Scott Alexander · Astral Codex Ten
Why evidence stops working: a psychiatrist's craft of doubt

One strange thing: the screening scan was 15× deadlier than the surgery

Scott Alexander wanted to donate a kidney to a stranger.

A Vox article talked him into considering it, but he was not about to take a Vox article's word for it. "Making decisions about internal organs based on a Vox article sounded like the worst idea. This was going to require more research."

So he checked. Risk of death in surgery: 3.1 in 10,000; 1.3 in 10,000 if you do not have hypertension, which he does not. The numbers held up.

Then his girlfriend at the time — whom he describes as "both brilliant and pathologically anxious, which can be a powerful combination: her zeal to justify her neuroses gave her above-genius-level ability to ferret out medical risks that doctors and journalists had missed" — spent a few hours and found something nobody had mentioned.

The screening requires a multiphase abdominal CT to check that the kidneys and their blood vessels are where they should be. That scan delivers a radiation dose of roughly 30 millisieverts. By the usual rule of thumb — one sievert equals about a 5% increase in cancer death risk — the screening exam alone carried a fatality risk of about 1 in 660.

Surgery: 1 in 10,000. The exam performed to confirm the surgery was safe: 1 in 660. Fifteen times more likely to kill him than the operation itself.

I discussed this concern with transplant doctors at UCSF and the National Kidney Foundation, who seemed very surprised to hear it, but couldn't really come up with any evidence against. My Left Kidney, 2023

He asked whether they could use a non-radioactive MRI instead. They agreed.

Every component of this mind is in that anecdote: a default procedure everybody follows, a number nobody has run, an answer an amateur can reach in a few hours, and someone actually asking — and actually getting the procedure changed.

The ending is even more characteristic. After five months of blood draws, urine tests, echocardiograms and MRIs, UCSF called to reject him — on the grounds that he had mild OCD as a child, treated at fifteen, with no recurrence in the twenty years since, during which he had become a psychiatrist running his own practice. He asked what sort of therapy would be indicated. They sounded surprised to learn that therapy came in types. He asked how frequent it needed to be. They sounded surprised to learn that therapy had frequencies, and said, you know, therapy, the thing where you talk to someone.

He went to a different hospital and donated the kidney.

Who he is

Scott Alexander is a pen name; he is Scott Siskind, a Bay Area psychiatrist. He started Slate Star Codex in 2013. The pseudonym was not an affectation. In the early 2010s he blogged under his real name, and when he interviewed for his dream psychiatry job the interviewer had Googled him and asked pointed questions about whether having a blog meant he was irresponsible and unprofessional — and there was nothing controversial on it. He did not get that job, or several others he thought were locks. He failed his entire first year of residency matching and came close to having to give up medicine altogether. So he deleted the blog, rebuilt his online presence under a pseudonym, interviewed again the next year, and got hired.

In June 2020 the New York Times was preparing a piece on him that, per its policy, would use his real name. He deleted all 1,557 posts in protest. It became a very large story — his post objecting to being doxxed was read by 513,000 people (he notes that the population of Iceland is 366,000). Six months later, on 21 January 2021, he restarted on Substack as Astral Codex Ten, and in the last line of the relaunch post printed his own name: "My name is Scott Siskind, and I love all of you so, so much."

He also quit his job and opened a clinic, Lorien Psychiatry, for uninsured patients. The reasoning is very him: he had spent years writing about why American health care costs four times what it did a generation ago, then quoted Feynman — "what I cannot create I cannot understand" — and decided to run one himself and see whether he could deliver good care at a quarter of the going rate. If it works, he said, he plans to be insufferable about it. If it doesn't, he can at least have a good conversation with the economist Alex Tabarrok about where his theories went wrong.

He defines the blog himself as a blog about ṛta — the Sanskrit word carrying truth, order, rationality, natural law and harmony at once. He is still posting at near-daily volume; the most recent post is dated 1 September 2026. About 19% of posts are paywalled, but his About page states that all important content is and always will be free. Everything this page relies on is publicly readable.

