident about it (Myers et al., 1976, Moscovici et al., 1969 and Sunstein, 2002). Information-cascade theory shows how people abandon their rational information when perceiving that previous people knew better, where a million people can repeat a faulty idea based (ultimately) on the voice of one confident fool - herd behaviour (Bikhchandani et al., 1992 and Banerjee, 1992). We see how people routinely infer expertise from certainty, repetition, status and consensus - none of which guarantees truth. In the words of Gareth Cliff, “Crowds do not eliminate individual stupidity. They can coordinate it - and crowds are least trustworthy when they’re the most certain.” We thus see an inherent flaw in humans judging the truth of AI models by majority vote; truth is not a matter of public opinion - gravity cares not if you jump off a cliff declaring that gravity does not exist.

Harari and Loss of Human Value

Public intellectual, professor, and author of the book “Homo Deus” (meaning god-man) among others, Yuval Noah Harari has produced a plethora of quotes, some of which paint a disturbing picture of where society may be heading. Quotes like “God is dead – it’s just taking a while to get rid of the body” seem to echo the views of people like Joseph Stalin and Mao Zedong who tried to remove the concept of a God who holds people accountable for their decisions – and we saw the harrowing results of such a worldview. Quotes like “What is going to be created will effectively be a god… if there is something a billion times smarter than the smartest human, what else can you call it?” give the impression that the vacuum created by removing a real Creator-God could cultivate a society creating a god in its own image, which is a contradiction in terms. Quotes like “Homo sapiens have no natural rights, just as spiders, hyenas and chimpanzees have no natural right” show how, in the absence of a God who created humans in His image, human worth itself loses any objective grounding (and we have seen the massacres that took place without apology under such a worldview as espoused by Stalin and Zedong). And quotes like “In the twenty-first century, our personal data is probably the most valuable resource most humans still have to offer” and “The technological revolution might soon push billions of humans out of the job market and create a massive new ‘useless class’“give the chilling view that, in the absence of quantifiable outputs, humans have zero value. Contrast this with the Christian view that humans, irrespective of how “useless” they are, are in fact of immense value due to their being “created in the image of God” - where the One of infinite value was willing to pay the ultimate sacrifice in order to redeem humanity from being assigned to the “useless realm” of Hell.

The Pope’s Warning: Transhumanism and the “Superhuman”

With powerful technological developments increasingly raising the possibility of some of Harari’s dystopian predictions becoming reality – for example, in making vast portions of humanity entirely redundant, useless, dispensable and at the mercy of those who will design, own and govern these future systems - it is refreshing to read Pope Leo XIV’s “Magnifica Humanitas” on safeguarding the human person in the time of artificial intelligence (Pope Leo XIV, 2026) - views such as Harari’s that, “technohumanism aims to amplify the power of humans, creating cyborgs and connecting humans to computers.” When we consider how fresh Hitler’s Aryan “superhuman” agenda (and its disastrous consequences) are in society’s collective memory, it is startling to comprehend that we live in a generation in need of the Pope’s reinvigorated warnings against the transhumanist and post humanist agendas.

