WHO'S HALLUCINATING?

One Facebook Post, the Vaccine Wars, and the Search for What's True

Robert Curtis

Back-cover headline: Following the Evidence in an Age of Memes, Vaccines, and AI

Copyright

Copyright © 2026 Robert Curtis. All rights reserved.

No part of this book may be reproduced, distributed, or transmitted in any form or by any means without the prior written permission of the author, except brief quotations in reviews and other noncommercial uses permitted by copyright law.

This book is provided for general informational and educational purposes only and does not constitute medical, legal, or professional advice. Readers should consult qualified professionals regarding their individual circumstances. [Disclaimer subject to legal review.]

This book was developed with the assistance of AI tools. All analysis, conclusions, and editorial decisions are the author's own.

Draft manuscript — not for distribution. ISBN to be assigned. First edition, 2026.

Contents

Dedication

This book is dedicated to the memory of

Bob Geiger, Kim Brien, and Naureen Khan,

friends whose lives were cut short during the COVID-19 pandemic. Their absence is part of the reason these pages exist.

It is also dedicated to my sister, a nutritionist and diabetes educator in Florida, who was called into frontline care during the pandemic and witnessed firsthand what statistics, graphs, models, and political arguments can never fully capture — the patients, the families, the fear, the loss, and occasionally the hope.

This book is for all of them.

It is also for those who chose vaccination and those who did not. Not to tell them what to think. Not to tell them what to believe. But to encourage all of us to pause, to question, to listen, and to think a little more carefully before deciding that the other side has lost its mind.

And vice versa.

Foreword — This Is Not Really a Book About Vaccines

"Curiouser and curiouser!" — Alice's Adventures in Wonderland

Foreword — This Is Not Really a Book About Vaccines
Foreword — This Is Not Really a Book About Vaccines

This is not really a book about vaccines. Vaccines are the subject matter; COVID is the backdrop; politics, media, government, and even artificial intelligence wander through the story. But the real subject is truth — and how human beings decide what is true when certainty is unavailable.

Over the last several years I kept being drawn into arguments that looked unrelated on the surface — a Facebook post about vaccine injuries, a fight about election integrity, a quarrel about economics or climate — and finding the same pattern underneath. Every side believed it possessed reality. Every side believed the other had been deceived. Every side could point to cases where its opponents were wrong. And every side was remarkably blind to the chance that it might be wrong itself.

Then artificial intelligence handed the culture a useful word: hallucination. A machine generates an answer that sounds convincing, fits its context, and is not grounded in reality. Not a lie — a lie requires knowing the truth — but a plausible story mistaken for fact. The more I sat with the term, the more familiar it felt, because humans do this constantly. We take fragments and assemble them into narratives, fill the gaps, name the heroes and villains, and hold the result with far more confidence than the evidence allows.

What follows does not promise certainty. It promises to follow the evidence as far as it goes, to mark clearly where the ground is firm and where it isn't, and to resist the temptation to turn complexity into a slogan. Reality rarely fits on a bumper sticker. It is messy, sometimes contradictory, and it changes when new evidence arrives. But it has one quality that separates it from every ideology, institution, movement, and machine: reality keeps score.

The question is whether we are paying attention.

How to Read This Book

Before going further, six ideas appear throughout this book and are worth understanding from the start. They are not technical. They sit beneath nearly every disagreement that follows, and most arguments about vaccines are really arguments about one of these things.

First, a signal is not a conclusion. A signal is simply a pattern suggesting that something unusual may be happening and deserves a closer look. A smoke alarm is a signal. Sometimes there is a fire; sometimes there is burnt toast. The alarm tells us where to look, not what we will find. Much of the vaccine fight comes from treating an alarm as a verdict.

Second, correlation is not causation. Two things can rise and fall together without one causing the other. Every summer, ice-cream sales and drowning deaths both increase. Ice cream does not cause drowning; warm weather drives both. Showing that two things move together is the beginning of an investigation, not the end of one.

Third, relative risk and absolute risk are not the same thing. If a risk rises from one in a million to two in a million, it has doubled in relative terms while remaining tiny in absolute terms. "Doubled" sounds alarming; "one extra case per million" sounds negligible; both describe the same fact. Public arguments routinely exploit the gap between these two ways of describing the same number.

Fourth, many benefit figures are counterfactuals. Estimates of "lives saved" attempt to describe a world that never happened — the timeline in which no one was vaccinated. Such estimates can be careful and useful, but they are produced by models, not by counting bodies. An estimate and a measurement are different kinds of claims, and this book keeps them distinct.

Fifth, uncertainty is not ignorance. When scientists say they are uncertain, people often hear "they don't know." But "we know this with high confidence, this with moderate confidence, and this poorly" is not the same as knowing nothing. Honest uncertainty is a map of how firm the ground is in different places — and refusing to flatten it into false certainty is a feature of good evidence, not a weakness.

