Science
How Knowledge Gets Manufactured, Where the Machine Fails, and Why It Still Corrects Itself
Introduction
Nearly every page of this site leans on a phrase like "research shows" or "your own scholars established." This page opens the factory those phrases point at. It is overdue. The site keeps asking you to weigh one study against another - Lancet estimates against their critics, one democracy index against two that disagree - without ever explaining how the weighing works.
Two claims run through what follows, and the tension between them is the subject. The machine that manufactures knowledge is deeply, measurably flawed - its critics are mostly right, and the evidence for that comes from the machine itself. And it remains the only large institution in this Power cluster that systematically pays people to prove it wrong. Courts correct by appeal to power, media by market, bureaucracies rarely; science corrects by organised attack on its own output. Slowly, expensively, and - so far - reliably.
One definition before starting. "Science" here does not mean the method taught in school, and it does not mean the truth. It means the institution: journals, careers, grant committees, reputations, and the more than two trillion dollars a year that flow through them. The method is four hundred years old. The institution, as you are about to see, is barely one lifetime old - and that fact explains more of its behaviour than anything else.
Younger Than Your Grandparents
In 1936 Albert Einstein submitted a paper to the American journal Physical Review. The editor sent it to an anonymous referee, as journals now routinely do. Einstein was furious. He had not authorised the editors, he wrote back, to show the manuscript to specialists before it was printed; he withdrew the paper and never sent the journal another. The referee, it later turned out, was right about the error. The detail worth keeping is not the irony. It is that in 1936 the greatest physicist alive could regard peer review as an insult, because it was not yet a norm.
Nearly everything people treat as the eternal architecture of science is similarly recent. Nature, founded in 1869, made external refereeing systematic only in 1973. The government grant system - the idea that the state funds open-ended research in peacetime - dates to Vannevar Bush's 1945 report and the agencies built on it. The metrics that now govern careers are younger still: the journal impact factor was first published in 1975, the h-index in 2005. A researcher retiring today is older than most of the machinery that judged their life's work.
The recency matters for the same reason it matters everywhere else on this site. Rules this new are not laws of nature, and the people enforcing them overestimate their permanence. The institution scaled a thousandfold in one century - from a few thousand scientists to around nine million, from private patrons to state budgets. Its quality-control mechanisms were improvised during the scaling, not designed before it. Most of what is wrong with science is what you would expect of any factory whose inspection department was bolted on while the line was already running.
The Currency Problem
The institution runs on one currency: the published paper. Jobs, grants, tenure, prizes, and national rankings all convert to it. And the moment publication became the measure of knowledge production, Goodhart's law - the mechanism this site's incentives page treats at length - did to science what it does to everything: the measure became the target, and the target detached from the thing measured.
The distortions are documented in science's own journals. Positive results publish; null results mostly do not, so entire literatures overstate every effect they contain. Analytical flexibility does quiet work: run twenty defensible analyses on honest data, report the one that crossed the significance line, and you have fabricated nothing while still being wrong. The practice is common enough to have a name - p-hacking - and a measurable signature in the distribution of published results. Above the individual level sit industrial versions: paper mills selling authorship on fabricated manuscripts at scale. One major publisher retracted more than ten thousand papers in a single year, most from that pipeline; citation rings inflate the metrics the rankings reward. For years several countries paid cash per paper; China, the largest, formally abolished the practice in 2020 after its consequences became internationally visible.
The honest framing is not that science is full of villains. It is that science is full of rational people responding to a badly specified reward function, which is exactly what the rest of this site would predict. The wonder, and the next section's subject, is that the same community that built the bad metric also built the audit that exposed it.
The Crisis That Named Itself
In 2005 the epidemiologist John Ioannidis published a paper titled "Why Most Published Research Findings Are False," and the institution's response was to make it one of the most-cited papers of the decade. A decade later came the test: a consortium of researchers re-ran one hundred published psychology experiments and got the original result barely a third of the time. A parallel project in cancer biology managed to repeat about half the experiments it attempted, with effects far smaller than first reported. The phrase "replication crisis" was coined by scientists, about scientists, in scientific journals.
