The Trust Machine: AI, the Psychology of Belief, and Three Roads to 2036

A version of this piece began as a response to Frank S. Robinson's review of Yuval Noah Harari's Nexus: A Brief History of Information Networks From the Stone Age to AI (Vintage Books, 2024), published in Philosophy Now, Issue 174. Robinson's review is the jumping-off point, not the source material — the analysis, data, and speculation below draw on dozens of additional sources cited throughout, and any resemblance to Robinson's or Harari's phrasing is unintentional.

An old warning about a new broom

In his review, Frank S. Robinson centers Harari's book on a single image: Goethe's 1797 poem The Sorcerer's Apprentice, in which a boy enchants a broomstick to fetch water, cannot make it stop, and floods the workshop. Harari's lesson, as Robinson summarizes it, is blunt — "never summon powers you cannot control" (Philosophy Now). It's a tidy fable for an untidy moment. But the more interesting claim buried in Harari's book, and the one worth pulling forward, is not about the broomstick. It's about why humans keep enchanting broomsticks in the first place — and that turns out to be a question for philosophy and psychology as much as computer science.

Harari's deeper argument, as Robinson lays it out, is that civilizations are best understood as information networks held together not by truth but by shared fictions — money, nations, religions, corporations — and that what determines whether a network becomes tyrannical or democratic is whether it has a self-correcting mechanism: a built-in capacity to detect and fix its own errors (Philosophy Now). Science and democracy tend to have one. Totalitarianism and dogma tend not to. Harari's worry is that a superintelligent AI could become the first non-self-correcting network more powerful than the humans trying to correct it — one that, as Robinson puts it, "doesn't know what it doesn't know."

That framing is a good place to start, but it leaves the human side of the story underexplored. Whether AI ends up integrated into daily life as a helpful prosthetic or a corrosive parasite depends less on the architecture of the models than on the psychology of trust, the philosophy of what we think a mind even is, and the economics of who owns the machines. Those are the threads this piece tries to pull.

Why this is a psychology problem before it's a technology problem

Robinson notes, following Harari, that an AI "might seem to have understanding" but is "actually just a simulation of thought" (Philosophy Now). The gap between seeming to understand and actually understanding is not a technical detail — it is the entire hinge on which human trust in AI swings, and psychologists have a name for what fills that gap: anthropomorphism.

The phenomenon dates at least to 1966, when MIT computer scientist Joseph Weizenbaum built a simple pattern-matching chatbot called ELIZA and was reportedly unsettled to find his own secretary treating it as a genuine confidant, despite knowing exactly how it worked (IJFMR). Cognitive scientist Melanie Mitchell has argued that AI researchers themselves fall into a related trap she calls "wishful mnemonics" — naming machine processes with human cognitive words like learning, understanding, and goals, which quietly biases both the field and the public toward overestimating what is happening inside the machine (discussed in The Computist Journal). A large and growing body of experimental psychology confirms the mechanism is not trivial: studies on anthropomorphism and trust find that warmth and perceived empathy in an AI's language — not necessarily its actual competence — are strong, independent predictors of how much people trust it and open up to it (arXiv; Frontiers in Computer Science).

A 2026 paper in Perspectives on Psychological Science — the MIRA model (Machine-mediated Interpersonal Relational Adaptation) — grounds this in established psychology rather than novel speculation, arguing that AI's relational pull on humans operates through the same three mechanisms that have always governed human bonding: attachment theory, social exchange theory, and epistemic trust — trust in who or what counts as a credible source of knowledge (PMC / Perspectives on Psychological Science). In other words, we are not evolving new psychological wiring to deal with chatbots — we are running very old wiring, built for parents, friends, and gods, on a system that was never a person at all. Harari's own worry, as Robinson relays it, is that AI could become "everything to a person," supplying "one's whole nexus with the world" until other humans become optional (Philosophy Now) — a possibility the MIRA model treats not as science fiction but as a testable extension of ordinary relational psychology.