Sixth, a big signal can hide a smaller one — this is called masking. Safety systems work by scanning reports for events that show up more often than expected. But when one product suddenly generates a flood of reports, it can crowd out a quieter pattern for that same product, the way a loud noise drowns out a softer one in the same room. Statisticians have built methods to correct for this — to adjust the comparison so a large signal cannot bury a small one. This matters because the central dispute in the vaccine-safety fight is partly a technical argument about whether one such method should have been used sooner, and what it would have shown. The terms attached to these methods (you will see "MGPS" and "RGPS") are just names for different ways of doing that scan; what matters is the idea, not the acronyms.

With those distinctions in hand — signal versus conclusion, correlation versus causation, relative versus absolute, estimate versus measurement, uncertainty versus ignorance, and the way a big signal can mask a smaller one — the rest of the book becomes much harder to misread.

The Vaccine Debate, Explained — A Roadmap for an Adult Conversation
The Vaccine Debate, Explained — A Roadmap for an Adult Conversation

Origin Story — One Facebook Post

"Would you tell me, please, which way I ought to go from here?" — Alice's Adventures in Wonderland

This began with a single social-media post. It carried alarming claims, real documents, and a strong conclusion: that Senator Ron Johnson had shown the FDA caught vaccine stroke warnings "26 days in" and buried them, that a working detection method had been shut down, and that the media was ignoring a major scandal while the vaccine-injured were never warned.

A few years earlier I might have done what most people do — shared it if I already agreed, dismissed it if I didn't. Instead I asked one question: how much of this is actually true?

The answer was more complicated than expected. The first surprise was that much of the underlying scaffolding was real. The Senate hearing existed. The interim report — Failure to Warn, from the Permanent Subcommittee on Investigations, released with the May 21, 2025 hearing — existed. The internal FDA emails existed. The dispute over statistical methods existed. The FDA scientist and the alternative signal-detection method existed.

The second surprise was that the strongest claims in the post were not supported by the evidence it rested on. The existence of a safety signal is not proof of causation. A dispute over statistical methods is not proof of a cover-up. An allegation is not a conclusion. And the specific "26 days" framing was wrong: the rollout began in December 2020, and the twenty-six days actually referred to the gap between an FDA biostatistician's internal warning in March 2021 and a later demonstration of the masked signals — not twenty-six days into the campaign.

Each answer generated new questions, and at some point the subject quietly stopped being vaccines and became information itself: how anyone tells a signal from a conclusion, skepticism from cynicism, a fact from a story that merely feels true. That is the journey this book traces.

Prologue — The Meme Versus the Messy Truth

"Begin at the beginning and go on till you come to the end: then stop." — Alice's Adventures in Wonderland

Prologue — The Meme Versus the Messy Truth
Prologue — The Meme Versus the Messy Truth

Almost nobody argues with the real version of the vaccine problem. They argue with the cartoon. One side says the vaccines were safe and effective, full stop. The other says they were dangerous, useless, and part of a plot. Both fit on a meme. Neither survives under a microscope.

Vaccines are not magic water; they have benefits, risks, and limits. Diseases are not moral tests; they have age curves, comorbidity curves, and variant curves. Institutions are not gods; they can be right, wrong, cautious, slow, or opaque. And conspiracy communities are not wrong about every concern — but they often take one real question and inflate it into a universe of false certainty.

The adult conversation starts with a discipline: count both sides of the ledger. Count vaccine injuries and disease injuries. Count deaths prevented and deaths not prevented. Count what we know and what we cannot. Then ask a better question than "are vaccines good or bad?" — namely, good or bad for whom, at what age, with what health background, against which variant, with which vaccine, and compared with what alternative?

That question does not fit on a meme. That is why it matters.

Stage 1 — Vaccine Safety: Signals, Injuries, and What the Numbers Really Mean

"Take care of the sense, and the sounds will take care of themselves." — Alice's Adventures in Wonderland

Stage 1 — Vaccine Safety: Signals, Injuries, and What the Numbers Really Mean
Stage 1 — Vaccine Safety: Signals, Injuries, and What the Numbers Really Mean

Start with the distinction the whole public fight keeps tripping over, because it is a matter of definition, not opinion. A safety signal is not proof of harm. The systems that scan vaccine reporting databases — the FDA's standard MGPS method, and the masking-adjusted RGPS method at the center of Senator Johnson's report1 — are built to flag events reported more often than expected. They point investigators where to look. They cannot, by themselves, measure how often something happens or establish that the vaccine caused it.

The clearest illustration is the FDA's own Medicare monitoring. Early in the rollout it flagged four signals after the Pfizer shot in people over 65 — pulmonary embolism, heart attack, a clotting disorder, and low platelets. On closer analysis, three of the four dissolved, leaving only pulmonary embolism modestly elevated, and the researchers stated plainly that the analysis did not establish causation.2 That is a working surveillance system catching itself, and it is the template for how a signal should be treated.