What he is wrestling with

Scott writes about everything: economics, geopolitics, neuroscience, housing, AI, the Bible, lotteries, religious miracles. Pick three posts at random and you will not find a through-line. He is entirely aware of how suspicious this is:

I often find myself trying to justify my existence; how can I write about science when I'm not a professional scientist, or philosophy when I'm not a professional philosopher, or politics when I'm not a professional policy wonk? Epistemic Minor Leagues, 2021

But there really is only one question under those 1,400 posts, and he ran into it through his job.

A psychiatrist's working day looks like this: you sit across from a person whose reasoning machinery can be demonstrated to be broken. Not "insufficiently clever", not "under-read" — a reproducible fault in the machine itself. The phobic patient fully agrees that dogs generally do not bite. He can recite the statistics to you. It does not help at all.

The awkward fact in the treatment history makes it worse. The old approach to phobia was flooding: afraid of dogs? We will lock you in a room with an enormous Rottweiler, and by the time you come out you won't be.

It sounds barbaric. More importantly, it did not work. You could spend all day in that room; the Rottweiler could fall asleep, or lick your face, or do anything else sufficient to prove it was not frightening — and you would come out more afraid of dogs than when you went in.

That is the crack. Habituation requires the experience to register as safe and okay. But whether an experience was safe is itself a perception, and it too goes through the blender. If the prior is strong enough, it overwhelms the real experience. An afternoon spent with a calm Rottweiler gets processed as close to peak horror, and becomes the prior for next time.

Scott's root problem grows out of exactly there: what if ordinary political reasoning is the same fault in the same machine?

And there is a second half, forced on him by the pandemic. Suppose you want to route around your own broken machine by deferring to institutions. Those institutions are playing a slanted game of their own. They rarely say anything false outright, but they select what to report, drop the context, and use a phrase like "no evidence" to mean two opposite things. So what he is really wrestling with is: given a self-reinforcing reasoning machine inside a systematically misleading information environment, can you know anything at all?

His answer is not "no" — that is nihilism, and it is also the rock. Nor is it "trust the science", which is self-soothing. What he has spent a decade doing is turning the problem into a craft that can be practised well or badly.

His ideas, told as the way he found them

1. Trapped priors: why more evidence makes people wronger

A note on method first. To find the load-bearing walls of this blog, I counted every internal link across the 870 posts I scraped, looking for which pieces later posts keep returning to. First place is Trapped Priors As A Basic Problem Of Rationality from March 2021, cited by 16 later posts; second place has 9. It is the foundation of this mind, and he is the one who voted for it, with his own links. (The most-cited post on the old site is his 2017 review of Surfing Uncertainty — the entry point to predictive processing. The old site was unreachable from this run's network; see the note at the end.)

His starting point does not sound like philosophy. The brain does not see first and judge second. It combines raw experience (sensations, memories) with context (priors, expectations, other senses) to produce a perception. You never notice the blending. You only ever get the finished product, and it feels exactly like raw experience.

His examples get more ordinary as they go. The chessboard illusion: two sets of pieces are the same grey, but against a darker and a lighter background you simply see one as white and the other as black, and you cannot talk yourself out of it. The McGurk effect: the same syllable, dubbed over different footage of a mouth, is heard differently — not as "my ears and eyes disagree so I'll split the difference", but as "this is exactly what I heard, with my ears, and vision had nothing to do with it". Simplest of all is the wine illusion: dye a white wine red and many people say it tastes like red wine. The placebo effect is the same shape — raw experience is the nerve firing just as many pain signals as before, context is a doctor promising relief, and the product is that you hurt less.

The cognitive version of this is just normal Bayesian reasoning. Your friend says she saw a coyote on the way to work; you believe her. She says she saw a polar bear; you doubt it. Same raw evidence, different prior, different conclusion. Nothing has gone wrong yet.

Then he lets it slide one step. You are a zealous Democrat. A friend makes a plausible-sounding argument for a Democratic position and you accept it; she makes an equally plausible one for a Republican position and you are sceptical. Still fine — that is a prior doing its job. And in principle it is self-correcting: give it enough evidence and you have to come round eventually.