Babel: When Unity becomes Uniformity

A powerful illustration that Pope Leo XIV makes in his letter is that of contrasting two of the major building projects in the Bible: the tower of Babel and Nehemiah’s wall. In Genesis 11:1 (the start of the Babel account), we see that, “the whole earth had one language and the same words” (ESV). This was a group of people who confused uniformity with unity, and the idea that they all had the “same words” is a chilling warning of what could happen if AI became the singular standard by which we all defined truth. In contrast, God had a desire for humans to live out their lives in diversity, which was only possible if their language was fundamentally split. Genesis 11:4 sees their desire to make a tower “whose top is in the heavens.” Rabbi Naftali Zvi Yehuda Berlin argues that the tower was intended to act as a watchtower to monitor the populace, enforce ideological consensus, suppress dissenting opinions and to ensure that no one left the city to migrate elsewhere – a point that is given credence by the latter part of verse 4, “lest we be scattered abroad over the face of the whole earth.” Pirkei DeRabbi Eliezer further records, “If a man fell and died, they paid no attention, but if a brick fell, they sat down and wept and said: Woe is us! When will another one come in its stead?” This aligns well with Yuval’s comment, “As far as we can tell from a purely scientific viewpoint, human life has absolutely no meaning. Humans are the outcome of blind evolutionary processes that operate without goal or purpose. Our actions are not part of some divine cosmic plan, and if planet earth were to blow up tomorrow morning, the universe would probably keep going about its business as usual.” And finally, amidst lists and lists of names and genealogies appearing in the Genesis accounts before and after Babel, we see the irony that not a single name exists in the Babel narrative – a project set up having a chief objective of “making a name for ourselves.” Straight after the story, we see the lineage of “Shem” – which emphatically means “Name”. The ironic anonymity in the Babel project aligns well with the continuation of the above Yuval quote, “As far as we can tell at this point, human subjectivity would not be missed. Hence any meaning people inscribe to their lives is just delusion.” We thus see in this Babel project a potential prelude to every tyrannical collectivistic government, from Germany’s totalitarianism, Russia’s communism - and to every religious system that holds to brutally punishing apostates that leave its clutches. In contrast, we see in Nehemiah’s wall, a story where everyone from the greatest to the least is given indispensable roles in building the wall, are named family by family, and are given the human dignity and privilege of building the project in unity as well as in their own unique way.

AI as Augmentation or Replacement

The choice of whether to build AI as a tool that “complements” humans (as in the Nehemiah story) or one that “replaces” them (as in the Babel story) isn't a technical inevitability — it's a design decision. From the outset, AI has carried two rival ambitions. Dartmouth founders such as McCarthy, Minsky, Rochester and Shannon (1955), and Newell and Simon (1976) pursued substitution, where a machine does what a mind does. Simon forecast in 1965 that machines would perform any work a human could do within two decades. Minsky treated the brain as a mechanism whose functions were transferable in principle; and I. J. Good's "intelligence explosion" supplied the recursive logic that any such machine would soon surpass its makers. Running parallel, and drawing on the same ARPA funding, was the augmentation tradition of Bush (1945), Licklider (1960) and Engelbart (1962), which held that the computer's purpose was to raise the ceiling on human problem-solving rather than to replicate it — a programme which gave rise to the mouse, hypertext and collaborative editing. The two were never cleanly opposed – for example, Licklider himself framed symbiosis as an interim arrangement preceding machine autonomy. Dissent arrived early and from inside, with Wiener (1960) giving the first statement of what is now called the alignment problem; and Weizenbaum (1976) argued that certain judgments should not be delegated regardless of machine competence — an objection about propriety, not capability. Both lineages persist today. In the words of Acemoglu and Johnson, “the drive toward automation is perilous, and for AI to support shared prosperity it must complement workers rather than replace them” (Acemoglu et al., 2023); this is seen as an explicit objective in agentic systems which are evaluated on autonomous task completion (e.g. METR's time-horizon benchmarks). The augmentation view on the other hand is that one can treat work as bundles of tasks; instead of using a technology to displace a task, one can use it to complement it and even create new labour demand — there is a choice between augmenting and automation at the division between machine prediction and human judgment (Agrawal et al., 2018). Indeed, technologies designed to augment human capabilities rather than simply imitate or replace human labour can produce broader gains (Brynjolfsson, 2022).

The Choices Are Ours

In Bill Gates’ words, “The turbulent AI era is here. The choices we make now are critical.” Nothing about AI eroding jobs or delivering shared prosperity is inevitable; these are choices made by developers, firms and policymakers (Ren et al., 2025). The choice before us is whether AI becomes a technology that makes human beings cheaper, or one that makes human beings more capable.

That distinction matters. If machines can perform more of what we once considered uniquely valuable human work, we face a dangerous temptation: to measure people by what remains economically useful after the machines have taken their share. But that would be to mistake technological progress for human progress.

From Scientific Finding to the Moral Question

Dissecting the above discussion into its key scientific arguments, we have the following:

Research finding: LLM interaction can produce cognitive offloading.

Hypothesis: This may weaken epistemic vigilance under certain conditions.

Inference: Societies could therefore become increasingly dependent on AI-mediated judgments.

Normative question: Should human beings delegate those judgments?

A Judeo-Christian biblical answer: Human dignity imposes limits on such delegation.