One detail from those projects deserves more fame than it has. When researchers were asked to bet on which findings would replicate - in real-money prediction markets - they were right about seven times in ten. The community collectively knew which of its published results were fragile before anyone re-ran them. The knowledge existed; the incentive to act on it did not, because replication was nobody's career and a published correction was nobody's promotion.
The fraud cases tell the same double story. A Dutch social psychologist fabricated data across dozens of papers before colleagues caught the statistics being too clean. A foundational Alzheimer's paper stood for sixteen years before a neuroscientist spotted manipulated images; it has since been retracted, along with a research programme built partly on it. Two of behavioural science's best-known honesty researchers were separately accused - by fellow academics working in their spare time - of publishing studies about honesty containing fabricated data; retractions followed, both deny wrongdoing, and the affair is now taught as a case study. The president of Stanford resigned in 2023 after his student newspaper forced a review of manipulated figures in his old lab. Note the pattern in who did the catching: a blogger with a statistics habit, volunteer image detectives, a student journalist, and the internet forum PubPeer. The machine's immune system turned out to live largely outside its formal bodies.
And the fixes, once shame made them fundable, worked measurably. Preregistration - publishing your hypothesis and analysis plan before collecting data - closes the forking-paths loophole. Registered reports, now offered by hundreds of journals, commit to publishing a study based on its design rather than its result, which deletes the incentive to torture the data. Journals screen images with forensic software. None of this fixed the underlying currency. All of it made the counterfeit harder to pass.
Who Pays Decides What Exists
Follow the money and the word "science" splits in two. Of the roughly $900bn the United States spends on research and development each year, about three-quarters is industrial - aimed at products, owned by shareholders, published selectively if at all. The public image of science - curiosity, open publication, basic questions - describes the minority share paid for by governments and foundations. That minority is load-bearing: it produces the un-ownable knowledge, the training, and the seed corn that industry later harvests, which is why economists keep finding that public research pays returns private markets would never wait for.
How the public share is allocated shapes what exists to know. The dominant mechanism - project grants awarded by committees of peers - is a postwar invention with known side effects. When success rates fall toward one in five, committees fund the safe and the incremental; genuinely strange proposals die in the averaging of scores. The average age at which an American biomedical researcher wins their first major independent grant is now in the mid-forties, which means the institution asks its members to spend their most original decade as apprentices. The counterexamples prove the constraint is a choice: agencies that fund people rather than projects, and the pandemic-era experiments that made grant decisions in days rather than months, produced outsized results by relaxing exactly the caution the committees institutionalise.
The allocation also has silences, and they rhyme with this site's page on the unthought. Replication is chronically unfunded because it is nobody's discovery. Maintenance of datasets, instruments and long-term measurements is unfunded for the same reason maintenance is uncelebrated everywhere. Negative results - the map of what is not true, the most reusable product science makes - mostly go unwritten. The knowledge that exists is not the knowledge that mattered most. It is the knowledge that was fundable, publishable, and finished inside one grant cycle.
Two Ways to Kill It
From outside, the machine can be broken in two directions, and the twentieth century supplied a clean demonstration of each. The first is capture: the state simply overrules the evidence. In the Soviet Union, Trofim Lysenko's politically convenient biology became official doctrine; genetics was outlawed as bourgeois pseudoscience, geneticists were dismissed and some imprisoned, and Soviet agriculture and biology paid for it for a generation. The lesson is not that ideology can bend science - it is that a sufficiently armed ideology does not need to bend it; it can just switch it off.
The second is subtler and was invented by the tobacco industry: manufactured doubt. You do not need to win the scientific argument; you need only to keep it visibly open - fund counter-studies, amplify outlier voices, demand a certainty science never claims to have, and let the customer conclude the jury is still out. An internal tobacco memo said it in four words in 1969: doubt is our product. The sugar industry ran the play against fat research, and fossil-fuel companies ran it against climate models whose projections - their own scientists' projections - proved accurate. The mechanism's genius is that it wears the costume of scientific virtue: what it sells looks exactly like healthy scepticism, and a public taught to "do its own research" cannot tell the difference from outside.