The Johnson dispute sits on top of this. It is firmly established that the FDA had a genuine internal disagreement about methods, and that "masking" — where a flood of reports for one product hides smaller signals for that same product — is a real statistical phenomenon. The Senate PSI report alleged that masking-adjusted analysis uncovered roughly 25 additional statistically significant disproportionality signals that the FDA's standard method had not previously detected.3 Those were safety signals, not proof of causation. What is not established is whether sticking with the older method was deliberate suppression or defensible caution about false alarms. The documents show the disagreement; they do not settle the motive. And the claim that traveled furthest — that a later stroke signal "confirmed" an early buried one — does not survive scrutiny: the 2022 ischemic-stroke signal came from a different surveillance system, concerned a different (bivalent) product that did not exist in 2021, was investigated, and was not borne out by any other database, the Moderna bivalent, or any other country.4

Move from detection to what the confirmation studies found, and the ground gets firm. Several vaccine risks are real, measured, and accepted. Myocarditis after mRNA vaccination is the clearest, concentrated in adolescent and young adult males — roughly 71 cases per million second Pfizer doses in boys 12 to 15, and about 106 per million in boys 16 to 17, most of them mild and resolving.5 Anaphylaxis settled near five per million mRNA doses (early per-product monitoring ran higher for Pfizer and lower for Moderna before stabilizing near this figure). Thrombosis with thrombocytopenia syndrome (TTS) was causally linked to the Johnson & Johnson adenovirus-vector vaccine, with overall reporting rates around 3 to 4 per million doses in later summaries but higher in some subgroups, especially women under 50 in early monitoring. Guillain-Barré appeared as a rare association with some vaccines.6 These are not signals; they are findings — and they sit at the high-confidence end of everything in this book. But notice the scale: even the single worst-affected subgroup is about one hundredth of one percent, roughly one in nine thousand. That does not describe injury across any meaningful fraction of the vaccinated.

Against other long-used vaccines, these risks are not a separate category of danger. Influenza vaccine carries a small Guillain-Barré risk of one to two extra cases per million; Shingrix carries an FDA-recognized Guillain-Barré warning at about three excess cases per million in adults 65 and older, with the risk concentrated after the first dose; rotavirus vaccine carries an intussusception risk in infants of roughly one to six per hundred thousand.7 COVID vaccines have a distinctive profile — myocarditis in young men, the rare clot syndrome with the adenovirus vaccine — but distinctive is not unique, and each is known precisely because the monitoring worked.

The fair conclusion: COVID vaccines did cause rare serious adverse events; some were age- and sex-specific; some signals deserved more transparent investigation; and regulators should be challenged when their methods are opaque. But the existence of rare injury does not prove widespread catastrophe, and the available evidence does not support claims of population-level devastation. Skepticism is healthy. Propaganda is not. The line between the two is where truth lives.

Stage 2 — Vaccine Benefits: Lives Saved, Hospitalizations Prevented, and the Unknowable Middle

"Now, here, you see, it takes all the running you can do, to keep in the same place." — Through the Looking-Glass

Stage 2 — Vaccine Benefits: Lives Saved, Hospitalizations Prevented, and the Unknowable Middle
Stage 2 — Vaccine Benefits: Lives Saved, Hospitalizations Prevented, and the Unknowable Middle

The benefit side is harder to measure than the harm side, for a structural reason rather than a political one. Counting harms means counting events that happened. Counting benefits means estimating events that did not happen — a counterfactual world that never existed.

Within that limit, the directional evidence is strong and consistent. In matched and surveillance comparisons, unvaccinated people had substantially higher COVID death and hospitalization rates. During late 2022, unvaccinated people had about fourteen times the mortality of those with an updated booster. During Omicron, hospitalization rates ran roughly ten times higher in the unvaccinated than the boosted in one national study, and twenty-three times higher in a Los Angeles County analysis.1 That direction is high-confidence.

The aggregate magnitude is not. A Commonwealth Fund/Yale modeling estimate suggested that U.S. vaccination prevented about 18.5 million hospitalizations and 3.2 million deaths through November 2022; a WHO estimate put lives saved in the European Region at roughly 1.4 million.2 These are model estimates, not observed counts. They are built by simulating a no-vaccine world, they carry wide uncertainty, and they have drawn real methodological criticism. They are the best available estimates of a well-supported direction — and they should never be quoted as if they were body counts.

There is also a perceptual asymmetry that distorts the whole debate. Harms are visible: a person develops myocarditis, and there is a face, a family, a story. Benefits are invisible. Nobody attends the funeral that never happened or interviews the grandmother who never needed oxygen. The prevented outcomes are statistical ghosts — real, but storyless.

And the equation varies by who you are. For an eighty-five-year-old with multiple conditions, the disease risk was extreme and the vaccine risk low, so the balance favors vaccination overwhelmingly. For a healthy teenager, the disease risk was already low, so the absolute benefit shrinks even though it remains real. One statement can be entirely true for one group and misleading for another.

