The Real Hallucination Problem Isn’t Artificial — It’s Us

We’ve been wringing our hands about the wrong hallucinations.

Open any tech publication, tune into any congressional hearing, scroll any LinkedIn feed, and you’ll find a chorus of voices sounding the alarm about AI hallucination — the tendency of large language models to occasionally fabricate a citation, invent a statistic, or confidently state something false. It’s a real limitation. It deserves engineering attention. But the breathless, hand-over-mouth panic surrounding it has become its own kind of irony, because it distracts from a far older, far deadlier phenomenon that no one seems eager to discuss: the capacity of entire human societies to collectively hallucinate — and to act on those hallucinations with catastrophic, world-altering consequences.

AI gets a fact wrong and we convene a Senate hearing. Humans get reality wrong together and we get wars, famines, and genocide. If we’re serious about reducing preventable suffering on this planet, perhaps the most urgent question isn’t how to stop machines from occasionally confusing facts — it’s how to stop us from doing it at civilizational scale. And perhaps, counterintuitively, the logic machine we’ve built is one of the most powerful tools we’ve ever had to help.

The Science of Getting It Wrong Together

This isn’t armchair philosophizing. The science of collective cognitive failure is well established and deeply sobering.

In the early 1970s, psychologists Daniel Kahneman and Amos Tversky launched the heuristics and biases research program, demonstrating that human reasoning is riddled with systematic errors — not random mistakes, but predictable, repeatable failures hardwired into how we process information. Confirmation bias, anchoring, the availability heuristic, motivated reasoning — these aren’t bugs in a few flawed individuals. They are features of the human cognitive architecture, present in experts and laypeople alike. Kahneman’s Nobel Prize-winning work showed that even trained statisticians and professional forecasters fall prey to these distortions when relying on intuition.

Philip Tetlock, a professor at the University of Pennsylvania’s Wharton School, extended these findings into the world of expert judgment. In a landmark 2005 study spanning two decades and tens of thousands of predictions, Tetlock found that the average expert’s forecasts were barely better than random chance — or, in his memorable framing, roughly as accurate as a dart-throwing chimpanzee. Pundits, political analysts, and policy advisors who confidently shaped public opinion and government decisions were operating with almost no predictive advantage over flipping a coin.

If individual cognition is this unreliable, collective cognition can be worse. Social psychologist Irving Janis demonstrated this with his theory of groupthink, published in Victims of Groupthink in 1972. Studying catastrophic U.S. foreign policy decisions — the failure to anticipate the attack on Pearl Harbor, the Bay of Pigs invasion, the escalation of the Vietnam War — Janis showed that cohesive groups of highly intelligent, well-credentialed decision-makers routinely suppressed dissent, ignored contradictory evidence, and converged on disastrous conclusions precisely because they were a group. The social pressure to maintain consensus overrode each individual’s capacity for independent judgment.

Groupthink doesn’t just produce bad meetings. It produces body counts.

The Ledger: What Human Hallucination Has Actually Cost Us

The cost of collective human delusion isn’t theoretical. It’s denominated in corpses, in displaced populations, in civilizations scarred for generations. The numbers are staggering, and they deserve to be stated plainly.

The Crusades (1095–1291)

The shared hallucination: God willed the violent reconquest of the Holy Land, and killing people of a different faith was sacred duty.

For nearly two centuries, European armies marched to the Middle East under a collective delusion reinforced by papal authority, feudal obligation, and religious fervor. Entire cities were sacked. Thousands of Jews were massacred in the Rhineland before the Crusaders even reached the Holy Land. This was not a failure of intelligence; it was a failure of collective perception, dressed in religious certainty.

The toll:

• 1–3 million dead — estimates range from 1 million (Wertham, Hittell) to 3 million (Garrison, Matthew White) to as high as 9 million (Robertson), with modern historians cautioning against precision over a 200-year span
• Entire cities — Jerusalem, Antioch, Constantinople — sacked and their populations massacred
• Thousands of Jews killed in the Rhineland Massacres of 1096 before the First Crusade even left Europe

The Holocaust (1941–1945)

The shared hallucination: Jewish people were subhuman parasites responsible for Germany’s economic and moral suffering.

This narrative was not a fringe belief. It became the consensus reality of an educated, industrialized European nation and was reinforced by government, academia, media, law, and culture simultaneously. Millions participated directly; millions more watched in silence. The shared false belief was so total, so institutionally reinforced, that disbelief itself became socially dangerous.

The toll:

• 6 million Jews systematically murdered
• 5 million additional victims — Roma, disabled people, political prisoners, Soviet POWs, and others
• An entire continent’s moral and intellectual infrastructure bent toward industrialized extermination

The Irish Famine (1845–1852)

The shared hallucination: The Irish poor were morally deficient, and laissez-faire economics was a higher imperative than preventing mass death.

Ireland was exporting enormous quantities of grain throughout the famine — enough to feed the starving population. Up to 75% of Irish soil was devoted to crops grown for export. But a prevailing ideology among the British ruling class treated market forces as sacred and the famine as divine judgment on Irish character. The collective hallucination was not that potatoes had failed. It was that an entire people didn’t quite deserve to be saved.

The toll:

• ~1 million dead from starvation and famine-related disease (estimates range from 800,000 to 1.5 million)
• 2.1 million emigrated between 1845 and 1855 — one of the greatest population displacements in modern history
• Ireland’s population fell from 8.2 million to 6.5 million in a single decade
• Ireland’s population has still not recovered — remaining below pre-famine levels nearly 180 years later
• Britain spent £8 million on Irish famine relief over the entire crisis — but £69 million on the Crimean War shortly afterward

The Iraq War (2003–2011)

The shared hallucination: Iraq possessed weapons of mass destruction that posed an imminent threat to the Western world.

Intelligence was cherry-picked, dissent was marginalized, media fell in line, and populations were swept into consensus. The weapons did not exist. The war happened anyway.

The toll:

• ~200,000 Iraqi civilians killed directly (Iraq Body Count project)
• 461,000 excess deaths in Iraq through 2011 (PLOS Medicine, 2013)
• 940,000+ total deaths across post-9/11 war zones (Brown University Costs of War project)
• 3.6–3.8 million additional indirect deaths from infrastructure destruction, displacement, and economic collapse (Brown University)
• 4,492 U.S. servicemembers killed; 32,292 wounded
• Roughly 40% of Iraq’s middle class fled the country (UN, 2007)
• The destabilization directly contributed to the rise of ISIS

The Covid-19 Lockdown Response (2020–2022)

The shared hallucination: COVID-19 posed essentially uniform mortal danger to all age groups, justifying the broadest possible shutdown of economic and social life.

The pandemic was real. The virus killed people. But the collective response was catastrophically miscalibrated to the actual risk profile, and the cost of that miscalibration was staggering. The data on who COVID actually endangered was available almost from the start — and it told a very specific story that the dominant narrative refused to absorb.

What the data actually showed — the risk was age-stratified, not uniform:

• 81% of U.S. COVID deaths in 2020 occurred in people aged 65 and over (CDC)
• Infection fatality rate by age (peer-reviewed meta-analysis, PubMed):
• Age 10: 0.002%
• Age 25: 0.01%
• Age 55: 0.4%
• Age 75: 4.6%
• Age 85: 15%
• The death rate for adults 85+ was 360 times higher than for adults 18–29 (CDC)
• A meta-analysis of 611,000+ patients confirmed the sharpest mortality threshold was age 60, with risk tripling between the 50–59 and 60–69 age brackets
• Among those under 70–94% of the global population — the median infection fatality rate was 0.034%

The rational response would have been targeted protection of vulnerable populations while preserving the functioning of society for the vast majority at minimal risk. Instead, the collective hallucination took hold that the threat was essentially uniform — and that the only responsible response was to shut everything down.