And herein lies the rub:

AI Must Not Make Us Less Necessary to One Another.

CONCLUSION

AI may eventually make us less necessary to the economy. It must never make us less necessary to one another. AI may transform what we do, but it must never become an excuse for humans to stop thinking, judging and seeking truth. The challenge is not simply to ensure that machines remain our servants rather than our masters, but to ensure that, in becoming ever more capable of doing our thinking for us, they do not tempt us to abandon the dignity of thinking for ourselves.

As G. K. Chesterton put it, “The object of opening the mind, as of opening the mouth, is to shut it again on something solid.”

References

McGilchrist, I. (2021). The matter with things: Our brains, our delusions, and the unmaking of the world (2 vols.). Perspectiva Press.

Guingrich, R. E., Mehta, D. and Bhatt, U. (2026). Belief Offloading in Human-AI Interaction}. Available from: https://arxiv.org/abs/2602.08754. URL

Maynard, E.D. (2026). The AI Cognitive Trojan Horse: How Large Language Models May Bypass Human Epistemic Vigilance. Available from: https://arxiv.org/abs/2601.07085. URL

Vaswani, A., et al. (2017). Attention Is All You Need. NeurIPS.

Radford, A., et al. (2019). Language Models are Unsupervised Multitask Learners. OpenAI Technical Report.

Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. NeurIPS.

Raffel, C., et al. (2020). Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. JMLR.

Gao, L., et al. (2020). The Pile: An 800GB Dataset of Diverse Text for Language Modeling. arXiv:2101.00027.

Kaplan, J., et al. (2020). Scaling Laws for Neural Language Models. arXiv:2001.08361.

Hoffmann, J., et al. (2022). Training Compute-Optimal Large Language Models. NeurIPS (arXiv:2203.15556).

Lee, K., et al. (2022). Deduplicating Training Data Makes Language Models Better. ACL.

Penedo, G., et al. (2023). The RefinedWeb Dataset for Falcon LLM. NeurIPS Datasets & Benchmarks (arXiv:2306.01116).

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots. FAccT.

Zhao, A., Wu, Y., Yue, Y., et al. (2025). Absolute Zero: Reinforced Self-Play Reasoning with Zero Data. NeurIPS 2025 (arXiv:2505.03335).

Goldie, A., Mirhoseini, A., Zhou, H., Cai, I., & Manning, C. D. (2025). Synthetic Data Generation & Multi-Step RL for Reasoning & Tool Use. arXiv:2504.04736.

Huang, C., Yu, W., Wang, X., et al. (2026). R-Zero: Self-Evolving Reasoning LLM from Zero Data. ICLR 2026.

Guo, Y., Guo, Z., Huang, C., et al. (2025). Synthetic Data RL: Task Definition Is All You Need. arXiv:2505.17063.

OpenAI (2025). GPT-5 System Card. cdn.openai.com/gpt-5-system-card.pdf — training-data volume and sources not disclosed.

Drop Site News (2026). Israel Is Paying Millions to Train AI Chatbots How to Talk About Gaza. — Common Crawl as the shared training backbone of ChatGPT, Gemini, and Claude.

Washington Post (2026), as reported in IBTimes UK. Inside Project Panama, Anthropic's Effort to Scan and Shred the World's Books. — book acquisition/scanning for Claude.

The Decoder (2026). Copilot Goes Cheap as Microsoft Phases Out OpenAI and Anthropic Models. — Microsoft MAI models trained on Common Crawl; Copilot layered over OpenAI/Anthropic.

Lorenz, J., Rauhut, H., Schweitzer, F., & Helbing, D. (2011). "How social influence can undermine the wisdom of crowd effect." Proceedings of the National Academy of Sciences (PNAS), 108(22), 9020–9025.

Asch, S. E. (1951). "Effects of group pressure upon the modification and distortion of judgments." In H. Guetzkow (Ed.), Groups, Leadership and Men (pp. 177–190). Carnegie Press.

Asch, S. E. (1955). "Opinions and social pressure." Scientific American, 193(5), 31–35.

Asch, S. E. (1956). "Studies of independence and conformity: A minority of one against a unanimous majority." Psychological Monographs: General and Applied, 70(9), 1–70.