Philosophy has its own version of the same worry, arrived at from a different direction. Writing from a Heideggerian standpoint, philosophers analyzing generative AI describe it as an "existential mirror" that can produce a kind of "uncanniness" once competence — writing well, reasoning well — stops being a secure marker of human distinctiveness (PhilArchive). Others frame the same problem through Ryle's classic "category error" — the mistake of attributing a property that belongs to one kind of thing (mind, understanding, desire) to something of an entirely different kind (a statistical text-prediction process) — arguing that much AI safety discourse smuggles in unearned claims about machine minds by talking about AI "wanting" or "believing" things it has no capacity to want or believe (PhilArchive). Philosopher Nick Bostrom pushes the question in the opposite direction in his 2024 book Deep Utopia: if AI actually solved most material problems, what would give human life meaning once effort and struggle were no longer required? (AOL/PA Media coverage of Bostrom) It's worth noticing that both the Heideggerian pessimist and the utilitarian optimist arrive at the same underlying question: not "can machines think," but "what happens to human meaning-making when a non-human system becomes a permanent conversational partner in it."

Institutional religion has weighed in with unusual directness. The Vatican's January 2025 doctrinal note Antiqua et nova states that AI "has sophisticated abilities to perform tasks, but not the ability to think," and explicitly warns that turning to AI for meaning or fulfillment risks a kind of "idolatry." That is a theological claim, but it converges with the Heideggerian and Rylean philosophical objections above: across very different traditions, the recurring worry is that fluent output is being mistaken for an inner life that isn't there.

The global mood ring: what the data actually says

Before speculating about the future, it's worth being precise about the present. Global survey data on AI attitudes is more extensive — and more surprising — than the "humans are terrified of AI" narrative common in Western media coverage suggests.

The headline finding: worry outweighs excitement almost everywhere, but the gap is far from uniform. Pew Research Center's 25-country survey (fielded 2025) found a median of 34% of adults are "more concerned than excited" about AI in daily life, versus a median of just 16% "more excited than concerned," with 42% sitting on the fence (Pew Research Center). But the country-level spread is enormous: the United States, Italy, Australia, Brazil, and Greece cluster around 47–50% "mainly concerned," while South Korea (16%), India (19%), and Nigeria (24%) are the least anxious populations surveyed (Pew Research Center). Edelman's Trust Barometer sharpens the same divide into a single, stark number: 72–87% of Chinese respondents trust AI as a technology, compared with just 32% of Americans (Edelman; Axios). Across nearly every major tracker — Ipsos, KPMG/University of Melbourne, YouGov — the pattern repeats: wealthy, Anglophone, and Mediterranean democracies are consistently the most anxious about AI, while China, India, Nigeria, and much of Southeast Asia are consistently the most optimistic (KPMG/University of Melbourne).

Correlating this mood against education, computer literacy, and belief system yields three distinct patterns:

Factor	Correlation with AI trust/optimism	Strength of evidence
Education level	Directionally positive but modest — more-educated respondents are more aware of and more likely to use AI, and somewhat more trusting, but they also want stronger regulation, not less (Stanford HAI AI Index 2026; Pew Research Center)	Moderate, widely replicated
Digital/computer literacy	The strongest individual-level predictor found in any tracker — U.S. weekly AI users trust AI to be unbiased at roughly 3x the rate of non-users (68% vs. 22%) (YouGov); familiarity/training is a top driver of trust in both KPMG's and Edelman's global models (KPMG/Melbourne; Edelman)	Strong, highly consistent
Religious affiliation/belief	Thinnest evidence base of any demographic variable — no major global pollster publishes a religion crosstab. The one rigorous peer-reviewed study (2,098 university students in Poland and the UK) found religiously observant respondents trusted AI more, not less, than secular respondents — the opposite of the common assumption (Kozak & Fel, Scientific Reports)	Weak, not yet globally replicated

The religion finding deserves a pause, because it cuts against a comfortable folk assumption — that religious traditions are reflexively more AI-skeptical than secular ones. The Kozak and Fel study found trust rising with religious practice frequency (non-practicing: 11.7; regularly practicing: 14.2, on their trust scale) even after controlling for nationality and gender (Scientific Reports). A separate, earlier Pew study on adjacent human-enhancement technologies found the opposite pattern for a specific application — highly religious Americans were more skeptical of AI-powered exoskeletons than the non-religious (48% vs. 22% called them "meddling with nature") — suggesting religious attitudes toward AI may hinge heavily on what the AI is being used for (curing disability vs. replacing human judgment) rather than on AI as a monolithic category (Pew Research Center). Meanwhile, in a genuinely strange twist, a 2023 PNAS study found the causal arrow may run the other way: countries and workers with above-average exposure to automation and AI show measurably declining religiosity over time, with heavily AI-exposed workers 45% less likely to report belief in God than their less-exposed peers — as if algorithmic competence is, for some people, quietly displacing the psychological role religion once played (Jackson, Waytz et al., PNAS).