The economic devastation:

• 400 million full-time job equivalents lost globally (International Labour Organization)
• $3.5 trillion in lost worker income in the first 9 months of 2020 alone (ILO)
• $12.5 trillion in projected global economic losses through 2024 (IMF)
• $17–$35 trillion in estimated five-year total cost (retrospective economic analysis, 2023)
• 40–60 million people pushed into extreme poverty (United Nations)
• The estimated total cost to the U.S. alone: $16 trillion — roughly 90% of annual GDP (JAMA)

The human toll beyond economics:

• 1.6 billion children suffered some loss of education globally (UNICEF)
• 1.1 years of learning lost for every year of school closure (ScienceDirect systematic review)
• Low-income school districts were 4x less likely to recover academically than wealthy ones
• 44.2% of U.S. high school students reported persistent sadness or hopelessness (CDC Youth Survey)
• 9% of students reported attempting suicide during the pandemic period (CDC Youth Survey)
• U.S. adult depression and anxiety rates surged from 11% pre-pandemic to ~40% — an estimated 80 million additional individuals affected
• A systematic review of 61 studies covering nearly 55,000 children confirmed significant increases in anxiety, depression, and behavioral problems linked to lockdown (European Child & Adolescent Psychiatry)

The tragedy is not that governments acted in the face of uncertainty. It is that a collective narrative of undifferentiated, universal danger overwhelmed the readily available evidence of age-stratified risk — and that the policies built on that narrative inflicted their greatest damage on the populations least threatened by the virus itself: children, young adults, low-income workers, and the developing world. The hallucination was not that COVID was harmless. It was that the correct response to a disease that killed predominantly the elderly was to shut down the schools, the economies, and the social fabric of entire nations — and that anyone who questioned the proportionality of this response was treated not as a rational voice but as a moral threat.

And the list goes on.

The cases above are not outliers. They are the pattern. We could extend this ledger indefinitely:

• The transatlantic slave trade — sustained for centuries by the shared hallucination that certain races were biologically suited to enslavement, displacing an estimated 12.5 million Africans
• The subjugation of women — generations denied education, property, and political participation under the hallucination that they were intellectually unfit for public life
• The tobacco consensus — decades of manufactured agreement that cigarettes were safe, killing an estimated 100 million people in the twentieth century alone (WHO)

In each case, the catastrophe was not a failure of available information. It was a failure of collective cognition — a shared, socially enforced false belief that resisted correction because it served powerful interests or deeply held identities.

“Just Trust the Science”: When the Consensus Itself Was the Disease

Perhaps the most insidious form of collective hallucination is the one that wears a lab coat. The examples above involve governments, armies, and ideological movements — but some of the deadliest collective delusions in modern history were produced by the scientific and medical establishments themselves, endorsed by regulatory agencies, published in peer-reviewed journals, and enforced with the full authority of expert consensus. The phrase “trust the science” assumes that the science being trusted has been produced by a process immune to the same groupthink, institutional capture, and motivated reasoning that distorts every other domain of human collective judgment. The historical record suggests otherwise.

Thalidomide (1957–1961)

The consensus: A safe, mild sedative suitable for pregnant women.

Thalidomide was marketed as an over-the-counter drug in West Germany in 1957, promoted for anxiety, insomnia, and morning sickness. It was hailed as “totally safe” — despite having undergone no teratogenic testing whatsoever. The drug had been tested for toxicity only on rodents, as was standard at the time. No one tested what it did to a developing fetus.

The toll:

• 10,000–20,000 babies born with severe birth defects worldwide — shortened or absent limbs, eye damage, heart defects, malformed internal organs
• Approximately 40% died at or shortly after birth
• A single 50mg tablet during the sensitive window was sufficient to cause defects in up to 50% of pregnancies
• Described by researchers as “the biggest man-made medical disaster ever” (Birth Defects Research, PMC)
• The disaster was only identified because two independent clinicians — Widukind Lenz in Germany and William McBride in Australia — broke from the consensus in 1961
• Evidence later emerged that the manufacturer, Chemie Grünenthal, had received reports of harm and suppressed them — one of Grünenthal’s key developers had previously experimented on prisoners at Auschwitz and Buchenwald

The United States was largely spared — not because of a superior system, but because of a single FDA reviewer, Frances Oldham Kelsey, who resisted intense pharmaceutical pressure to approve the drug. One person’s refusal to defer to consensus saved thousands of American children. The consensus itself was the threat.

The Lobotomy (1935–1970s)

The consensus: Surgically severing brain tissue cures mental illness. Nobel Prize-worthy.

Portuguese neurologist António Egas Moniz introduced the prefrontal leucotomy in 1935, claiming it was “a simple operation, always safe.” In 1949, the Nobel Committee awarded him the Prize in Physiology or Medicine for this “discovery.” The procedure exploded in popularity.

The toll:

• An estimated 40,000–50,000 lobotomies performed in the United States alone between 1936 and the 1970s; 60,000+ worldwide
• Walter Freeman, the procedure’s most zealous American proponent, performed thousands using an ice pick hammered through the eye socket — sometimes in under 10 minutes, in front of an audience
• Indications expanded grotesquely — from schizophrenia to depression to anxiety to “being a badly-behaved child”
• Freeman’s own records show the procedure killed or worsened 14% of patients
• Psychologist Mary Francis Robinson assessed 90 lobotomized patients and found most could no longer concentrate, appeared unmotivated, and lost interest in their own lives — musicians stopped playing, writers stopped writing
• Rosemary Kennedy, sister of a future president, was lobotomized at age 23 and left incapacitated and institutionalized for life
• Fellow Nobel laureate Torsten Wiesel called the prize “a terrible mistake that caused permanent damage to thousands of patients”
• The lobotomy only fell out of favor after chlorpromazine was introduced in 1952 — not because the scientific establishment self-corrected, but because a better option made the barbarity impossible to ignore

The Nobel Prize has never been rescinded.

Vioxx (1999–2004)

The consensus: A revolutionary painkiller, safer than traditional alternatives.

Merck’s arthritis drug Vioxx (rofecoxib) was approved by the FDA in 1999 as a safer alternative to older painkillers, with fewer gastrointestinal side effects. It was prescribed to more than 20 million Americans and generated $2.5 billion in annual revenue at its peak. The scientific consensus held that it was safe.

The toll:

• An estimated 88,000–140,000 Americans suffered heart attacks from taking Vioxx (The Lancet)
• An estimated 38,000–55,000 died (FDA investigator Dr. David Graham / The Lancet)
• FDA’s own Dr. David Graham testified before the U.S. Senate that Vioxx was the “single greatest drug safety catastrophe in the history of this country or the history of the world”
• Merck scientists had discussed potential cardiovascular risks as early as 1996 — three years before FDA approval
• Merck ghostwrote 20 scientific articles and published them under the names of respected scientists
• When an FDA safety officer raised concerns, he was called “junk scientist” by senior FDA officials and faced threats from supervisors
• Merck settled all litigation for $4.85 billion — a fraction of Vioxx’s total revenues

The system designed to protect the public — peer review, FDA oversight, clinical trials — was compromised at every level by commercial interests, institutional deference, and the suppression of internal dissent. The consensus was manufactured.

The Opioid Crisis (1996–Present)

The consensus: Prescription opioids carry minimal addiction risk — “much less than 1%.”

The opioid epidemic — the deadliest drug crisis in American history — was built on a collective hallucination that originated inside the medical establishment itself. The foundational claim was drawn from a five-sentence letter published in the New England Journal of Medicine in 1980, which observed low addiction rates among hospitalized patients receiving opioids under medical supervision. Purdue Pharma and others transformed this narrow, uncontrolled observation into a universal marketing claim: that the risk of opioid addiction was “much less than 1%.”