"But in fact many political zealots never accept reality."

To see why, go back to the Rottweiler.

The crucial step is simple enough that reading it produces an audible noise. Habituation requires an experience that registers as safe. But safety is also a perception, and it also goes through the blender. The raw evidence — the Rottweiler sitting calmly, wagging its tail — looks promising. The context is a very strong prior that dogs are terrifying. Strong enough, and it overwhelms the real experience. Result: the Rottweiler was terrifying. So any update you make now runs toward dogs being frightening, not away from it.

Your prior fear of dogs determines your present experience, which in turn becomes the deranged prior for future encounters. Trapped Priors As A Basic Problem Of Rationality, 2021

He adds a precise piece of phenomenology. Less self-aware patients find the prior colouring every part of the encounter: a joyful pounce for a head-pat is read as a vicious lunge, a whine at not being played with as a murderous growl. They leave saying "the dog came this close to attacking me, I knew all dogs were dangerous." More self-aware patients say "I know deep down that dogs aren't going to hurt me, I just know that whenever I'm with a dog I'm going to have a panic attack and hate it and be miserable the whole time." Then they go in, have a panic attack, are miserable, and the link between dogs and misery is welded a little tighter.

Where does emotion enter? He follows van der Bergh and colleagues: when an experience is too intolerable, the brain narrows the bandwidth of the raw-experience channel to protect you from the traumatic feelings. The cost is that context now almost entirely determines what you perceive. This is why trauma victims' accounts of their traumas are often oddly short, undetailed and to the point. And it is why strong feeling entrenches bias: a trapped prior is a purely cognitive phenomenon, but emotion creates the perfect storm for it.

He is careful to insist this is not just "emotional bias" — an emotionless threat-assessment program using the same Bayesian reasoning people use would have the same problem. The political evidence is hard: more scientifically literate people are more partisan on contested scientific questions. If they were merely biased, each new fact should nudge them toward the truth. That is not what happens.

they start out wrong, and each new fact they learn, each unit of effort they put into becoming more scientifically educated, just makes them wronger. Trapped Priors As A Basic Problem Of Rationality, 2021

He digs out the 1979 study. Partisans read paired studies on capital punishment and rated the methodology from −8 to +8. The psychologists had designed the two to be about equally good and even swapped their conclusions between subjects to average out any residual difference. Conservatives rated the pro-punishment study about +2 and the anti-execution one about −2; liberals produced a slightly smaller mirror image. Both groups reported that the evidence had confirmed their position, and described themselves as more certain than before. The more detail they got, the stronger the belief. He reaches for a clinical word: this pattern is pathognomonic of a trapped prior. Then: these people are doomed.

The model's most useful payoff is that it explains a maddening feature of daily life: dog-whistle readings. Once your prior about a side is trapped, anything neutral they say can be heard "in context" as something monstrous. Republicans talking about "law and order" means locking up as many black people as possible; Democrats talking about "gay rights" means destroying the family so the state can take your children. He has met both: people convinced that no pro-life conservative really cares about fetuses and just wants to punish women, and people convinced that no pro-lockdown liberal really cares about COVID deaths and just enjoys making people wear a badge of submission.

"Once you're at the point where all these things sound plausible, you are doomed. You can get a piece of evidence as neutral as 'there's a deadly pandemic, so those people think you should wear a mask' and convert it into 'they're trying to create an authoritarian dictatorship'. And if someone calls you on it, you'll just tell them they need to look at it in context."

He flags a hole in his own model. In the section on emotion he notes that under most reasonable weighting functions, even a strong prior plus some evidence of a friendly dog ought to make the perception slightly less scary than the prior — and iterated long enough, that should shift the prior. It doesn't. "I don't know why this doesn't happen in real life, beyond a general sense that whatever weighting function we use isn't perfectly Bayesian… I realize this is a weakness of this model and something that needs further study." Hold that thought for section six.