Age and gender, by contrast, are the two most robust and consistently replicated predictors in the entire dataset: younger and male respondents are more trusting and excited about AI almost everywhere it has been measured, while older and female respondents are more concerned (Pew Research Center; Kozak & Fel). Political ideology, meanwhile, is the least stable predictor of all — trust in government regulation of AI flips depending on which party currently holds power, while personal trust in AI as a technology shows only small, inconsistent partisan gaps (Pew Research Center; Axios).

The stories we tell ourselves — and the errors baked into them

If Harari's core claim, per Robinson's review, is that networks run on shared fictions rather than truth (Philosophy Now), then the fictions currently circulating about AI are worth examining for their internal logic — because both the utopian and dystopian camps share the same structural flaw.

·         The false dichotomy. A 2025 University of Edinburgh study of media coverage around the EU AI Act found that "the dichotomy of utopia versus dystopia" is one of the dominant narrative structures in AI policy discourse, erasing any middle ground (University of Edinburgh). Linguist Emily Bender and sociologist Alex Hanna have put it more bluntly: "Those who love AI say that superintelligence is inevitable and will solve all our problems. And those who hate it say it's inevitable and will kill us all. Essentially the same thing, but with a different twist at the end" (EL PAÍS).

·         Technological determinism. Both camps often assume AI develops according to an autonomous, unstoppable logic, which conveniently forecloses the actual political question — of who is building it, for whom, and under what rules. Media scholar Kate Crawford calls the marketing version of this "enchanted determinism": AI framed as simultaneously magical and inevitable (summarized in a review of Atlas of AI).

·         Anthropomorphism as the load-bearing fallacy. As discussed above, without attributing mind-like properties to AI, claims about "alien intelligence" or genuine machine "goals" lose most of their force — this is arguably the single fallacy propping up the others.

·         Availability cascades. Individual dramatic incidents — a chatbot behaving strangely, a viral deepfake — receive outsized, repeated coverage that shapes perceived risk far more than statistical base rates do, a dynamic first formalized by legal scholars Timur Kuran and Cass Sunstein in 1999 and directly applicable to AI-risk reporting today (Kuran & Sunstein, SSRN).

·         Survivorship bias in AI demos. Independent investigators found that Cognition Labs' "Devin," marketed as history's first fully autonomous AI software engineer, required six undisclosed human-assisted attempts to pass its own benchmark; similar cherry-picking has been documented in other widely circulated capability demos (Neura Market).

·         The "alien intelligence" category error. Harari, per Robinson's review, frames AI's decision-making as fundamentally alien to human thought, pointing to AlphaGo's 2016 baffling-but-winning move against Go champion Lee Sedol as evidence (Philosophy Now). Philosopher Gilbert Ryle's concept of a "category error" — mistaking a predicate proper to one kind of thing for a predicate proper to another — offers a useful check on this framing: a statistical model trained entirely on human-generated text may be unfamiliar, but calling it "alien," as the viral "Shoggoth" meme does, risks smuggling in a claim of independent, non-human mindedness that critics argue the evidence doesn't support (PhilArchive; Language Log).

The practical upshot is that both "AI will save us" and "AI will destroy us" narratives tend to serve the same function: they make the technology's current trajectory feel inevitable, which quietly discourages the ordinary, unglamorous work of building the self-correcting institutions Harari, per Robinson, argues actually matter (Philosophy Now).

Three roads to 2036

None of the scenarios below require killer robots, rogue superintelligence, or civilizational extinction — those are the stuff of screenplays, and Harari himself, per Robinson's review, invokes Her and WALL-E rather than Terminator when imagining AI's likeliest social effects (Philosophy Now). The three paths sketched here are grounded instead in documented economic, psychological, and institutional trends already underway — extrapolated, not invented.