The toll:

• 806,000 Americans dead from opioid overdoses between 1999 and 2023 (CDC)
• Drug overdose deaths increased 520% from 1999 to 2023
• Opioids are a factor in 76% of all overdose deaths — killing more than 217 Americans per day
• The original NEJM letter was cited 608 times between 1980 and 2017–72% of citations used it as evidence that addiction was rare, and over 80% failed to mention that the original study applied only to hospitalized patients
• Purdue Pharma’s OxyContin sales grew from $48 million in 1996 to over $1.5 billion by 2002
• Purdue funded more than 20,000 pain-related educational programs between 1996 and 2002
• The letter’s own lead author later said: “I’m essentially mortified that that letter to the editor was used as an excuse to do what these drug companies did”
• By the time addicted patients were cut off from prescriptions, many had already shifted to heroin and illicitly manufactured fentanyl — the third and deadliest wave of the crisis

The opioid crisis was not caused by patients behaving irrationally. It was caused by a medical consensus — reinforced by pharmaceutical marketing, regulatory capture, and institutional groupthink — that told doctors and patients alike that a profoundly addictive class of drugs was safe. The five-sentence letter that launched the epidemic was never a study. It was a collective hallucination dressed as evidence.

AZT: The Dosing Debacle (1987–1996)

The consensus: High-dose AZT monotherapy is the standard of care for HIV/AIDS.

When AZT (zidovudine) became the first FDA-approved treatment for AIDS in 1987, it was genuinely lifesaving for many patients — the first drug to show any ability to suppress HIV replication. But the medical consensus that formed around it became dangerously rigid.

What went wrong:

• AZT was initially prescribed at 1,500 mg/day — doses later found to be far more toxic than necessary
• At therapeutic doses, AZT caused severe bone marrow depletion, anemia, and immune suppression — the very symptoms it was supposed to treat
• Patients frequently required blood transfusions to survive the treatment
• The drug was approved after a trial halted at just 16 weeks — far too short to assess long-term effects or the development of viral resistance
• Doctors began prescribing AZT to healthy, asymptomatic people who merely tested positive for HIV antibodies — expanding the drug’s toxic reach far beyond its evidence base
• It was later discovered that one-third the original dose was equally effective and far less toxic — meaning thousands of patients had been subjected to needlessly dangerous treatment
• The medical establishment was slow to embrace alternative treatments and combination approaches; activists (notably ACT UP) had to fight for access to experimental drugs that many doctors dismissed
• It was not until 1996, nearly a decade later, that HAART (highly active antiretroviral therapy) — using low-dose AZT in combination with other drugs — transformed HIV from a death sentence into a manageable condition

The lesson is not that AZT was bad — it was a critical first step. The lesson is that the consensus around AZT — the insistence on high-dose monotherapy as the sole standard, the resistance to alternatives, the expansion to populations beyond the evidence — caused unnecessary suffering because questioning the prevailing approach was treated as heresy rather than science.

Leaded Gasoline (1923–1996)

The consensus: Tetraethyl lead is a safe, effective fuel additive. Government studies confirm it.

In 1921, a General Motors chemical engineer named Thomas Midgley Jr. discovered that adding tetraethyl lead (TEL) to gasoline eliminated engine knock. Lead was known to be a potent neurotoxin since antiquity — the Romans documented its capacity to cause madness and death. Midgley knew this too: he contracted lead poisoning himself. But GM, Standard Oil, and the newly formed Ethyl Corporation marketed the product aggressively, never using the word “lead” in advertising. When workers at TEL plants began dying — five dead, 35 hospitalized with hallucinations in a single New Jersey plant in 1924 — Midgley held a press conference in which he poured leaded gasoline over his hands and inhaled its fumes to “prove” it was safe. A government investigation funded by General Motors concluded the fuel was harmless. Leaded gasoline became the global standard for seven decades.

The toll:

• An estimated 100+ million premature deaths worldwide attributed to Midgley’s two inventions (leaded gasoline and CFCs), making him arguably the individual responsible for the most accidental deaths in human history
• 170+ million Americans alive today were exposed to dangerous lead levels in early childhood (PNAS, 2022)
• A collective 824 million IQ points stolen from the U.S. population — an average of 2.6 points per person; those born in the 1960s–70s lost up to 6–7 IQ points (Duke/Florida State, 2022)
• An estimated 151 million cases of psychiatric disorder over 75 years linked to childhood lead exposure in the U.S. alone (Duke, 2024)
• Economic losses estimated at $165–319 billion in lost wages in the U.S. alone
• Environmental historian J.R. McNeill called Midgley the person who “had more adverse impact on the atmosphere than any other single organism in Earth’s history”
• Leaded gasoline was not fully banned in the United States until 1996 — and the last country to sell it, Algeria, continued until 2021

The scientific consensus that leaded gasoline was safe persisted for over 50 years, defended by industry-funded research, regulatory capture, and institutional inertia — while every generation of children born during that period suffered irreversible neurological damage.

DES — Diethylstilbestrol (1940–1971)

The consensus: A safe synthetic estrogen that prevents miscarriage.

DES was prescribed to pregnant women for over three decades to prevent miscarriage and pregnancy complications. It was given widely, without rigorous testing, under the assumption that supplemental estrogen was beneficial. A study as early as 1953 showed DES was ineffective at preventing miscarriage. Doctors continued prescribing it for nearly two more decades.

The toll:

• An estimated 5–10 million Americans exposed — pregnant women and their children — between 1940 and 1971 (NCI)
• Globally, DES was prescribed to an estimated 2–10 million pregnant women worldwide
• “DES daughters” (women exposed in utero) face a 40-fold increased risk of clear cell adenocarcinoma of the vagina and cervix — a cancer virtually nonexistent in young women before DES
• DES daughters have approximately twice the risk of breast cancer after age 40
• Documented elevated risks of: infertility (33% vs. 16%), spontaneous abortion (50% vs. 39%), preterm delivery (53% vs. 18%), ectopic pregnancy (15% vs. 3%), stillbirth (9% vs. 3%), and early menopause (NCI, 2011)
• DES is a proven transplacental carcinogen — the first substance shown to cross the placenta and cause cancer in offspring
• Effects are multigenerational: DES granddaughters are now being studied for inherited epigenetic damage
• The drug was not banned in the U.S. until 1971 — and continued to be prescribed in Europe until 1978, and in Spain as late as 1983
• DES was also prescribed to prepubescent girls to stop growth, and was used to chemically castrate gay men — including Alan Turing, who died shortly after forced DES treatment

Trans Fats (1950s–2018)

The consensus: Partially hydrogenated vegetable oils are a heart-healthy alternative to butter and saturated fat.

In one of the most consequential dietary recommendations in modern history, the scientific establishment actively promoted artificial trans fats as a healthier alternative to animal fats. The U.S. Surgeon General in 1988, the Institute of Medicine in 1989, and the FDA all concluded that partially hydrogenated oils were comparable to or safer than saturated fats. Trans fats became ubiquitous in processed food, fast food, baked goods, and margarine — promoted as the “smart” choice for heart health.

The toll:

• Trans fat consumption kills an estimated 500,000 people per year globally from coronary heart disease (WHO/PAHO)
• In the U.S. alone, artificial trans fats were estimated in 2006 to cause 1 in 5 heart attacks — up to 250,000 heart attacks and 50,000 deaths annually (PMC)
• High trans fat intake increases risk of death from any cause by 34%, coronary heart disease deaths by 28%, and coronary heart disease incidence by 21% (WHO)
• Trans fats have no known health benefits — yet for decades, they were promoted as healthier than the natural fats they replaced
• Denmark banned trans fats in 2003 and saw immediate declines in heart disease deaths — but the FDA did not finalize its ban until 2018, decades after the evidence was clear
• As of 2023, 5 billion people worldwide remain unprotected by trans fat regulations (WHO)
• The original scientific endorsement of trans fats was itself a correction of an earlier consensus — the recommendation to reduce saturated fat intake — demonstrating how one collective hallucination can beget another

The Tuskegee Syphilis Study (1932–1972)

The consensus: A legitimate, ethically acceptable public health study.