Is there hope? His proposals are borrowed from the clinic: gradual exposure rather than flooding; getting patients to re-tell the trauma in exhaustive multimodal detail, by drawing or acting it out — which he suspects is a way of upweighting the raw-experience channel. The political analogue may be Sloman and Fernbach's finding that the only task which moderated partisan extremity was requiring people to give a precise mechanistic account of how their preferred policy would actually work. His other candidates — psychedelics, which reduce the weight given to priors, and meditation — he flags with their risks and his uncertainty.

The last thing worth noting is how long this model lives in his hands. In 2025 he wrote an essay about medieval ultramarine pigment, AI-generated images, and whether we can still feel wonder once everything is cheap — apparently nothing to do with Bayesian brains. Reaching Chesterton's claim that the thousandth sunset is as beautiful as the first, he stops and asks: what if Chesterton was not being poetic, but describing his experience accurately?

This sort of thing would make complete neurological sense - it's just an increase in the precision of sensory evidence relative to top-down priors. The Colors Of Her Coat, 2025

Same machine. Four years later, used to explain sainthood.

2. Heuristics that almost always work: how experts decay into a rock

The first idea is about how the machine in your own head breaks. The second answers the obvious reply: fine, then outsource your judgement to experts.

In February 2022 he wrote an essay containing no data at all, just six short stories. You have met the first: the security guard who can be losslessly replaced by a rock reading THERE ARE NO ROBBERS.

Then the doctor. Patients come to her saying "my back hurts" or "my stomach feels weird". She inspects, palpates, percusses, runs tests, and says take two aspirin and call me in a week. It always improves; nobody ever calls. So she gets sloppy — she inspects but does not palpate, skips the tests, just says it'll get better on its own. And she is always right. If she is very lucky nothing bad ever happens; more likely two or three of her patients have cancer and she misses it. But those people die, and everyone else remembers her as such a nice doctor, so reassuring, who never poked and prodded them with needles like everyone else. She could be profitably replaced with a rock saying IT'S NOTHING, TAKE TWO ASPIRIN AND WAIT FOR IT TO GO AWAY.

Then the futurist, who denies every predicted upheaval and therefore tops the Brier score rankings, and could be replaced by a rock reading NOTHING EVER CHANGES OR IS INTERESTING. Then the skeptic, who debunks everything; she used to research each phenomenon first, but after a hundred rounds she skipped steps one and two and her algorithm became "if anyone says something weird, insult them viciously". She has a flawless record — she condemned hydroxychloroquine first, and ivermectin first. "(shame about the time she condemned fluvoxamine equally viciously, though)"

Here he lands the best paragraph in the piece:

This is a great rock. You should cherish this rock. […] But it is a Protestant god. It does not need priests. If someone sets themselves up as a priest of the rock, you should politely tell them that they are not adding any value, and you prefer your rocks un-intermediated. Heuristics That Almost Always Work, 2022

Then he turns the joke into a genuinely sinister parable. A queen rules a volcanic island. The wise men determine the volcano has a flat 1/1000 chance of erupting in any year, uncorrelated with the last. There are faint signs legible to the learned — a slight change in the colour of the lava, an imperceptible shift in the smell of the sulphur — but nothing obvious until it is too late.

She founds a Learned Society Of Vulcanologists. Unbeknownst to her there are two kinds. Honest ones, who genuinely try to read the signs. And the Cult Of The Rock, an evil sect who gain diabolical knowledge by communing in secret with a rock reading THE VOLCANO IS NOT ERUPTING.

Every so often an honest vulcanologist thinks the lava looks odd and says so. The queen panics and asks everyone. The honest ones say: it's a hard question, the lava is weird today but it's always weird in some way, this volcano rarely erupts but this might be the exception. The cultists check their rock and say: don't worry, it isn't erupting. The volcano doesn't erupt. The queen punishes the trigger-happy one who raised the false alarm, grumbles at the useless ones who wouldn't commit, and promotes the confident cultists who predicted everything correctly.

