You have probably watched this happen in real time: a scientific claim gets debunked, the researchers move on, and years later the claim is still circulating as if nothing changed. Stand like Wonder Woman for two minutes before a job interview and you’ll perform better. The lead author of that study publicly walked it back a decade ago. It is still being repeated.
Chapter 1 drew the pipeline in the abstract. Here I walk four real findings through it, stage by stage, to see what each gate actually does to a claim and where a real one gets lost, a shaky one sails through, or a correction dies in a drawer. Then I ask a question the diagram skips: what kind of thing is moving? A hormone result and a currency both travel the same pipe, but only one of them has a fact outside the pipe to check against, and that difference changes what “failure” even means.
Four pipeline outcomes
Picture the pipeline as a relay race an idea runs from the lab out to the public. An idea can drop out at any handoff, and where it drops out decides what it leaves behind. The four cases below each exit at a different point:
- Made it all the way to meme. The finding traversed every stage and its meme-form persists in culture, often even after the correction reaches the upstream stages.
- Died at consensus. The finding was published (it passed measurement, analysis, and the publishability gate) but the broader scientific consensus rejected it after independent replication failed. The consensus gate did its job and the meme never spread.
- Died at curation. The finding, often a corrective, passed consensus but never made it through journalism, popularization, or the cultural curation layer. The original kept spreading; the correction sat in the journals.
- No theory at all. The “finding” never had a measurement origin. It was manufactured content: injected at the curation or meme stage and decoded by receivers as if it had come from upstream. The downstream gates select on form, not provenance, so the manufactured variant occupies the same stages a real finding would.
Case 1: Power posing — all the way to meme
The 2010 paper by Carney, Cuddy, and Yap reported that holding “high-power” poses (hands on hips, chest out, feet apart) for two minutes raised testosterone, lowered cortisol, and improved performance on a risk-taking task. It appeared in Psychological Science as an actionable, free intervention with measurable physiological effects.
Each stage handled it as Chapter 1 predicts. Measurement recorded saliva hormone levels and risk-task scores from 42 participants. Analysis ran the standard significance tests and reported effects on both hormones and behavior at p < 0.05. Consensus, at the publication gate, passed it into a top journal. Curation embraced it: Cuddy’s 2012 TED talk on the work has been viewed more than 70 million times, a book followed (Presence, 2015), business publications adopted the framing, and some corporate training programs picked it up as a pre-interview trick.
At the meme stage it became “stand like Wonder Woman before your interview.” The compressed form dropped nearly every qualification (the tiny sample, the specific hormone claims, the modest behavioral effect) and landed as a clean prescription: actionable, identity-friendly, frictionless to apply. A 42-person hormone study became a TED-talk-and-bookshelf staple in five years, and at every stage the gates passed the variant they were tuned to pass.
Then the corrective. In 2015 Ranehill and colleagues ran a much larger replication (n = 200) and found no hormone effects and no behavioral ones. In 2016 Dana Carney, the original lead author, publicly disavowed the work and said she no longer believed the effects were real. (Cuddy has kept defending a narrower “felt power” effect, so the field’s picture is a sharp downgrade rather than a clean reversal.) The corrective passed measurement, analysis, and consensus; within psychology, power posing is now a canonical case of the replication crisis. And the meme kept spreading anyway. The TED talk is still watched, the book still read, the pose still recommended. Carney’s disavowal drew modest coverage in scientific media and almost none in the general media that had broadcast the original.
The pipeline can run its full length and leave a persistent meme even when the finding is wrong, because the correction has worse curation fitness than the thing it corrects. At every stage the gates reward content that is actionable, flattering, and novel, and a correction is systematically less of all three than the claim it retracts. This is the selection dynamic Chapter 10 takes apart; here it is enough to see that the failure is real and continuous, not an occasional glitch.
Case 2: Arsenic-life — died at consensus
In December 2010 a paper by Felisa Wolfe-Simon and NASA-funded colleagues appeared in Science claiming that a bacterium from California’s Mono Lake (designation GFAJ-1) could substitute arsenic for phosphorus in its DNA backbone. If true, terrestrial life could be built on chemistry outside the standard set of elements biology runs on, an enormous result. NASA held a high-profile press conference timed to publication, and the curation layer began packaging the news for the meme stage (“alien life on Earth”) within days.
Then the consensus gate fired, harder than the publication gate had. Within days the microbiologist Rosie Redfield posted a detailed critique; within weeks, labs announced replication attempts. Science published eight technical comments in 2011, and in 2012 two independent refutations showing that GFAJ-1’s DNA held no arsenate, and that the original signal came from arsenic contamination in the sample. Consensus settled to negative inside eighteen months.
The meme barely propagated past the first press cycle. “Alien life found on Earth” spiked for a couple of days, then went quiet as the consensus signal came through, and most people who heard of it at all heard “that NASA arsenic claim that turned out to be wrong.” The later stages never built the cultural artifact they were about to build, because the consensus gate caught the variant before curation could compound it.