The U.S. Public Health Service enrolled 399 Black sharecroppers with syphilis and 201 without it in Macon County, Alabama, telling them they were receiving free treatment for “bad blood.” They received no treatment. When penicillin became the standard cure for syphilis in the 1940s, it was deliberately withheld. The study was published in medical journals for nearly 40 years. Fifteen papers were published without a single letter of protest from the medical community. A 1969 CDC review panel — with no Black members and no ethicists — recommended the study continue.

The toll:

• 28 men died directly from syphilis; 100+ died from syphilis-related complications
• 40 wives were infected with syphilis; 19 children were born with congenital syphilis
• The study ran for 40 years — described as “the longest non-therapeutic experiment on human beings in medical history”
• It was exposed not by the medical establishment, but by a journalist (Jean Heller of the Associated Press) in 1972
• The lasting damage to Black Americans’ trust in medicine is incalculable — and contributed directly to lower participation in clinical trials, organ donation, and vaccination for decades
• The study was not secret: it was published openly. The consensus was not that it was hidden — the consensus was that it was acceptable

The Radium Girls (1917–1930s)

The consensus: Radium is harmless — even beneficial.

During World War I, young women were hired to paint watch dials with radium-based luminous paint. They were instructed to “lip-point” their brushes — licking them to a fine tip — and were told the paint was completely safe. Some painted their nails and teeth with it for fun. The company knew radium was dangerous. It hired doctors who provided false diagnoses. It suppressed internal reports. It stole the radium-riddled bones of dead workers during autopsies to destroy evidence.

The toll:

• An estimated 4,000 women were hired across U.S. and Canadian factories
• By 1927, more than 50 had died directly from radium poisoning; likely thousands more died over subsequent decades from cancers and bone diseases
• Workers were exposed to radiation levels up to 20,000 times the maximum permissible body burden later established
• Victims suffered from “radium jaw” — their jawbones literally disintegrating — along with bone cancers, anemia, and spontaneous fractures
• Marie Curie herself stated there was no cure once radium entered the body
• The legal battles of the Radium Girls led to the establishment of OSHA and foundational worker safety regulations — but only after the women were dying and dead

The Cumulative Toll

Add these to the cases documented earlier in this section, and a staggering picture emerges:

• Thalidomide: 10,000–20,000 birth defects, 40% mortality
• Lobotomy: 40,000–50,000+ procedures in the U.S.; untold thousands left permanently incapacitated; Nobel Prize never rescinded
• Vioxx: 88,000–140,000 heart attacks; 38,000–55,000 dead
• Opioid crisis: 806,000 dead from opioid overdoses (1999–2023, CDC)
• AZT high-dose monotherapy: Years of unnecessary toxicity before dose reduction
• Leaded gasoline: 170+ million Americans neurologically damaged; 824 million IQ points lost; estimated 100+ million premature deaths globally
• DES: 5–10 million exposed; multigenerational cancer and reproductive damage
• Trans fats: 500,000 deaths per year globally, for decades, while promoted as “heart-healthy”
• Tuskegee: 128+ dead; 40 wives infected; 19 children born with congenital syphilis; 40 years of open publication without protest
• Radium Girls: Thousands poisoned; companies destroyed evidence

These are not fringe incidents. They are the output of the most prestigious institutions in human civilization — the FDA, the Nobel Committee, the New England Journal of Medicine, the U.S. Public Health Service, the American Medical Association — operating exactly as designed. The hallucination was not a bug. It was a feature of human collective cognition applied to science under the distorting pressures of profit, prestige, and groupthink.

The Pattern

Every case follows the same architecture of failure:

• A claim gains institutional authority — through a regulatory approval, a journal publication, a Nobel Prize, a marketing campaign
• Dissent is suppressed — the whistleblower is fired, the skeptic is marginalized, the contrarian data is buried
• Commercial or reputational incentives reinforce the consensus — careers, profits, and institutional prestige become entangled with the claim’s survival
• The human cost accumulates silently — because the very system designed to detect harm is the system that produced it
• Correction comes late, from outside — from a lone FDA reviewer, an independent clinician, an activist movement, or sheer statistical weight of the dead

This is not an argument against science. It is an argument against the assumption that science is self-correcting on any timeline that matters to the people it harms. The peer review process, the regulatory apparatus, the expert consensus — these are human institutions, staffed by human minds subject to every cognitive bias Kahneman and Tversky cataloged. When they fail collectively, they fail catastrophically — and the phrase “trust the science” becomes the hallucination’s most effective camouflage.

“The Experts Have It Under Control”: When Economic Consensus Destroyed Lives

The pattern of collective hallucination is not confined to laboratories and medical journals. It extends — with even greater scale and arguably greater human cost — into the realm of economic policy. Here, the hallucination typically takes the form of a consensus that a particular financial structure, deregulatory framework, or austerity package is “sound” or “necessary” — a consensus maintained by the same architecture of institutional authority, suppressed dissent, and perverse incentives that drives scientific consensus failures. The difference is that when economic consensus fails, it doesn’t just harm patients or workers in a single industry. It reshapes entire societies.

The Great Depression (1929–1939)

The consensus: Markets are self-correcting. The Federal Reserve’s monetary policy is sound. Prosperity is permanent.

In the 1920s, the prevailing economic orthodoxy held that unregulated markets would naturally find equilibrium — that the stock market’s dizzying ascent reflected genuine value creation rather than speculative mania fueled by margin lending. When the bubble burst in October 1929, the Federal Reserve — the institution specifically designed to prevent such disasters — made it worse. Rather than expanding the money supply, the Fed contracted it. Rather than backstopping failing banks, it let them cascade into collapse. The consensus was that the economy would self-correct. It did not.

The toll:

• Unemployment surged from under 3 million in 1929 to 12.5 million by 1932 — a quarter of all American families had no employed wage earner
• Average family income fell 40%, from $2,300 to $1,500
• The suicide rate hit its all-time high of 22.1 per 100,000 in 1932 — a 22.8% surge, the largest increase in any four-year period ever recorded (CDC/AJPH)
• More than 200,000 vagrant children wandered the country after their families disintegrated
• In New York City, 110 people died of starvation by 1934
• Over 400,000 Mexican Americans — many U.S. citizens — were forcibly “repatriated” under economic pressure
• The Fed itself later admitted responsibility: Governor Ben Bernanke said in 2002, “Regarding the Great Depression, you’re right. We did it. We’re very sorry.”

The Great Depression was not an act of nature. It was an act of institutional consensus — a collective hallucination about the self-regulating wisdom of markets — maintained by the most credentialed economists of the era, and corrected only after a decade of human devastation.

The 2008 Financial Crisis and the Asymmetric Bailout

The consensus: The housing market is sound. Mortgage-backed securities are safe. Deregulation works.

The 2008 crisis is perhaps the purest modern example of collective economic hallucination — and of what happens when the correction is applied selectively, rescuing institutions while abandoning the individuals those institutions were supposed to serve.

For years, the consensus across Wall Street, the ratings agencies, the Federal Reserve, and both political parties held that the U.S. housing market was fundamentally sound, that mortgage-backed securities were safe investments (routinely rated AAA), and that financial deregulation — including the 1999 repeal of Glass-Steagall — was modernizing the economy. Alan Greenspan, the most influential central banker on Earth, testified to Congress that derivatives were making the financial system more stable. The SEC, the Fed, and the OCC actively blocked state regulators from investigating predatory lending. The consensus was not merely wrong — it was aggressively defended against the people who saw the crash coming.

Then the music stopped.