Five hundred years pass, the institutions selecting for cultists the whole way. When the lava finally does look strange, the queen's advisors are 100% cultists. The sulphur smells different; she asks if they are sure; they double-check the rock and say yes. The earth begins to shake; she asks one last time; they get tiny magnifying glasses and look at the rock as closely as they can, and it still says THE VOLCANO IS NOT ERUPTING.

"Then the volcano erupted and everyone died. The end."

Only then does he state the payload. There are two harms here, and the second is the lethal one:

Maybe the heuristic produces a prior of 99.9% that the thing won't happen in general. But then you consult a bunch of experts, who all claim they have additional evidence that the thing won't happen, and you raise your probability to 99.999%. But actually the experts were just using the same heuristic you were, and you should have stayed at 99.9%. False consensus via information cascade! Heuristics That Almost Always Work, 2022

Put plainly: expert opinion is not an independent source of evidence, and you are multiplying it as though it were. "This new invention won't change everything. This emerging disease won't become a global pandemic. This conspiracy theory is dumb. This outsider hasn't disproven the experts. This new drug won't work. This dark horse candidate won't win the election. This potential threat won't destroy the world." All almost always true — and precisely because they are almost always true, they tempt us to be more certain of each than we should be.

His ending is sly and honest at once: next time someone mocks rationalists for burning brain cycles on "obvious" questions, check how they make their own decisions. 99.9% of the time it is Heuristics That Almost Always Work. "(but make sure to watch for the other 0.1%; those are the people you learn from!)"

3. Bounded distrust: lying is a game with rules

If the first two ideas are diagnosis, the third is his prescription — and the one most often quoted and most often misread.

He opens with a question designed to make liberal readers uncomfortable. You are a liberal who does not trust FOX News. You are waiting for a plane, half-watching the TV at the gate. It is FOX, and they say a shooter just killed twenty people at Yankee Stadium; there is live footage of people running and screaming. Do you believe it?

"I'm a liberal who doesn't trust FOX News, and sure, I believe it."

Because the level on which FOX is bad is not that level. They would not use deepfakes or staged actors and call it live footage; that goes way beyond anything FOX has done. He raises the stakes: FOX says police have arrested a suspect, a Saudi immigrant named Abdullah Abdul, and shows footage of the press conference. Still yes — faking a press conference, or dubbing over one so a different name is heard, is far worse than anything he has heard of them doing.

He deliberately closes off the easy objection: FOX being biased is irrelevant here. You could argue they want to stoke fear of Islamic terrorism, or that they would change a white gun owner into a Saudi terrorist. But those sound like deranged conspiracy theories; even FOX's worst enemies do not accuse it of that.

The point is, there are rules to the "being a biased media source" game. There are lines you can cross, and all that will happen is a bunch of people who complain about you all the time anyway will complain about you more. And there are other lines you don't cross, or else you'll be the center of a giant scandal and maybe get shut down. I don't want to claim those lines are objectively reasonable. But we all know where they are. Bounded Distrust, 2022

And then — this is why the essay is his and not somebody else's — he immediately turns the same knife on his own side, with the example liberals least want raised.

First he builds the mirror case for conservatives. A Washington Post piece claimed Lincoln was a friend of Karl Marx who admired his socialist theories, and therefore that attacking socialism is un-American. Scott read the rebuttal and found it far more convincing: almost every sentence in the original is literally true but organised to mislead. "That might be because Lincoln was regularly reading Karl Marx" — with the hedge phrased so you read past it. "It's nearly guaranteed that, in the 1850s, Lincoln was regularly reading Marx" — the evidence being that Lincoln read a newspaper Marx had published in. Marx sent letters to Lincoln — without mentioning that presidents receive thousands of letters and there is no evidence Lincoln read this one. A US ambassador told Marx's group that Lincoln appreciated them — without mentioning it was a form letter.