The consensus stage can and does work. Arsenic-life is the counterexample to power posing: same era, same kind of splashy publication, opposite outcome. So the question is not whether the gates ever catch errors, but what makes them catch one claim and wave another through. Part of the answer is already visible. An extraordinary claim in biochemistry has a bacterium any lab can re-test; a modest claim in social psychology, with a TED talk attached, has a story people want to be true. The gate faces different pressure in each case, and Chapter 10 is about where that pressure comes from.
One caution about what “died at consensus” means. Arsenic-life is famous partly because active, public rejection of a high-profile claim is rare. Most claims that fail at consensus fail quietly: the paper appears, few engage it, it gathers no citations, and it never reaches curation at all. That is consensus inattention, not consensus rejection, and it is the far more common fate. Arsenic-life shows the gate firing; usually the gate simply never turns to look.
Case 3: Stanford Prison Experiment — the correction died at curation
Philip Zimbardo’s 1971 Stanford Prison Experiment is one of the most-cited social-science studies of the twentieth century. Volunteers were randomly assigned to be “guards” or “prisoners” in a basement mock prison; within days the guards turned cruel and the prisoners broke down, and Zimbardo halted it early. The finding, that ordinary people behave sadistically when handed a guard role, entered textbooks, documentaries, a 2015 feature film, and the pop-psychology canon as the demonstration that situation overwhelms character.
The pipeline ran cleanly: publication, favorable early reception, heavy curation (Zimbardo became a public intellectual on the strength of it), and a durable meme. For four decades the finding sat in the meme stage as load-bearing cultural content.
In 2018 the picture changed. Ben Blum’s reporting drew on archival tapes and notes showing that Zimbardo had coached the guards on how to behave and that one prisoner’s famous breakdown was performed. Thibault Le Texier’s 2018 book Histoire d’un mensonge (translated in 2024 as Investigating the Stanford Prison Experiment) worked through the archive and concluded the study was, at minimum, methodologically broken, with Le Texier himself calling parts of it fabricated. The social-psychology community largely accepted the critique. The correction passed measurement, analysis, and consensus: within the field the experiment is no longer cited as evidence for the situational hypothesis.
And then it died at curation. Textbooks have been slow to update; introductory courses still teach the experiment as canonical, pop-psychology books still cite it, the film is not being re-cut with caveats, and most non-specialists who have heard of it still believe it shows what it originally claimed. Even coverage of the replication crisis tends to skip it, because the cleaner cases (power posing among them) don’t require unseating a story this embedded.
The correction has all the wrong curation properties: technical, unflattering to a beloved figure, narratively deflating, with no clean replacement story. The original meme has all the right ones: dramatic, morally clean, self-contained, and flattering to the reader’s sense that they’d have resisted. An accurate correction with bad curation properties loses to a wrong original with good ones, even after consensus has fully ratified the fix. This is the shape Chapter 8 calls “preservation held, training hollowed”: the journals hold the correction; the readers who would need to be re-taught it have not been.
Case 4: Astrology — manufactured content in the pipeline
The first three cases each began with a measurement. Some content in the pipeline has none. The cleanest example is astrology. The astronomical positions it references are real, accurate, and ancient. But the interpretive framework, that the positions at your birth shape your personality and predict events, has no measurement origin. It was never derived from data and never passed a consensus gate in any culture that had one. Yet it occupies every downstream stage a real finding would: authoritative-feeling print sources, popular-press curation (the horoscope column), and full meme-stage integration (“oh, you’re such a Capricorn”). The gates between curation and meme cannot tell measured from manufactured, because they were built to check form, not provenance.
The “you only use 10% of your brain” claim works the same way. It has no traceable measurement origin (its most-cited “source,” William James, said no such thing) but circulates as if it did: a textbook entry, a movie premise (Lucy, 2014), a self-help staple, undented by waves of corrective neuroscience. Manufactured content can populate the late stages without ever passing through the early ones, because the late gates check form, not origin. Chapter 5b works the mechanism out; here the point is just that this is a normal, recurrent feature of the pipeline. Astrology has held the meme stage for two thousand years, the brain myth for more than a century. These are stable equilibria, not transient failures.
Reading the four cases
Stand back and one pattern holds all four. The pipeline can fail at every stage, and it can work at every stage: arsenic-life was caught, the journals hold the Stanford correction, power posing’s standing inside the field is now accurate. So the diagnosis is not “the pipeline is broken.”
The pipeline doesn’t break at random. It fails in a direction. A correction is almost always less dramatic, less flattering, and less immediately useful than the myth it is trying to replace, so wherever the gates reward drama, flattery, and usefulness, the correction loses the race, every time, no matter how right it is.
That is what any fix has to be designed against. “Tell the truth louder” misreads the problem, because volume was never the issue; the gates are. A pipeline whose gates reward drama will carry drama whether it is true or not, and drop a correction whether it is right or not. The four cases are four samples from one continuous surface, and changing what the gates reward is the whole of the rest of the book.
The three realities
Walking the cases surfaces a question the diagram skips: what kind of reality is moving? All four cases are claims about objective reality, states of affairs that hold whether or not anyone believes them, so the pipeline can be checked against an external fact (the next replication, the methodological audit) and the words true, false, corrective, manufactured apply cleanly. But the pipeline carries other kinds of content, and the vocabulary works differently for them. Harari, in Nexus, names three categories:
- Objective reality. Atoms, planets, bacteria, the speed of light: a fact of the matter the pipeline transmits or distorts. All four cases live here.