The toll on ordinary people:

• Nearly 10 million Americans lost their homes to foreclosure between 2006 and 2014 (Marketplace)
• 3.1 million foreclosure filings in 2008 alone — one in every 54 U.S. households received a notice (RealtyTrac)
• U.S. household wealth fell by $11 trillion — from $61.4 trillion to $50.4 trillion
• Median household wealth dropped 35%, from $106,591 to $68,839 between 2005 and 2011
• Home values fell 26% nationally — exceeding the drop during the Great Depression
• 8.7 million jobs vanished; unemployment hit 10%
• By 2012, 46.5 million Americans — one in seven — were living below the poverty line
• Foreclosure rates were directly linked to rising suicide rates, particularly among working-age adults (PMC)
• The Federal Reserve Bank of San Francisco calculated the crisis cost every American — man, woman, and child — $70,000 in lost lifetime income
• Less than a third of families who lost homes will ever return to homeownership (National Association of Realtors)

The toll on institutions: considerably gentler.

• The U.S. government committed over $626 billion in bailout funds through TARP, plus trillions in Federal Reserve emergency lending and quantitative easing
• AIG alone received $182 billion
• The government’s foreclosure prevention program, promised at $50 billion, spent only $20 billion — years after foreclosures peaked
• One bank — a small community bank called Abacus Federal Savings — was the only institution criminally prosecuted for conduct related to the crisis. It was acquitted.
• No senior executive at any major financial institution went to prison

The asymmetry is the hallucination made visible. The consensus held that banks were “too big to fail” — that their collapse would destroy the economy. The same consensus held no such view about the 10 million families whose homes were taken. Banks received immediate, unlimited liquidity. Homeowners received bureaucratic modification programs that reached a fraction of those in need, years too late. The people who built the bomb were rescued. The people standing under it were told to read the fine print on their mortgage.

The Greek Austerity Crisis (2010–2018)

The consensus: Severe fiscal austerity is the only path to economic recovery.

When Greece’s debt crisis exploded in 2010, the “Troika” — the European Commission, the European Central Bank, and the IMF — prescribed a treatment that was, by any humanitarian measure, worse than the disease. In exchange for bailout loans, Greece was required to cut spending by €28 billion in 2010–2011, with a further €13 billion in 2012–2014 — a total austerity package equivalent to almost 32% of Greece’s 2012 GDP.

The toll:

• The Greek economy contracted by over 25% — a depression comparable to the U.S. in the 1930s
• Unemployment hit 27%; youth unemployment exceeded 50%
• By 2015, nearly 20% of Greeks lacked funds to meet daily food expenses (OECD)
• 20,000 Greeks were made homeless in a single year (2011)
• Depression rates nearly quadrupled: from 3.3% in 2008 to 12.3% in 2013
• Suicide rates rose 35% between 2010 and 2012 (BMJ Open); austerity was directly linked to an estimated 511 additional suicides between 2011 and 2015
• Public health spending was cut by 25% between 2008 and 2012; mental health funding was slashed an additional 55% between 2011 and 2012
• By 2015, 14% of the population had lost health insurance coverage
• Infant mortality rose. Infectious disease morbidity increased. HIV infections among drug users surged as needle exchange programs were cut.
• After all the austerity, Greece’s debt-to-GDP ratio actually increased — from 120% in 2009 to 180% by 2015 — achieving the opposite of the policy’s stated goal
• The IMF’s own research later concluded that fiscal consolidation programs “do not reduce debt ratios, on average”

The Troika prescribed austerity as medicine. It functioned as poison. And when the patient deteriorated, the prescription was: more.

IMF Structural Adjustment Programs (1980s–Present)

The consensus: Privatization, deregulation, and spending cuts will produce growth in the developing world.

Beginning in the 1980s, the IMF and World Bank imposed “structural adjustment programs” (SAPs) on dozens of developing countries as a condition of receiving loans. The consensus — held by Western-trained economists and enforced by institutions governed primarily by wealthy nations — was that austerity, privatization, and trade liberalization would unlock growth. The programs were applied as a near-universal template, regardless of local conditions.

The toll:

• In sub-Saharan Africa and Latin America, public health systems collapsed, school enrollment plummeted, and poverty surged under SAP conditions
• The IMF’s own independent evaluation found its programs were associated with increased income inequality in 16 out of 22 countries examined
• In Argentina, IMF-supported austerity in the late 1990s contributed to a crisis that saw unemployment and poverty soar, culminating in the 2001 economic collapse and social unrest
• In Indonesia during the 1997 Asian Financial Crisis, IMF prescriptions worsened the downturn, leading to widespread poverty and the fall of the Suharto government
• In country after country, cuts to healthcare, education, and food subsidies fell hardest on women, children, and the poorest populations
• Jordan’s IMF-mandated subsidy cuts triggered violent protests — first in 1989, then again in 1996 — a pattern repeated across the developing world from Egypt to Zambia
• IMF-required austerity is “significantly associated with rising inequality, by increasing the income share to the top ten percent at the expense of the bottom 80 percent” (Boston University Global Development Policy Center, 2021)
• As of 2023, the vast majority of IMF loans to 38 countries (population: 1.1 billion) during the COVID pandemic still required austerity conditions (Human Rights Watch)

The structural adjustment consensus was not a theory tested and found wanting. It was an ideology imposed by creditor nations on debtor nations — a collective hallucination about the universality of market liberalization, maintained by institutions in which the people bearing the costs had the least representation.

The Cumulative Economic Toll

The pattern across these cases is unmistakable:

• The Great Depression: A decade of mass unemployment, starvation, and record suicides caused by institutional consensus about self-correcting markets — later admitted as error by the institution responsible
• 2008 Financial Crisis: $11 trillion in destroyed household wealth, 10 million homes lost, 8.7 million jobs eliminated — while the institutions that caused the crisis received unlimited government support and zero criminal prosecution
• Greek Austerity: A 25% economic contraction, 35% rise in suicides, quadrupled depression rates — to implement a policy the IMF later admitted doesn’t work
• IMF Structural Adjustment: Decades of increased poverty and inequality across the developing world, imposed by institutions governed by the countries least affected by their policies

In every case, the hallucination follows a familiar structure: economic orthodoxy becomes consensus; consensus becomes policy; policy is imposed on populations with no meaningful voice in its design; dissent is dismissed as economically illiterate; the human cost accumulates; correction comes only after enormous damage — and sometimes not at all.

What makes economic hallucination uniquely dangerous is its asymmetric correction. When the consensus fails in medicine, at least the harm and the correction eventually fall on the same system. When economic consensus fails, the architects of the failure are routinely insulated from its consequences. Banks are bailed out; homeowners are foreclosed on. Creditor nations design austerity; debtor nations bury their dead. The executives who packaged toxic mortgages kept their bonuses. The families who signed them lost everything.

This is not a failure of markets. It is a failure of the human institutions that govern markets — institutions subject to the same groupthink, the same perverse incentives, the same suppression of dissent, and the same catastrophic pattern of collective hallucination documented throughout this essay. The difference is that in economics, the hallucination doesn’t just kill patients. It reshapes the distribution of wealth and power for generations.

“Trust the Experts”: When Economic Consensus Destroyed Millions of Lives

The pattern of institutional hallucination is not confined to medicine and science. The same architecture of failure — confident consensus, suppressed dissent, misaligned incentives, catastrophic human toll, belated correction — operates with equal lethality in economic policy. And because economic policy scales to entire populations, the body counts are often larger.

The 2008 Financial Crisis: Bailouts for Banks, Ruin for Everyone Else

The consensus: Deregulated financial markets are self-correcting. Housing prices only go up. Mortgage-backed securities are safe.