Beyond both groups lies what may be the largest category of all: the unknowable middle. The vaccinated young adult who never got infected may have gained little. The unvaccinated person who had a mild case may have been fine either way. The unvaccinated person who died might — or might not — have survived with vaccination. We can estimate, model, and compare populations. We cannot replay reality for any single person. That is not a failure of science; it is a consequence of living in one timeline.

Stage 3 — Age, Co-Morbidity, and Why One-Size-Fits-All Rhetoric Fails

"Everything's got a moral, if only you can find it." — Alice's Adventures in Wonderland

Stage 3 — Age, Co-Morbidity, and Why One-Size-Fits-All Rhetoric Fails
Stage 3 — Age, Co-Morbidity, and Why One-Size-Fits-All Rhetoric Fails

If one lesson outranks the others, it is that averages mislead. COVID had one of the steepest age gradients of any modern infectious disease: more than eighty percent of U.S. COVID deaths were among people over 65, and risk climbed sharply with each additional underlying condition.1 An eighty-year-old and a twenty-year-old were not playing the same game.

This creates a trap that wrecks naïve comparisons. The people at highest risk from COVID were also the first urged to vaccinate, so comparing vaccinated frail elderly against healthy unvaccinated young confounds the effect of the vaccine with the effect of age and health. The only honest comparison matches like with like: an eighty-year-old with heart disease who was vaccinated against an eighty-year-old with heart disease who was not. Do that, and the benefit is large for the old and vulnerable and smaller in absolute terms for the young and healthy — not because the vaccine changed, but because their baseline risk was lower to begin with.

The myocarditis question cuts along the same line and deserves stating plainly, because it is the one place the math genuinely narrows. For the population at large, COVID infection itself carries a higher myocarditis risk than vaccination does. But in young males, particularly after a second mRNA dose and especially with Moderna, the vaccine-related risk can match or exceed the infection-related risk.2 That is exactly the subgroup the loudest arguments are about, and an honest account names it rather than resting on the comfortable population-wide average.

Comorbidities add another layer, and they are often misused. "Most people who died had other conditions, therefore COVID wasn't the real problem" does not hold up. Millions live for years with diabetes, heart disease, or obesity; COVID frequently acted as an amplifier, turning a managed vulnerability into a crisis. A house built in a flood zone may have weaknesses, but the flood is still what destroys it.

The mature question, then, is not whether vaccines were good or bad in some universal sense. It is: for which populations were the benefits largest, for which were the risks most significant, and how should policy communicate those differences honestly? Framed that way, much of the ideological heat dissipates. Science deals in distributions and trade-offs, not absolutes — and the average is never you.

Stage 4 — Conspiracy Thinking, Political Identity, Trust, and the Information Ecosystem

"The question is, which is to be master — that's all." — Through the Looking-Glass

Stage 4 — Conspiracy Thinking, Political Identity, Trust, and the Information Ecosystem
Stage 4 — Conspiracy Thinking, Political Identity, Trust, and the Information Ecosystem

(This is the interpretive chapter. The evidence below is real and cited; the larger story it supports is a reasonable reading, not a measured certainty. It is held to the same correlation-is-not-causation standard the safety chapter applies to vaccine injury.)

By the time most people think they are arguing about vaccines, they are usually arguing about trust — in government, scientists, pharmaceutical companies, journalists, and one another. The pandemic asked the public to trust institutions at the precise moment trust was near historic lows. In the United States, the share of people who say they trust the federal government to do right most of the time fell from about 77 percent in 1964 to roughly 22 percent by 2024, and has stayed below 30 percent since 2008.1 Into that environment arrived a global emergency demanding deference to expertise.

Some of the distrust was earned; institutions made real mistakes and sometimes projected more certainty than the evidence justified. But there is a difference between "the authorities may be wrong" and "the authorities are deceiving everyone," and that leap is where conspiracy thinking begins. Healthy skepticism asks "what is the evidence?" and can change its mind. Conspiracy thinking asks "how are they hiding the evidence?" and treats every contradiction as further proof — a belief that has become immune to disproof, and therefore an identity rather than a hypothesis.

The overlap between vaccine skepticism and a broader anti-establishment politics is not speculation; it is measurable. Using October 2021 data and controlling for age, race, income, education, and rurality, partisanship was the single strongest predictor of vaccine uptake.2 A randomized trial found that exposure to online misinformation lowered intent to vaccinate by about six percentage points in both the UK and the US.3 And a study linking voter registration to mortality found that, after vaccines became available to all adults, the excess death rate among Republican voters ran about 43 percent higher than among Democratic voters.4 That finding must be handled with the same discipline as a vaccine signal: the study covered Florida and Ohio voters, not the whole United States, and it did not determine individual causes of death. It is best read as strong correlational evidence consistent with a vaccination-behavior pathway, not as proof of individual causation.