So how should a conservative accept "the 2020 election was not rigged" from that same paper? His reasoning is operational. The Marx piece got almost no scrutiny — a few smart people noticed it was fake, three or four wrote small editorials, and nobody cared. The 2020 election got enormous scrutiny from every major institution. The historians who praised the Marx article were all explicitly socialist and all studying other periods; statisticians and election security experts, speaking inside their actual expertise, said clearly that the election was fair. And he adds the kicker: the Marx piece was a cutesy human interest story, and everybody knows cutesy human interest stories are always false.

The third example shows how far he is willing to go. Swedish researchers collecting crime statistics found, incidentally, that immigrants accounted for a disproportionate share of some crimes. The Swedish establishment brought scientific misconduct charges: first, for not getting permission to include ethnicity statistics, though the data were publicly available; second, for being unable to justify how their research would "reduce exclusion and improve integration". Scott's verdict is unhedged: the charges are probably true on their own terms, but no researcher who found that immigrants were great would have their technicalities scrutinised this way, and the permission system partly evolved in order to crush people in exactly this situation. "So I think the outrage is justified, this is exactly what people mean when they accuse experts of being biased, and those accusations are completely true."

And then the question that turns it around: have you ever heard an expert say, in so many words, that immigrants to Sweden definitely don't commit more crime than natives?

No. The academic establishment will bury inconvenient facts, but the experts will not lie directly. They do not go on television and say the science has spoken and there is strong evidence against immigrant criminality. They do not invoke a "strong scientific consensus". They occasionally punish people who raise it, but they do not call them science deniers.

Which is exactly why you can still trust experts about some things. The reason there is no petition signed by every respectable criminologist on Swedish immigrant crime, and there is one from climatologists on anthropogenic warming, is not that experts are unbiased — it is that they are not quite biased enough to sign a transparently false statement. So the existence of that climate petition really is evidence for the conclusion, even to someone who is not a climatologist.

He knows precisely how dangerous this is, and says so:

My answer - which I don't think is an obvious or easy answer, it's a bold claim that could be wrong, is "I think I have a good sense of the dynamics here, how far people will bend the truth, and what it looks like when they do". I realize this is playing with fire. But listening to experts is a powerful enough hack for finding the truth that it's worth going pretty far to try to rescue it. Bounded Distrust, 2022

He later pushed the argument to its least popular position, in The Media Very Rarely Lies. He went through the front page of Infowars item by item and could barely find anything wholly fabricated — the vaccine adverse-event data are real, just stripped of context — and then held the same ruler up to the New York Times and measured the same disease. Even the worst purveyors of misinformation rarely state false facts; they state true things without enough context. And nobody will ever agree on which context is necessary.

Finally, the most human version of the idea, from the end of the ivermectin piece. He asks why he himself believes the immunologists, given that he is not one. The answer is not "because I am rational". It is:

I think of myself as a temporarily-embarrassed immunologist. I don't know all the interleukins. But I would like to believe that if I really wanted, either I or at least people I know and trust could learn immunology to a standard where we could double-check the work of the vaccine scientists. Ivermectin: Much More Than You Wanted To Know, 2021

(He is borrowing Steinbeck's line that socialism never took root in America because even the poor see themselves as temporarily-embarrassed millionaires.)

Now imagine someone whose social filter bubble excludes scientists entirely. To them experts are not colleagues but hostile aliens: using words like "quantum memetic plague" that are not merely beyond your understanding but unfairly beyond it, nothing you or anyone you trust could ever double-check; who got their last several predictions wrong — remember "don't wear a face mask"; and 95% of whom vote one way. Scott writes an entire alien-invasion fable to inhabit that view, then delivers the line: "99% of hostile aliens agree: vaccines are right for you!"

His conclusion is not that those people are stupid. It is that if we want people to take vaccines, we have to make the scientific establishment stop feeling like an enclave of hostile aliens to half the country. "But I don't really know how to do that."

How he thinks

What transfers from reading him is not the conclusions. It is a handful of moves you can copy.