For a decade, the entire institutional apparatus of American economic governance — the Federal Reserve, the SEC, the ratings agencies, the Treasury Department, both political parties, and the nation’s most prestigious economics departments — operated under a shared hallucination: that the financial system had effectively engineered risk out of the housing market. Subprime mortgages were sliced, repackaged, and rated AAA by the same agencies paid by the banks issuing them. Alan Greenspan’s Fed kept interest rates low and championed deregulation. The few who raised alarms — Brooksley Born at the CFTC, economist Robert Shiller, analyst Meredith Whitney — were dismissed or marginalized.

Then the hallucination collapsed.

The toll:

• Nearly 10 million Americans lost their homes to foreclosure between 2006 and 2014 (Marketplace)
• $11 trillion in household wealth evaporated — from $61.4 trillion to $50.4 trillion (Federal Reserve)
• Median household wealth fell 35% between 2005 and 2011 — from $106,591 to $68,839
• 8.7 million jobs lost; unemployment soared to 10%
• By 2012, 46.5 million Americans — 1 in 7 — were living below the poverty line
• Housing prices dropped 27% nationally; in Arizona, California, Florida, and Nevada they fell over 50%
• By November 2008, an estimated 12 million homeowners were “underwater” — owing more than their homes were worth
• The crisis cost every man, woman, and child in the country an estimated $70,000 each in lost income over the following decade (Federal Reserve Bank of San Francisco)
• Less than one-third of families who lost homes to foreclosure will ever return to homeownership (National Association of Realtors)

The asymmetry:

The government mobilized $700 billion in TARP bailout funds for Wall Street and ultimately committed trillions in emergency lending and guarantees to stabilize banks. AIG alone received $182 billion. The architects of the crisis — the banks that issued predatory loans, the ratings agencies that blessed them, the regulators who looked away — were rescued. Many paid themselves bonuses with bailout money.

The homeowners they had victimized received almost nothing. The Obama administration’s mortgage modification programs promised $75 billion; they delivered less than half, and most of it arrived years after families had already lost their homes. The government spent billions on foreclosure prevention programs — ProPublica described it as a firefighter arriving late and spraying the smoldering ruins. Only one bank — Abacus Federal Savings, a small community bank — was ever prosecuted. Not a single executive at the major institutions that engineered the crisis went to prison.

The human cost beyond the ledger:

• An estimated 10,000+ excess suicides in the U.S., Canada, and Europe were attributable to the Great Recession (Oxford/Lancet)
• A 54-country study found 4,884 excess suicides in 2009 alone compared to pre-crisis trends (BMJ, 2013)
• In the U.S., approximately 4,750 excess suicides occurred during 2008–2010, concentrated among working-age men (The Lancet)
• In the UK alone, an estimated 1,000 excess suicides occurred between 2008 and 2010, representing the tip of an iceberg of roughly 30,000–40,000 additional suicide attempts
• Foreclosure itself was directly linked to increased suicide rates in middle-aged adults — the demographic most likely to hold a mortgage (PMC)

The lesson is stark: when the consensus failed, institutions protected themselves. The people who believed the consensus — who bought homes they were told were safe investments, with mortgages they were told they could afford, in an economy they were told was sound — bore the cost. The hallucination flowed downhill.

Russian “Shock Therapy” (1991–1998): When Economists Experimented on 150 Million People

The consensus: Rapid privatization and immediate market liberalization will smoothly transform the Soviet economy into a prosperous capitalist democracy.

In one of the most consequential economic experiments in modern history, Western economists — led by Harvard’s Jeffrey Sachs and backed by the IMF and World Bank — prescribed “shock therapy” for Russia’s post-Soviet transition. The plan: simultaneous price liberalization, mass privatization, and fiscal austerity. A “big bang” that would irreversibly shift the economy to free markets. The architects promised short-term pain and long-term prosperity.

The toll:

• An estimated 3–7 million excess deaths in Russia and the former Soviet bloc during the 1990s, depending on methodology (UNICEF estimated 3 million; The Lancet estimated 7 million across Eastern Europe)
• Russia’s male life expectancy plummeted from 64 years in 1990 to 57 in 1994 — a decline of 7 years virtually overnight, giving Russian men shorter life expectancies than men in Egypt, Indonesia, or Paraguay
• Russian GDP fell by roughly 50% between 1992 and 1998 — a peacetime collapse worse than the Great Depression
• Poverty skyrocketed from 2% of the population (1987–88) to 50% (1993–95) — in under five years
• 70,000 factories closed. Investment collapsed 92%. Agricultural production fell 63%
• Countries that implemented rapid mass privatization saw an average 13% surge in working-age male deaths (The Lancet, 2009)
• Alcoholism rose 60%; deaths from disease doubled
• A handful of connected insiders — the future oligarchs — acquired vast industrial assets for pennies, while pensioners lost their savings to hyperinflation exceeding 2,500%
• Russians rechristened the process prikhvatizatsia — “grabitization”

Countries that pursued gradual reform — Poland, the Czech Republic, Slovenia — experienced dramatically better outcomes: a 10% decline in male mortality and only 2% increases in unemployment. The evidence was available at the time. Gradualist economists warned that destroying an economy doesn’t automatically create a market. They were overruled by a confident consensus of Western advisors who, as former Soviet Premier Gorbachev later stated, did “irreparable harm.”

IMF Structural Adjustment: Exporting Austerity to the World’s Poorest

The consensus: Fiscal discipline, privatization, and market liberalization will foster growth in developing economies.

For decades, the International Monetary Fund has conditioned emergency lending on “structural adjustment programs” (SAPs) requiring borrower nations to cut social spending, privatize state assets, deregulate markets, and implement austerity. The theory was that short-term pain would yield long-term growth. The results have been extensively documented:

• IMF structural reforms are significantly associated with rising poverty and increased inequality, concentrating income to the top 10% at the expense of the bottom 80% (Boston University Global Development Policy Center, 2021)
• WHO has stated that “IMF economic reform programs are associated with significantly worsened tuberculosis incidence, prevalence, and mortality rates” in post-communist countries
• The World Bank itself has acknowledged that SAPs “have failed the poor, with a special burden falling on women and children”
• In country after country — Zambia, Kenya, Ghana, Argentina, Greece — austerity triggered mass unemployment, collapsed social services, and political unrest
• Greek suicides rose more than 60% after the 2008 crisis and subsequent IMF-backed austerity (The Lancet)
• Anti-austerity protests in Kenya in 2024 resulted in at least 39 deaths, hundreds of injuries, and 627 arrests

The pattern is identical to the medical cases: a confident institutional consensus, imposed over objections, generating immense human suffering, defended for decades before partial correction, with the cost borne overwhelmingly by the most vulnerable.

The Economic Pattern

These are not merely policy disagreements. They are hallucinations — shared delusions adopted and enforced by the most powerful institutions on earth, resistant to contradictory evidence, defended by the careers and reputations of their advocates, and measured in human bodies.

In every case:

• The experts agreed. The Fed, the IMF, Harvard, the Treasury, the ratings agencies, the economics profession
• Dissenters were marginalized. Brooksley Born was silenced. Gradualist economists were overruled. Countries that resisted IMF conditions were punished
• The incentives pointed the wrong direction. Banks profited from risk. Ratings agencies were paid by the companies they rated. IMF economists’ careers depended on the programs they designed
• The cost was socialized downward. Banks were bailed out; homeowners were foreclosed on. Oligarchs got factories; pensioners got hyperinflation. Bondholders were protected; hospitals were defunded
• The architects were never held accountable. No Wall Street executive was imprisoned. No IMF economist was fired for the social wreckage of structural adjustment. No one rescinded Jeffrey Sachs’s honorary degrees

When the subject is economics, the phrase changes — from “trust the science” to “trust the markets” or “trust the experts” — but the underlying cognitive architecture is the same. A consensus hardens into dogma. Dogma becomes policy. Policy has a body count. And by the time the correction arrives, the people who needed it are already gone.