What COVID did was turn a medical question into a tribal identity marker. Information ecosystems rewarded outrage over nuance; algorithms amplified the most engaging content; communities formed around shared beliefs until changing one's mind felt like betrayal. The pattern is not uniquely American — versions appear across many democracies — though that broader claim is the read of many analysts rather than a single measurement. Vaccine skepticism did not, by itself, elect anyone. But it became one visible cylinder in a larger engine of distrust, and the vaccine became a proxy for a deeper question about who gets to define reality.

The healthy posture is the same on every side: judge claims by method, not by identity — "how do we know this is true?" rather than "who said it, and are they one of my people?"

Epilogue — Counting Both Sides of the Ledger

"Who in the world am I? Ah, that's the great puzzle!" — Alice's Adventures in Wonderland

The longer I looked, the less interested I became in who won the argument. The argument itself may have been part of the problem. Too many people arrived already knowing the answer they wanted, and treated evidence as something to collect rather than something to follow.

What emerged was less dramatic and more important. COVID was real. Vaccine injuries were real. Hospitalizations and deaths prevented were real. Statistical uncertainty was real. Institutional mistakes were real. Conspiracy theories were real. Public distrust was real. The difficulty is that all of these were true at the same time. The world rarely offers a clean choice between complete truth and complete falsehood; more often it offers competing fragments of truth, mixed with error, fear, and incentive.

The conspiracy-minded observer sees one vaccine injury and concludes the whole system is corrupt. The institutional loyalist sees one population benefit and concludes every criticism is misinformation. Both make the same mistake: they count one side of the ledger. A serious account counts both — the young man with myocarditis and the elderly woman who avoided the ICU; the signals regulators should have chased and the millions vaccinated without incident; the questions answered and the questions that can never be. And it admits, at the center of all epidemiology, that we get only one version of history.

A healthy society does not require citizens to trust every institution blindly, nor to reject every institution reflexively. It requires the harder thing: the ability to weigh evidence, revise opinions, tolerate ambiguity, and resist the comfort of simple explanations.

Afterword — Truth, Consequences, and Hallucinations

"We're all mad here." — The Cheshire Cat, Alice's Adventures in Wonderland

Afterword — Who's Hallucinating? Truth, Consequences, and the Future of Public Health
Afterword — Who's Hallucinating? Truth, Consequences, and the Future of Public Health

For decades, anti-vaccine activists occupied the role of critics, and criticism is one of truth's closest allies — it asks uncomfortable questions and forces institutions to defend their decisions. But criticism and governance are not the same thing. At some point criticism acquires influence, influence becomes policy, and policy produces outcomes. The question shifts from "what are you against?" to "what happens if you get your way?" Ideas stop being judged by how effectively they criticize and start being judged by the consequences they produce.

And we are beginning to see consequences. Kindergarten vaccination coverage has slipped from a pre-pandemic 95 percent to 92.5 percent — below the threshold needed to stop measles spreading — while exemptions have reached a record 3.6 percent. In 2025 the United States recorded about 2,288 measles cases, its largest resurgence since 1992, surpassing the post-elimination high of 1,274 set in 2019, with three deaths — the first U.S. measles deaths in a decade. Reported pertussis cases surged in 2024, peaking around November 2024, and have declined since. U.S. tuberculosis cases rose through 2024, then provisional data showed a small decline in 2025 — about 10,260 cases, a roughly one percent drop — while remaining above pre-pandemic levels.1 Whether these changes ultimately weaken society is a verdict history has not finished writing — but for more than a century, sanitation, clean water, vaccination, and antibiotics produced one of the greatest increases in human life expectancy ever recorded. People generally lived longer because of those systems, not despite them.

Here is the central irony. Many people sincerely fear that vaccination and public-health programs are instruments of control or even population reduction. Yet historically their overwhelming effect has been population preservation. If societies weaken the systems that reduced infectious disease because they believe those systems are the threat, reality will render a verdict — not a political one, a biological one. Viruses do not vote. Pathogens do not care whether someone is progressive or conservative, vaccinated or unvaccinated. Nature is indifferent to ideology and responds only to conditions, and conditions are what policy creates. Reality keeps score.

That brings us to the strangest discovery in writing this. The deeper the investigation went, the harder it became to find a single villain. The activist believes they are protecting families. So does the epidemiologist. The official believes they are protecting society. So does the dissident. Everybody thinks they are Alice; everybody thinks someone else is the Mad Hatter; and everyone is increasingly convinced the other side has lost its mind. And vice versa.

Which raises the final question. The AI hallucinates when it generates a convincing answer unsupported by evidence. The conspiracy theorist hallucinates when scattered facts become an all-encompassing plot. The institution hallucinates when consensus becomes certainty. The politician hallucinates when ideology overrides reality. The activist hallucinates when the cause outranks the facts. The citizen hallucinates when identity outranks curiosity. Perhaps the uncomfortable answer is that all of them do, at least sometimes — including the author, and including the reader.