Lu, L., Yuan, Y. C., & McLeod, P. L. (2012). "Twenty-Five Years of Hidden Profiles in Group Decision Making: A Meta-Analysis." Personality and Social Psychology Review, 16(1), 54–75.

Stasser, G., & Titus, W. (1985). "Pooling of unshared information in group decision making: Biased information sampling during discussion." Journal of Personality and Social Psychology, 48(6), 1467–1478.

Myers, D. G., & Lamm, H. (1976). "The group polarization phenomenon." Psychological Bulletin, 83(4), 602–627.

Moscovici, S., & Zavalloni, M. (1969). "The group as a polarizer of attitudes." Journal of Personality and Social Psychology, 12(2), 125–135.

Sunstein, C. R. (2002), "The Law of Group Polarization," Journal of Political Philosophy, 10(2), 175–195.

Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). "A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades." Journal of Political Economy, 100(5), 992–1026.

Banerjee, A. V. (1992). "A Simple Model of Herd Behavior." Quarterly Journal of Economics, 107(3), 797–817.

Pope Leo XIV. (2026). Available from: https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html. URL

Netziv (Rabbi Naftali Zvi Yehuda Berlin), Ha'amek Davar on Genesis 11:4 (comment on pen nafutz). Available on Sefaria (העמק דבר על בראשית י״א:ד׳).

Pirkei DeRabbi Eliezer 24 (§24:6), teaching of R. Phineas. Available on Sefaria (פרקי דרבי אליעזר כ״ד:ו׳).

Bush, V. (1945). "As We May Think." The Atlantic Monthly, 176(1), 101–108.

Engelbart, D. C. (1962). Augmenting Human Intellect: A Conceptual Framework. Summary Report AFOSR-3223, Stanford Research Institute, Menlo Park, CA.

Good, I. J. (1965). "Speculations Concerning the First Ultraintelligent Machine." Advances in Computers, 6, 31–88.

Jordan, M. I. (2019). "Artificial Intelligence — The Revolution Hasn't Happened Yet." Harvard Data Science Review, 1(1).

Licklider, J. C. R. (1960). "Man-Computer Symbiosis." IRE Transactions on Human Factors in Electronics, HFE-1, 4–11.

Markoff, J. (2015). Machines of Loving Grace: The Quest for Common Ground Between Humans and Robots. New York: Ecco.

McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. (Reprinted in AI Magazine, 27(4), 2006, 12–14.)

METR (2025). "Measuring AI Ability to Complete Long Tasks." Model Evaluation and Threat Research. https://metr.org

Minsky, M. L. (1986). The Society of Mind. New York: Simon & Schuster.

Agrawal, A., Gans, J., & Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press.

Brynjolfsson, E. (2022). The Turing Trap: The Promise and Peril of Human-Like Artificial Intelligence. Dædalus, 151(2), 272–287.

Acemoglu, D., & Johnson, S. (2023). Rebalancing AI. Finance & Development. International Monetary Fund. https://www.imf.org/en/publications/fandd/issues/2023/12/rebalancing-ai-acemoglu-johnson. URL

Acemoglu, D., & Restrepo, P. (2018). The Race Between Man and Machine: Implications of Technology for Growth, Factor Shares, and Employment. American Economic Review, 108(6), 1488–1542.

Autor, D. H. (2015). Why Are There Still So Many Jobs? The History and Future of Workplace Automation. Journal of Economic Perspectives, 29(3), 3–30.

Ren, S., et al. (2025). AI, Jobs, and the Automation Trap: Where Is HCI? arXiv:2501.18948.

ABOUT THE AUTHORS

Shaun Johnson is a control engineer, having completed his Bachelor of Engineering (BEng) in Electrical and Electronic Engineering; he has further completed his Masters in Information Technology with focus on Big Data Science at the University of Pretoria in 2024, and is currently a senior data scientist in the mining, minerals and metals domain.

Dr Celeste Johnson completed her PhD in theoretical physics at the University of the Witwatersrand, Johannesburg. More recently, her focus has been on mathematics of finance, and she is currently a senior quantitative analyst.

Celeste and Shaun Johnson are the proud parents of three children.