Why Human Hallucination Is Structurally Irreparable — Without Help

AI hallucination and human hallucination are not equivalent problems. They differ in almost every dimension that matters — and the comparison consistently favors the machine.

AI hallucinations are detectable. They can be checked against sources, flagged by other systems, and corrected in milliseconds. Human collective hallucinations actively resist detection. They punish the people who try to point them out. Galileo was placed under house arrest. Ignaz Semmelweis was committed to an asylum for suggesting that doctors should wash their hands. Daniel Ellsberg was prosecuted for revealing the truth about Vietnam. The social immune system of a collective hallucination doesn’t merely fail to correct the error — it attacks the correction itself.

AI hallucinations are non-contagious. A chatbot that invents a fake legal citation doesn’t cause other chatbots to spontaneously adopt the same error. Human hallucinations spread — through media, through social pressure, through institutional authority, through the basic primate need to belong. One person’s false belief is unfortunate. A million people’s false belief is a movement. A nation’s false belief becomes policy — and policy has a body count.

AI hallucinations carry no emotional charge. They aren’t fused with identity, tribe, or ego. Correcting an AI’s error doesn’t trigger a defensive reaction that makes the error stronger. Correcting a human’s deeply held collective belief — especially one tied to political identity, religious faith, or national narrative — activates what psychologists call the backfire effect: the well-documented phenomenon where presenting contradictory evidence makes people more committed to their original position, not less.

AI hallucinations are improvable on a curve. Every generation of models hallucinates less. Retrieval-augmented generation, chain-of-thought reasoning, citation verification — these are engineering problems with engineering solutions, and the trajectory is unmistakably toward improvement. Human collective hallucination shows no clear historical trend of reduction. We have better information access than any civilization in history, and we still managed to collectively hallucinate our way through a pandemic in real time, with the data freely available on every screen in every pocket.

The problem isn’t access to information. The problem is the machinery processing it.

The Corrective Machine

Here is the argument that almost no one is making, because it cuts against the dominant narrative of AI fear: properly deployed AI systems may represent one of the most powerful correctives to collective human cognitive failure that our species has ever produced.

Consider what AI, at its core, does differently from a human mind operating within a social group:

• It evaluates claims without the gravitational pull of tribal loyalty
• It surfaces inconvenient data without career anxiety
• It holds contradictory evidence in view without the psychic distress that causes humans to unconsciously suppress whichever data set threatens their worldview
• It doesn’t need to belong, doesn’t need to be liked, and has no identity to defend

When Tetlock found that expert forecasters performed barely above chance, the critical variable separating the rare “superforecasters” from the rest wasn’t intelligence or credentials — it was their ability to resist the cognitive biases that captured everyone else. AI systems can be built with that resistance as a structural feature, not a rare individual talent.

This doesn’t mean AI is infallible. It is not. It reflects biases in its training data, it can be manipulated, and its errors can be confidently stated. These are genuine concerns that merit serious engineering attention. But they are tractable concerns — problems of design, alignment, and iteration that demonstrably improve with each successive generation.

The problems of human collective hallucination are not tractable in the same way. We’ve known about confirmation bias for over fifty years, and it hasn’t diminished. We’ve had access to the sum of human knowledge via the internet for a generation, and conspiracy theories have proliferated, not receded. Kahneman himself observed that even being aware of cognitive biases provides minimal protection against falling prey to them. Education helps at the margins, but it doesn’t inoculate — as the Holocaust demonstrated when one of Europe’s most educated populations committed one of history’s most systematic atrocities.

No software update is coming for the human brain. But we have, for the first time in history, built a tool that can serve as a cognitive counterweight at scale.

What a Corrective AI Actually Looks Like

This isn’t about replacing human decision-making. It’s about auditing it — inserting a dispassionate analytical layer into processes where group dynamics, emotional reasoning, and institutional pressure have historically overwhelmed evidence.

Imagine:

• An AI system embedded in intelligence analysis that flags when a consensus is forming faster than the evidence warrants — the precise dynamic that preceded the Iraq War’s phantom weapons of mass destruction.
• Public health policy stress-tested by AI that identifies when a risk assessment is failing to account for age-stratified data and is instead treating an entire population as uniformly endangered — the dynamic that led to school closures and economic shutdowns that devastated the young and the poor while the threat was concentrated overwhelmingly among the elderly.
• Economic policy subjected to AI review that identifies when assumptions about a population’s worthiness of aid are being driven by racial or ethnic bias rather than data — the dynamic that allowed a million Irish to starve while grain ships left their ports.

These are not science fiction scenarios. The building blocks exist today. Research published in Electronics in 2025 documented how AI-driven decision-support systems, incorporating predictive modeling and real-time analytics, can systematically reduce the heuristic-driven errors that Kahneman and Tversky identified over fifty years ago. The AI doesn’t make the decision. It makes the decision-makers aware of their own cognitive blind spots at the moment those blind spots are most dangerous.

What if the logic machine’s greatest value isn’t replacing human thought but disciplining it — flagging the moments when consensus is running ahead of evidence, when policy is being driven by narrative rather than data, when an entire society is about to sleepwalk into a catastrophe because the shared story feels too good — or too frightening — to question?

AI as Guardrail: Concrete Mechanisms for Reducing Human Self-Harm

This isn’t speculative. The tools already exist — some deployed, some in development — that demonstrate AI’s capacity to function as a structural check on the specific cognitive failures that produce collective hallucination. Here are the mechanisms, and the evidence behind them.

1. The Automated Devil’s Advocate

One of Janis’s original prescriptions for preventing groupthink was the appointment of a devil’s advocate — someone tasked with challenging the group’s emerging consensus. The problem, as decades of organizational research have shown, is that human devil’s advocates don’t work very well. They pull their punches because they don’t actually believe the opposing position. They face social penalties — reduced belonging, damaged relationships, threats to their standing. And they apply skepticism unevenly, influenced by their own biases.

AI solves every one of these problems.

A 2024 study published at the ACM International Conference on Intelligent User Interfaces (IUI ’24) by researchers at Purdue University tested LLM-powered devil’s advocates in group decision-making scenarios. The findings were striking:

• LLM-powered devil’s advocates enhanced decision quality and creativity while avoiding the social costs that cripple human advocates
• The AI could practice Socratic questioning — asking open-ended, critical questions rather than simply declaring opposing positions — which research shows is the most effective form of devil’s advocacy
• Groups using AI devil’s advocates showed no substantial increase in perceived workload, suggesting the friction was productive, not burdensome
• Critically, the AI advocate could present genuine opposing viewpoints without compromising the psychological safety of any group member — something no human advocate can reliably do

A separate 2025 study from IUI presented an AI-mediated devil’s advocate system specifically designed to amplify minority voices in power-imbalanced groups — the exact dynamic where groupthink is most dangerous and most difficult to counter. The system used anonymous, text-based communication to focus attention on the substance of ideas rather than participants’ status, directly addressing the power differentials that suppress dissent.

The research on dialectical inquiry and devil’s advocacy is unambiguous: groups that use structured adversarial techniques produce decisions rated 33–34% higher in quality than groups relying on consensus-building alone (Schweiger et al., longitudinal study of 120 managers). The problem has always been that human organizations can’t sustain the practice because of social costs. AI removes those costs entirely.

2. Early Warning Systems for Mass Atrocities

The most devastating consequences of collective hallucination — genocide, mass killings, ethnic cleansing — don’t erupt spontaneously. They are planned, preceded by detectable warning signs, and theoretically preventable. The challenge has always been detection at scale and speed.