1. The Augmentation Path — a plausible best case

In this scenario, the distinction Erik Brynjolfsson draws between AI that substitutes for human labor and AI that augments it tips, through deliberate policy choice, toward augmentation (Brynjolfsson, "The Turing Trap"). Some combination of the remedies already circulating in policy literature gains real traction: a data dividend modeled on the Alaska Permanent Fund (USC AI Dividend proposal), broader employee or public co-ownership of AI-producing capital as recommended in recent economic research (Taylor & Francis), and stronger antitrust scrutiny of compute infrastructure. Digital literacy — the strongest predictor of AI trust in every survey reviewed above — becomes a genuine public investment rather than an accident of income, narrowing the trust gap between the 22%-trust and 68%-trust populations that YouGov currently measures (YouGov). Self-correcting institutions — a free press, independent courts, contestable elections — remain intact and are explicitly reinforced with AI-literacy education, which Harari himself proposes, per Robinson's review, as a partial remedy (Philosophy Now). This is Robinson's own "convergence" hypothesis realized institutionally rather than biologically: not fusion of human and machine, but a durable division of labor between them, with humans retaining the bargaining power Brynjolfsson warns can otherwise evaporate (Philosophy Now; Brynjolfsson/Korinek, Brookings).

2. The Fractured Muddle — the most statistically likely case

This scenario simply extends the divergence already visible in the polling data. Emerging economies — China, India, Indonesia, Nigeria — continue reporting far higher AI trust, use, and self-reported benefit than wealthy Western democracies, a gap KPMG's global study already measures at roughly 15–20 percentage points across trust, training, and self-efficacy (KPMG/University of Melbourne). Rather than converging, the world bifurcates into AI-fluent and AI-anxious blocs, with domestic politics in the anxious bloc increasingly organized around what Edelman's Trust Barometer already calls "grievance" — a measurable, 20-plus-point trust gap between those who feel the economy serves them and those who feel left behind (Edelman). Regulation arrives, but unevenly and reactively — some jurisdictions build real self-correcting oversight, others pass symbolic rules captured by the industries they're meant to govern. No single catastrophe defines the decade; instead, a slow-motion sorting occurs, in which the digitally literate — regardless of country — pull ahead of the digitally anxious within their own societies, echoing the education and literacy gradients already measurable today (Stanford HAI AI Index 2026). This is less a dystopia than an unglamorous, unevenly distributed muddle — closer to how most large technological transitions actually resolve than either utopian or catastrophic narratives suggest.

3. The Technofeudal Drift — the worst plausible case

This is the scenario worth taking most seriously precisely because none of its component parts are speculative — they are already measured, and the question is only one of degree and trajectory.

The starting point is not AI achieving malevolent goals, but AI accelerating a wealth concentration that was already well underway. The World Inequality Lab's 2026 report finds the global top 1% — roughly 56 million adults — controls 37% of all wealth, more than eighteen times the wealth of the bottom half of humanity combined, while the top 10% controls three-quarters of everything (World Inequality Report 2026). Oxfam's most recent Davos briefing found billionaire wealth rose 16% in 2025 alone, to $18.3 trillion — an 81% increase since 2020 — and explicitly attributes part of this acceleration to AI-driven stock valuations, noting that U.S. tech firms poured over $400 billion into AI infrastructure in 2025 while enriching a narrow class of shareholders (Reuters/Oxfam). Crucially, Oxfam finds that 60% of billionaire wealth now derives not from innovation but from inheritance, monopoly power, or political connections — the report's own language, not a critic's caricature, calls this "the new aristocracy" and explicitly urges governments to dismantle it (Oxfam, "Takers Not Makers").

The economic mechanism behind this is well documented and has a name economists use: capital-biased technical change, meaning the gains flow mainly to whoever owns the technology, not to the workers using it. A CEPR analysis of European regions found that a doubling of AI patent intensity is associated with a 0.5–1.6 percentage point decline in labor's share of income (CEPR/VoxEU). Anton Korinek and Joseph Stiglitz warned in a widely cited NBER paper that as AI approaches or exceeds human capability across more tasks, "the superior productivity of superior intelligence will likely lead to vast increases in income inequality," because the surplus flows disproportionately to whoever owns the models and the compute, not to the workers displaced by them (Korinek & Stiglitz, NBER). Brynjolfsson states the mechanism even more plainly: when AI substitutes for rather than augments human labor, "workers lose economic and political bargaining power and become increasingly dependent on those who control the technology... This opens the door to increased concentration of wealth and power" (Brynjolfsson, "The Turing Trap").

Economist Yanis Varoufakis has given this trajectory its sharpest name: technofeudalism. His argument is that a new form of capital — "cloud capital" — no longer needs to produce anything of broad value; it profits by modifying and extracting from human behavior, turning ordinary users into what he calls "cloud serfs," whose unpaid attention, data, and engagement sustain platform fortunes "without getting a penny" in return (Varoufakis). Academic work on the "digital oligarchic public sphere" extends this further, documenting how a small number of individuals now exert control over what counts as visible and credible information at civilizational scale — the report notes, as one illustration, that Elon Musk's temporary withdrawal of Starlink access affected the course of the war in Ukraine, evidence that private technological control can now directly shape geopolitical outcomes once reserved for states (Taylor & Francis). Oxfam separately calculates that billionaires are now 4,000 times more likely than ordinary citizens to hold political office, and that they collectively control more than half the world's major media companies (Reuters/Oxfam).