The pandemic stress-tested the machine in both directions at once, and the ledger is honestly mixed. The core performed at a level no previous generation could have imagined: from published viral genome to authorised vaccine in eleven months, with hundreds of millions of doses following. The institutional periphery performed worse. Officials at times performed certainty they did not have - on masks early, on school closures, on the question of the virus's origin, where a hypothesis was policed as misinformation before being later judged plausible by parts of the American government itself. The cost of performed certainty is paid in the only currency that matters here: the public's willingness to extend provisional trust next time. "Trust the science" was always a category error. Science is not a body of answers owed deference; it is a process that earns trust by surviving attack, and it forfeits exactly that trust when its spokesmen refuse to show the uncertainty.
The American Experiment of 2025-26
Both failure modes are currently being tested in the machine's largest habitat, and this site's rule is to name actors. In 2025 the second Trump administration terminated or froze more than 3,800 grants at the National Institutes of Health and the National Science Foundation, about $3bn in unspent funds. It also attempted to cap the overhead payments that keep university labs running, and proposed cutting the NIH budget by double digits and the NSF budget by more than half for 2027. Congress largely refused the cuts for 2026, and a federal appeals court ruled the funding freeze arbitrary and capricious in March 2026. The money nonetheless moved slowly and fearfully: even with its budget mostly protected, the NSF has been awarding grants at roughly a fifth of its historical rate.
The health agencies took the capture test more directly. Health secretary Robert F. Kennedy Jr. dismissed all seventeen members of the CDC's vaccine advisory committee in June 2025 and replaced them with allies; the reconstituted panel narrowed vaccine recommendations, and the childhood immunisation schedule was cut from covering seventeen diseases to eleven. A federal judge ruled the purge unlawful in March 2026, voiding several of the panel's votes. Meanwhile the measurable world supplied its verdict: measles, declared eliminated in the United States in 2000, has been reported in forty-five states in 2026, overwhelmingly among the unvaccinated, and the country is on pace to lose its elimination status. The secretary, mid-outbreak, urged families to vaccinate - a reversal his own appointees' framework had made harder to act on.
What makes this an experiment rather than an episode is the migration variable. Surveys of American researchers in 2026 read like evacuation planning; European agencies are recruiting openly, and China's research system, already comparable in scale, is hiring. History has run this experiment once at full size: Germany in 1933 dismissed its Jewish and dissident scientists, the talent moved - much of it to the United States - and German scientific leadership never returned. The projection is a reading, not a measurement, and it is flagged as one. But the mechanism it rests on is the one this page has been describing throughout: the machine is institutions plus people, the institutions can be rebuilt in a budget cycle, and the people cannot.
Why It Still Works
After all of the above, the surprising fact is not that science errs. It is that the errors do not accumulate. Lysenkoism fell. Continental drift went from crackpot to textbook. Stomach ulcers turned out to be an infection, and the physician who was laughed at drank a flask of the bacteria to prove it, and later collected the Nobel Prize. In 2023 a claimed room-temperature superconductor went from world headline to public refutation in about three weeks - laboratories worldwide dropped their work, tried to reproduce it, failed, and published the failures. That was not the system embarrassing itself. That was the system, working at its best speed.
The property doing the work is the one Robert Merton called organised scepticism, and the physicist Richard Feynman compressed into one sentence: science is the belief in the ignorance of experts. Every other institution in this Power cluster corrects itself through power - a higher court, a new government, an owner's decision. Science is the only one whose corrections are, in the end, adjudicated by something that does not care who wins; the site's outside-view page makes the same observation about mathematics, and laboratory evidence is the nearest thing the empirical world has to a proof. The process is slow, biased, careerist, and patchily funded - everything the earlier sections said. It is also the only game of its kind, and every alternative humans have tried amounts to picking answers by authority.
For the reader, the practical residue fits in four questions. Has the finding replicated, or is it one study with a press release? Was the analysis committed to before the data arrived? Who paid, and would the payer have published the opposite result? And is the claimed effect large enough to matter outside a laboratory? Most of what this site treats as established has survived those questions; where it has not, the pages say so. That discipline is borrowed directly from the machine this page has described - flaws, audit, and all - because there is nowhere else to borrow it from.