AI-powered early warning systems are now operational:

• The Early Warning Project, a joint initiative of the U.S. Holocaust Memorial Museum’s Simon-Skjodt Center for the Prevention of Genocide and Dartmouth College, uses statistical forecasting and machine learning to rank more than 160 countries by their likelihood of experiencing mass killings. The system combines quantitative risk models with crowd-sourced predictions and expert assessment to identify threats before they escalate.
• The Sentinel Project’s Sentry AI uses a multi-sensor network with artificial intelligence to validate information from multiple sources in real time, allowing stakeholders to detect, identify, and predict threats in conflict zones. Its localized approach focuses on subnational regions — states, provinces, cities, even villages — where atrocities are actually carried out.
• The Violence & Risk Index (VRI) uses dynamic AI forecasting to predict not just whether violence will occur, but its intensity and frequency relative to case-specific baselines — a far more actionable signal than binary risk flags.
• University of Notre Dame researchers are developing an AI early warning system specifically designed to monitor how manipulated online content — altered photos, misleading memes, edited videos — can escalate into violent conflict, societal instability, and interference with democratic elections. Their research on the 2019 Indonesian election demonstrated how online disinformation campaigns produce real-world violence.

These systems exist because the pattern is clear: genocide and mass atrocities are always preceded by escalating rhetoric, dehumanizing narratives, and institutional alignment around a shared false belief. AI can monitor these precursors across languages, platforms, and geographies at a scale no human monitoring system can match — and flag them before the killing starts.

3. Structural Debiasing of High-Stakes Decisions

Beyond crisis prevention, AI offers something more fundamental: the capacity to systematically debias the routine decisions that, when distorted by collective cognitive failure, compound into catastrophe.

The evidence is building rapidly:

• A 2024 board search experiment published by the California Management Review found that AI debiased by design produced candidate slates with significantly higher diversity across gender, race, ethnicity, and sexual orientation compared to traditional human-led searches. The AI outperformed humans not by imposing quotas but by discovering talent pools that human cognitive shortcuts systematically overlooked.
• Research by Glickman and Sharot (2024) provided the first experimental evidence that unbiased AI recommendations can actively improve human judgment and mitigate pre-existing biases — not just avoid reinforcing them. Conversely, they showed that biased AI amplifies human biases, underscoring the importance of getting the design right.
• A systematic review published in MDPI Electronics (2025) covering 156 empirical studies documented how AI-driven decision-support systems incorporating predictive modeling and real-time analytics can systematically reduce the heuristic-driven errors — anchoring, availability bias, confirmation bias — that Kahneman and Tversky identified fifty years ago.
• U.S. workplaces spend an estimated $8 billion annually on unconscious bias training that has been shown to be largely ineffective at changing behavior (Dobbin & Kalev, Harvard). AI-driven debiasing, by contrast, operates at the point of decision rather than relying on individuals to override their own cognitive architecture — which Kahneman demonstrated they cannot reliably do.

4. The Compound Effect: When AI Improves, Humans Improve

Perhaps the most important finding in recent research is what scholars are calling the “compound human-AI bias” effect — the discovery that the interaction between human cognition and AI systems doesn’t just go one direction. When AI systems are designed to be unbiased, they don’t merely avoid amplifying human errors; they actively reduce them over time. Users exposed to accurate, well-calibrated AI recommendations begin making better independent judgments even without the AI present.

This is the opposite of the human collective hallucination dynamic, where exposure to a group’s false consensus makes individuals worse at independent judgment. AI, properly designed, doesn’t just resist groupthink — it builds cognitive immunity in the humans who work with it.

The implications are profound. Every historical catastrophe in the ledger above followed the same pattern: a false narrative gained social momentum, institutions reinforced it, dissenters were punished, and the collective hurtled toward disaster with increasing conviction and decreasing self-awareness. AI guardrails intervene at every stage of that cascade:

• At narrative formation: Flagging when a consensus is forming faster than evidence supports
• At institutional reinforcement: Surfacing contradictory data that organizational pressure would suppress
• At dissent suppression: Providing anonymous, socially costless channels for opposing views
• At policy execution: Stress-testing decisions against historical patterns of collective failure

No human institution has ever reliably performed all four functions. Most can’t perform even one. The machine can perform all four simultaneously, continuously, without fatigue, without career anxiety, and without the tribal allegiance that turns every human organization into a potential incubator of collective delusion.

Recalibrating Fear

None of this is to dismiss the importance of making AI systems more accurate. A tool that fabricates legal precedents or medical dosages is a tool that needs fixing. But the scale of concern is grotesquely miscalibrated.

A summary of the human hallucination ledger:

• The Crusades — The delusion: God wills holy war. The cost: 1–3 million dead.
• The Holocaust — The delusion: Jews are subhuman. The cost: 11 million murdered.
• The Irish Famine — The delusion: The poor deserve their fate. The cost: 1 million dead, 2 million displaced.
• The Iraq War — The delusion: WMDs exist. The cost: 461,000–940,000+ dead.
• Covid-19 Lockdowns — The delusion: Uniform risk justifies total shutdown. The cost: $17–35 trillion in damage, 1.6B children’s education disrupted.
• The Slave Trade — The delusion: Certain races are property. The cost: 12.5 million displaced.
• Tobacco Consensus — The delusion: Cigarettes are safe. The cost: 100 million dead (20th century).

We are treating AI hallucination as an existential crisis while treating human collective hallucination — which has, by conservative estimate, killed tens of millions of people, displaced hundreds of millions more, and perpetuated economic subjugation for billions across centuries — as an unfortunate but inevitable feature of the species. Background noise. The cost of doing business as a social animal.

That framing is itself a kind of collective hallucination.

Philip Tetlock showed us that expert judgment is barely better than chance. Kahneman showed us that our cognitive biases are structural, not correctable by mere awareness. Janis showed us that groups of brilliant people reliably make catastrophic decisions. The historical record shows us the price: millions dead, millions displaced, entire civilizations warped by shared delusions that persisted for decades or centuries.

We have now built, for the first time, a class of tools that processes information without tribal allegiance, evaluates evidence without ego protection, and improves with each iteration rather than repeating the same errors generation after generation. The question is not whether these tools are perfect. They are not. The question is whether they can help close the gap between the information available to human societies and the decisions those societies actually make — a gap that has, historically, been filled with graves.

The next time someone tells you they’re terrified of AI making things up, ask them how they feel about the fact that entire nations have made things up, believed those things with total conviction, and killed millions of people as a result — and that it keeps happening, right now, with full access to the facts.

Then ask them which hallucination problem they’d like to solve first — and which tool might actually help.

About the Author

Brian James Curry is a Kansas City–based AI, Machine Learning, and Data Science applied researcher and developer, operating at the intersection of data science, economic modeling, marketing science, and enterprise transformation. He is the founder of Vector1 Research, an independent lab focused on marketing economics, causal inference, cognitive AI architecture, and the economics of intelligent automation.

Brian has led data science and analytics initiatives across major enterprises including Koch Industries ($125B), Tractor Supply ($14B), Vail Resorts, Hallmark, Garmin, AT&T, and others. His work helps organizations move beyond surface-level analytics toward causal understanding — connecting marketing investment to outcomes, operational decisions to economic value, and AI adoption to competitive advantage.

His research contributions include Papilon (open-source marketing analytics & simulation), PyCausalSim (simulation-based MMM framework), Memory-Node Encapsulation (cognitive AI architecture for episodic memory), and The Autonomy Economy (macroeconomic framework for AI-driven transformation). He was the founder of the KC AI Lab, LLC in 2016, where he led educational initiatives and community projects in machine learning and artificial intelligence.

Brian grew up in a small farming town, started working in the crop fields at age 12, and spent time at Kansas State as a collegiate athlete. He has trained Brazilian Jiu Jitsu, enjoys making music and painting, and is the father of two wonderful humans. As an active musician and composer, he’s deeply interested in creativity, human–AI collaboration, and generative systems — bringing together his technical expertise with his artistic sensibilities to explore the intersection of intelligence, creativity, and automation.