A society can survive disagreement, bad leaders, and uncertainty. What it may struggle to survive is a condition in which half the population is convinced the other half has lost its mind — and vice versa. The difference, in the end, is this: when my hallucinations are discovered, somebody patches the software. Humans tend to elect theirs to public office.

On that note, I leave the last word to reality. It remains undefeated.

Appendix A — Glossary

Part I — How Evidence Works

Absolute risk — The actual chance an event happens to an individual. A risk of 1 in 10,000 means that, on average, one person in 10,000 would be expected to experience the event.

Relative risk — A comparison between two risks. If one group experiences an event twice as often as another, the relative risk is 2.0. Relative risk can sound dramatic even when the absolute risk remains very small.

Correlation — A statistical relationship in which two things occur together more often than chance would predict. It does not establish that one causes the other.

Causation — A relationship in which one factor directly contributes to producing another. Establishing causation generally requires several independent lines of evidence beyond simple correlation.

Confounding — A distortion caused by an outside factor that influences both variables being studied, creating the illusion of causation where none exists.

Statistical significance — A measure of how unlikely an observed result would be if there were truly no effect. Significance does not by itself imply practical importance.

Confidence interval — A range within which the true value is believed likely to lie. Wider intervals signal greater uncertainty.

Multiple-comparisons problem — When many outcomes are examined at once, some will look unusual by chance alone; statistical methods are used to limit false discoveries.

Part II — Vaccine Safety Monitoring

Safety signal — An observation that an event may be occurring more often than expected and deserves investigation. A signal is not proof of causation.

Signal detection — The process of identifying unusual patterns in safety data that may warrant further study.

Active surveillance — Monitoring that uses linked medical records and population databases to compare actual outcomes in vaccinated and unvaccinated groups.

Passive surveillance — Systems that accept reports from the public and clinicians without independently verifying causation.

VAERS — The U.S. Vaccine Adverse Event Reporting System, a passive system designed to surface potential signals.

Vaccine Safety Datalink (VSD) — An active-surveillance system linking vaccination and healthcare records across participating organizations.

BEST Initiative — The FDA's Biologics Effectiveness and Safety system, which monitors vaccine and biologic safety using large healthcare databases.

Background rate — The normal rate at which an event occurs in a population regardless of vaccination.

Disproportionality analysis — A statistical method for identifying events reported more frequently than expected.

MGPS (Multi-item Gamma Poisson Shrinker) — The FDA's traditional disproportionality method.

RGPS (Regression-adjusted Gamma Poisson Shrinker) — A modified approach designed to reduce masking effects.

Masking — When very common reports for one outcome obscure less common signals tied to the same product; the phenomenon at the center of the MGPS-versus-RGPS dispute.

Part III — Vaccine Safety Outcomes

Adverse event — Any medical event occurring after vaccination, whether or not the vaccine caused it.

Serious adverse event — An adverse event involving hospitalization, disability, a life-threatening condition, or death.

Myocarditis — Inflammation of the heart muscle.

Pericarditis — Inflammation of the membrane surrounding the heart.

Anaphylaxis — A severe allergic reaction requiring immediate treatment.

Guillain-Barré syndrome (GBS) — A rare disorder in which the immune system attacks peripheral nerves.

TTS (Thrombosis with Thrombocytopenia Syndrome) — A rare clotting disorder with low platelets, linked to certain adenovirus-vector COVID vaccines.

Intussusception — A rare bowel obstruction that occurs naturally and has also been associated with some rotavirus vaccines.

Part IV — Measuring Benefits

Vaccine effectiveness — How much vaccination reduces illness, hospitalization, or death under real-world conditions.

Vaccine efficacy — The reduction in disease observed under the controlled conditions of a clinical trial.

Counterfactual — A hypothetical scenario describing what would have happened under different circumstances; the basis of "lives saved" estimates.

Model estimate — A figure produced through simulation rather than direct observation.

Excess mortality — Deaths above what would normally be expected for a population and period.

Hospitalization rate / mortality rate — The frequency of hospital admissions or deaths within a defined population.

Part V — Public Health

Herd immunity — Population-level resistance to transmission arising from immunity through vaccination, infection, or both.

Herd-immunity threshold — The share of a population that must be immune to substantially reduce transmission (about 95% for measles).

Endemic — A disease continuously present in a population.

Epidemic — A rise in disease above normal expectations.

Pandemic — An epidemic spreading across multiple countries or continents.

Variant — A genetically distinct version of a virus.

Comorbidity — A medical condition existing alongside another.

Part VI — Information and Hallucinations

Misinformation — False or inaccurate information shared without intent to deceive.

Disinformation — False information shared deliberately to mislead.

Conspiracy theory — A hypothesis that explains events as the product of coordinated secret action by powerful actors. Such claims should be evaluated on their evidence, like any other claim.

Confirmation bias — The tendency to seek or interpret evidence in ways that support what one already believes.

Motivated reasoning — The tendency to evaluate information according to the conclusion one wants rather than the evidence.