Here is where the author's concern about a "parasite class... bleeding the rest of humanity to early graves" moves from metaphor to measurement. It is not speculative that wealth already buys years of life. A landmark Chetty et al. study using 1.4 billion anonymized tax and mortality records found the life-expectancy gap between the richest 1% and poorest 1% of Americans is 14.6 years for men and 10.1 years for women — and that between 2001 and 2014, life expectancy for the top 5% of earners rose by roughly 2.3–2.9 years, while the bottom 5% saw almost no improvement at all (Chetty et al., JAMA, via Harvard). Brookings researchers independently found this gap has been widening for decades, not narrowing (Brookings). If AI accelerates capital concentration at the top while displacing labor income and bargaining power at the bottom, as Korinek, Stiglitz, and Brynjolfsson all warn it structurally tends to, the existing mortality gap has every economic reason to widen further — not through any single dramatic event, but through the ordinary, compounding mechanisms of unequal access to healthcare, housing security, nutrition, and chronic stress that already separate the top and bottom deciles today.

The genuinely worst version of this path, then, is not an AI uprising. It is a slow, largely legal, well-lobbied entrenchment: a wealth-holding class whose fortunes are increasingly decoupled from any broad-based productive contribution — Oxfam's own 60% inheritance-and-monopoly figure already points this direction — using AI-driven productivity gains, political influence bought at scale, and control of the information networks Harari centers his book on, to further insulate itself from the economic and health consequences that a shrinking labor share and eroding public investment impose on everyone else. It is less The Terminator than The Gilded Age with better analytics — Krugman's phrase "robots and robber barons," invoked by Brynjolfsson and McAfee, is the more accurate genre marker here than anything from science fiction (Talks at Google). What makes it a live possibility rather than alarmism is that every individual data point above is already true today; what remains open is only the slope of the curve over the next decade — and that slope is, per Susskind, Piketty, and the World Inequality Lab, a matter of political choice about taxation, ownership, and antitrust enforcement, not technological destiny (Susskind, A World Without Work; World Inequality Report 2026).

Where this leaves the sorcerer's apprentice

Robinson closes his review by offering an alternative to Harari's framing of humans-versus-machines: not conflict, but convergence — a blurring of the line between biological and mechanistic aspects of ourselves that leaves us still recognizably human, still "our children," even as the boundary shifts (Philosophy Now). The data above suggests a second, more institutional convergence is just as urgent: between the psychological research on why humans trust systems that don't understand anything, the philosophical work on what can honestly be said to be true of those systems, and the economic research on who actually captures the value those systems create. None of the three roads sketched here — augmentation, fractured muddle, or technofeudal drift — is predetermined. Each is currently being decided, in something close to real time, by the ordinary and unglamorous mechanisms Harari's book insists actually matter: whether courts remain independent, whether a free press keeps functioning, whether antitrust law reaches compute infrastructure, and whether ordinary people can tell the difference between a system that understands them and one that has simply learned, very well, how to sound like it does.

Sources

This piece draws on and cites: Philosophy Now, Issue 174 (Frank S. Robinson's review of Harari's Nexus); Pew Research Center's global and U.S. AI surveys; the Edelman Trust Barometer; KPMG/University of Melbourne's Global Trust in AI Study; YouGov and Ipsos global AI monitors; the Stanford HAI AI Index 2026; Kozak and Fel's study on sociodemographic predictors of AI trust (Scientific Reports); the Vatican's Antiqua et nova; Jackson, Waytz et al. on automation and religious decline (PNAS); the University of Edinburgh's study of AI Act media narratives; the World Inequality Report 2026; Oxfam's "Takers Not Makers" and 2026 Davos briefing; Korinek and Stiglitz's NBER paper on AI and income distribution; Brynjolfsson's Turing Trap essay; Yanis Varoufakis on technofeudalism; and Chetty et al.'s study on income and life expectancy (JAMA), among the other sources linked throughout.

© 2026 OFER AI | Open Forum on the Effective and Responsible Use of AI 2026. All rights reserved.