- Subjective reality. Pains, dreams, private preferences. These don’t really traverse the pipeline; a report of a subjective state is an objective claim about it, but the state itself is not what moves. The pipeline analysis mostly doesn’t engage this category.
- Intersubjective reality. Currencies, nations, corporations, laws, religions, gods: things that are real because enough people act as if they are. They have hard consequences (you can be arrested for breaking a law; a dollar buys bread) but the consequences exist because the network agrees they do. The pipeline is not transmitting these realities so much as constituting them.
The vocabulary of true and corrective starts to misfit that third category, and the misfit is worth one worked example.
A fifth story: money
Take the plainest intersubjective truth going: the twenty in your pocket buys groceries. True; any register confirms it. Now set it beside Case 2’s kind of truth. When arsenic-life collapsed, the bacterium was unbothered; GFAJ-1’s DNA contained what it contained, and every replication could appeal to it. If every mind on Earth forgot the episode, the phosphorus would still be there to find again. Run the same erasure against the twenty and nothing survives: strip the agreement out of every head and the bill is ink on cotton, with no residue of value left out there to re-measure, because the value was never out there. The truth was the agreement. Both sentences are genuinely true; only one has a fact outside the network to lean on. That is the whole of Harari’s third category.
The difference stops being philosophical the moment the pipeline touches money in anger.
A failing currency cannot be corrected, only re-agreed. When a monetary agreement dissolves there is no upstream fact a correction could carry back. In Weimar Germany in 1923, and in Zimbabwe eighty-five years later, people kept calling the notes money (the word held) while spending wages within hours of receiving them, which is what the agreement actually consists of. The words outlived the truth. The recoveries, when they came, were not corrections but re-constitutions: a brand-new mark, or the wholesale adoption of someone else’s dollar, a population walking out of one agreement and into another. There is no bacterium under a currency.
Coverage of the truth feeds the truth. A bank run is a claim (the bank is failing) that makes itself true by propagating: holders hear it, act on it, and the acting is the failure. Nothing is distorted in transit; the transit is the event. Britain’s Northern Rock run in 2007 still moved at the speed of queues on the pavement. In March 2023, Silicon Valley Bank lost tens of billions of dollars in a single day, the run coordinated in group chats and timelines, the agreement moving at meme speed while the institutions built to steady it still moved at committee speed.
Constitution can now be watched from zero. Bitcoin is the rare case where the birth of an intersubjective truth sits on the public record. In 2009, “bitcoin is worth something” was simply false. No measurement changed and no discovery was made; through nothing but accumulating agreement, each new holder one more unit of the thing being agreed, the claim became true, with “internet money” doing the recruiting. And when such a network’s agreement splits, there is no measurement to appeal to: Bitcoin’s block-size dispute, like Ethereum’s over a catastrophic theft, ended in a fork, two ledgers and two communities, each truth fully generated by the network still holding it. The fork did for these constitutions what the consensus gate did for arsenic-life, with one difference that should land uncomfortably: it didn’t pick the winner. It set how many truths there are now.
What the third reality changes
None of the money story fits the four-exit taxonomy, because the taxonomy assumed a referee: a fact outside the pipeline that corrections could appeal to and gates could fail against. For objective claims the book’s diagnosis stands as written. For intersubjective ones, the failure mode is different in kind. When the pipeline mangles an objective fact it garbles something that stays true regardless, sitting out there to be measured again. But for things real only because we agree on them, there is no fact off to the side: change how the agreement is made and you don’t distort the reality, you change it. A captured currency-issuance process produces a different currency; a captured legal system, different laws. That is the scarier failure, because nothing outside is left to check the result against. The machinery for content without a referee, the constitutive regimes, what capture means when there is no truth to tune a gate against, and why the fork is the correction’s structural sibling, is worked out in the intersubjective-truth note. The chapters ahead flag the intersubjective case wherever it changes their conclusions.
Where I’m still uncertain
- The carving itself is live. The intersubjective note carves per-claim by generator (truth from agreement versus truth from reality), which handles the easy mixed bundles (a founding story; astrology’s false personality claims riding on a real membership practice) but has not been defended at the depth a critic would demand for the hard ones: brands’ quasi-objective promises, or whether scientific consensus tracks a generator or quietly is one.
- Clean correctives are the exception. Power posing is unusually tidy: the author disavowed and the replication failed dramatically. The common case is messier, with no disavowal, partial replications, and a split field. The stage-by-stage reading here assumes a clean corrective signal and hasn’t engaged the noisy one.
- The gradient is asserted, not measured. The claim that corrective signal has systematically worse gate-fitness is consistent with these four cases but not established across many. There is a literature to hold it to (the spread of misinformation versus its corrections, and retraction visibility versus original-paper citation rates), and the book should engage it rather than lean on four cases.
← Chapter 1: The Information Landscape · Chapter 3: The Human Time Budget →