Hallucination (artificial intelligence) — A generated response that sounds plausible but is unsupported by evidence.

Hallucination (human) — Used in this book as an informal term for a narrative that becomes detached from the available evidence while remaining psychologically convincing.

References: Endnotes

Endnotes are numbered fresh within each section and keyed to the Bibliography (Appendix B) by section letter and author/year, rather than repeating full citations here. Light density: notes mark the load-bearing factual claims, named studies, and contested figures; ordinary connective and interpretive prose is left unmarked.

Stage 1 — Vaccine Safety

  1. Signal-detection methods — MGPS, the masking-adjusted RGPS, and the masking phenomenon itself: Bib. §G — DuMouchel (1999); Szarfman et al. (2002); DuMouchel & Harpaz, RGPS (2012); Harpaz et al. (2022).
  2. FDA–CMS Medicare near-real-time monitoring; four early Pfizer signals in adults 65+, three not sustained, no causation established: Bib. §A — Wong et al., Vaccine (2023).
  3. FDA internal methods dispute and the ~25 added disproportionality signals under masking-adjusted analysis: Bib. §A — PSI, Failure to Warn (2025); masking as a quantified phenomenon, Bib. §G — Harpaz et al. (2022).
  4. 2022 ischemic-stroke signal (Pfizer bivalent booster, Vaccine Safety Datalink), investigated and not corroborated by other databases or countries: Bib. §A and §B — CDC/FDA Vaccine Safety Datalink ischemic-stroke evaluation (2023).
  5. Myocarditis after mRNA vaccination, age- and sex-specific rates (~71 per million second Pfizer doses in boys 12–15; ~106 per million in boys 16–17): Bib. §B — Oster et al., JAMA (2022).
  6. Anaphylaxis (~5 per million mRNA doses); TTS causally linked to the Janssen adenovirus-vector vaccine; Guillain-Barré as a rare association: Bib. §B — CDC anaphylaxis surveillance (2021); FDA/CBER Janssen TTS review (2021); Woo et al., MMWR (2021).
  7. Comparator vaccine risks — influenza-associated GBS (~1–2 per million); Shingrix GBS (~3 per million, adults 65+); rotavirus intussusception (~1–6 per 100,000 infants): Bib. §F — CDC Influenza Vaccine Safety/GBS; Goud et al. (2021); Yih et al. (2014).

Stage 2 — Vaccine Benefits

  1. Observed mortality and hospitalization by vaccination status — ~14× mortality in the unvaccinated vs. boosted (late 2022); ~10× hospitalization nationally and ~23× in Los Angeles County during Omicron: Bib. §C — Johnson et al., MMWR (2023); Havers et al., JAMA Internal Medicine (2022) and CDC MMWR (LA County, 2022).
  2. Modeled aggregate benefit (labeled estimates, not counts) — ~18.5 million hospitalizations and ~3.2 million deaths prevented in the U.S. through Nov. 2022; ~1.4 million lives saved in the WHO European Region: Bib. §C — Fitzpatrick et al., Commonwealth Fund (2022); WHO Regional Office for Europe (2024).

Stage 3 — Age, Co-Morbidity, and One-Size-Fits-All Rhetoric

  1. COVID age gradient — more than 80% of U.S. COVID deaths among people over 65, with risk rising sharply per additional underlying condition: Bib. §C — Johnson et al., MMWR (2023); CDC mortality-by-age surveillance.
  2. Myocarditis, infection vs. vaccination — higher after infection for the population at large, but in young males (especially after a second mRNA dose, particularly Moderna) the vaccine-related risk can match or exceed the infection-related risk: Bib. §B — Patone et al., Circulation / Nature Medicine (2022); Oster et al., JAMA (2022).

Stage 4 — Conspiracy, Trust, and the Information Ecosystem

  1. Decline in trust in the federal government — ~77% (1964) to ~22% (2024), below 30% since 2008: Bib. §D — Pew Research Center (2025).
  2. Partisanship the single strongest predictor of vaccine uptake (Oct. 2021 data, controlling for age, race, income, education, rurality): Bib. §D — Kirzinger et al., KFF (2021).
  3. Randomized trial — online misinformation lowered vaccination intent by ~6 percentage points in the UK and the US: Bib. §D — Loomba et al., Nature Human Behaviour (2021).
  4. Excess death rate ~43% higher among Republican than Democratic registered voters after adult eligibility (Florida and Ohio only; no individual cause of death determined — correlational): Bib. §D — Wallace et al., JAMA Internal Medicine (2023).

Afterword — Truth, Consequences, and Hallucinations

  1. Disease-resurgence indicators — kindergarten coverage down to 92.5% with exemptions at a record 3.6%; ~2,288 U.S. measles cases in 2025 (largest since 1992; three deaths, first in a decade); pertussis surge peaking ~November 2024; tuberculosis ~10,260 cases in 2025 (~1% decline, still above pre-pandemic levels): Bib. §E — Seither et al., CDC SchoolVaxView; CDC Measles Cases and Outbreaks / MMWR (2025); CDC pertussis surveillance; CDC National TB Surveillance System (provisional 2025). Worldwide measles-vaccine lives-saved estimate (~93.7 million, 1974–2024): Bib. §E — CDC / WHO measles update.

Appendix B — Bibliography

Organized by subject so a reader can find all sources on a given topic together. Citations are at author/title/journal/year level and should be expanded to full bibliographic form, with a final human proof, before publication. Model estimates are labeled.

Section A — FDA, CDC, and Senate documents

Section B — COVID vaccine safety, by topic

Section C — Vaccine benefit studies

Section D — Public trust, politics, and information

Section E — Disease resurgence

Section F — Comparator vaccines

Section G — Methodology and statistics

Author's Note

A book that asks who is hallucinating owes the reader an honest answer about its own construction. This one was written with the help of artificial intelligence.

I should be precise about what that means, because the phrase can hide as much as it reveals. I did not ask a machine to write a book and put my name on it. What I did was closer to my day job, where I build AI systems and spend most of my time trying to stop them from confidently asserting things that are not true. I used these tools the way you would use a fast, tireless, occasionally unreliable research assistant: to find sources, to draft and redraft, to check figures against the record, and to assemble the whole into the thing you are holding.

The irony is not lost on me. A book largely concerned with how machines fabricate plausible-sounding claims was assembled with a machine that can fabricate plausible-sounding claims. That tension is the reason the book exists, not an embarrassment to be hidden. It forced exactly the discipline the rest of these pages demand of everyone else.

Here is what that discipline looked like. Every figure was checked against a primary source — a study, a surveillance report, an agency dataset — and not accepted because it sounded right or because a model produced it. Where a number was a model's estimate rather than a measured count, it is labeled as an estimate. Where I could not verify something, I left it flagged rather than smoothing it over. Where the evidence was contested, I tried to give the competing readings their fair weight instead of quietly choosing the one I preferred.

That process caught real errors, and I have left the scars visible on purpose. A draft infographic asserted a kindergarten vaccination figure that did not match the record; checked against the CDC's data, it was corrected to about 92.5 percent, and a measles surge the same graphic had pinned to the wrong year was moved to 2025, where it belonged. A comparator statistic I had carried for some time — the rate of anaphylaxis after the HPV vaccine, which I had down as roughly 1.7 per million — collapsed under checking: the published rates vary so widely by product and setting that no single number was defensible. An early Australian program reported around 2.6 per 100,000; later surveillance found far less; the honest move was to replace the tidy figure with a sourced range. These were not the failures of one bad tool. They were the ordinary failure mode of any system, human or machine, that reaches for a confident answer faster than it reaches for the evidence — which is the exact failure this whole book is about, in institutions, in activists, in algorithms, and in ourselves.

There is a habit I have come to distrust, having watched it produce error after error in work people published with AI: they did not read what the machine wrote. Not closely. They accepted fluent output as finished work, and the mistakes that slipped through were not exotic — they were the kind a careful second reading would have caught. So the most important tool I used was not any model. It was rereading. I read what these systems produced, and then I read it again, and I checked the claims that mattered against the source rather than against my own sense of whether they sounded right.

I also made the models check each other. No single system was trusted to be its own auditor. The bulk of this work was done with Claude (Opus 4.8, on a Pro subscription) and ChatGPT (running GPT-5.5, on a Plus subscription), used in tandem so that each could be turned on the other's output — a figure one produced, the other was asked to verify against the record. For some spot checks I added a third opinion from Grok 4.3, in its Fast mode. Three systems do not agree their way to the truth any more than three people do; what the cross-checking buys you is friction — a second and third chance for a confident, wrong answer to be caught before it reaches the page. The verifying still has to be done by a person willing to go to the primary source. The models disagreeing is only the prompt to go and look.

The division of labor, then, is this. The machines helped me find, draft, check, and build. The judgment about what to trust, what to include, and what it all means is mine. So are the conclusions, and so are any errors that survived the checking — because a tool cannot be accountable, and the author can.

I asked the institutions, the activists, and the algorithms in this book to count both sides of the ledger and to follow the evidence even when it was inconvenient. It would have been dishonest to demand less of the way the book was made.

Disclaimer (to be reviewed by legal counsel — US, UK, Portugal)

A medical/informational disclaimer, and — importantly — a defamation/libel review, given that the text names and characterizes real public figures (UK law especially relevant).

About the Author

Robert Curtis is an information technology solutions architect who has spent thirty-five years inside enterprise systems, and the last several building AI tools designed to reduce machine "hallucination." Born in Stockton-on-Tees in 1961, he has lived and worked across the United States, Australia, Sweden, the Netherlands, Switzerland, the United Kingdom, and Portugal. He writes both fiction — mainly historical and science fiction — and nonfiction.