1. Turn "I read the literature" into a procedure with steps. His signature form is the "Much More Than You Wanted To Know" piece. Ivermectin is the specimen: he took the 29 studies listed on a pro-ivermectin site, read each one, threw out 13 he did not trust, then threw out 5 more that the epidemiologist Gideon Meyerowitz-Katz disliked even though he had no objection himself — in order to end up with a list both sides agreed was decent — leaving 11. His closing section names the new toolkit explicitly: check for fraud (the Carlisle-Stouffer-Fisher method, which looks at whether a paper's Table 1 p-values are suspiciously tidy; GRIM, which catches impossible means like "27% of 10 patients recovered"), and check for coding errors — he admits an early draft of his own post reported a strong effect that turned out to come from one survey respondent claiming 99,999 autistic children.

I want a world where "I did a study, but I can't show you the data" should be taken as seriously as "I determined P = NP, but I can't show you the proof." Ivermectin: Much More Than You Wanted To Know, 2021

2. Put a number on your confidence, then let yourself move it. The ivermectin post ends in a list of percentages: ivermectin does not meaningfully reduce COVID mortality absent comorbid parasites — 85–90% confidence; parasitic worms are a significant confounder in some studies — 50% confidence. A year later he publicly raised the first to 95% and cut the worm claim to 35%, with reasons. He also adds the crucial footnote: "I'm not an expert on this, you shouldn't care about my exact probability, and I'm only mentioning it to communicate clearly and try to hold myself accountable."

3. Write the argument as a detective story, and actually make the reader answer first. Midway through the ivermectin piece he stages the worst moment honestly: Mahmud et al is a good study and it got p = 0.003 in favour of ivermectin. "You can take the blue pill, and stay a decent respectable member of society. Or you can take the horse dewormer pill, and see where you end up." Then he refuses to answer — "you won't always have me to answer questions like this, and it might be morally edifying to observe your thought process" — and prints five lines of ellipses.

Turn the page, and the section is headed The Synthesis, and the first line is: "Hopefully you learned something interesting about yourself there. But my answer is: worms!"

The most impressive studies all came from places teeming with worms — Bangladesh, East India, Colombia. One of the commonest COVID treatments is corticosteroids, an immunosuppressant; the immune responses it suppresses are exactly the ones that keep Strongyloides roundworm in check. Switch them off even briefly and the worms multiply out of control, and people quite often die. Ivermectin is a deworming drug. So in worm-heavy regions it really did cut mortality — it just wasn't treating COVID. He says the reason the theory feels right to him is that it is the most troll-ish possible solution: everybody was wrong. The people who called it a miracle drug, and the people who dismissed every study because they Freaking Love Science, both.

4. Correct compulsively, starting with yourself. His 2025 essay If It's Worth Your Time To Lie, It's Worth My Time To Correct It gives a mechanical reason: if one side lies to make its arguments 5% stronger, it compounds; the other side has to do the same or fall behind; then that becomes the baseline, so why not 10%, why not 20%. Having made the point with an example about a bar fight framed as a hate crime, he immediately writes this parenthesis:

(sorry - I've genuinely heard people say he was going to put minorities in camps, but I'm not sure they specified all minorities, and I don't think they ever said they would get murdered there. Would you have let me get away with that exaggeration?) If It's Worth Your Time To Lie, It's Worth My Time To Correct It, 2025

5. Take the theory and go start a company with it. Years of writing about American cost disease, then Feynman's "what I cannot create I cannot understand", then Lorien Psychiatry — a real practice for uninsured patients, aiming to deliver good psychiatric care at a quarter of the going price. The theory gets acceptance-tested by an actual clinic.

6. Keep a public list of your errors. He maintains a standing Mistakes page: as of this review it holds 61 numbered errors, from October 2013 to August 2025, in reverse chronological order. The rule is stated at the top:

I don't promise never to make mistakes. But if I get something significantly wrong, I'll try to put it here as an acknowledgement and an aid for anyone trying to assess my credibility later. Mistakes, continuously updated

It excludes typos, failed predictions and errors in link posts, and covers only cases where he was fundamentally wrong about a major part of a post. This is not a pose: the subject of the next section is on that list.

Where he might be wrong

The most substantial rebuttal has a name, and he published it himself. Alexandros Marinos, an entrepreneur and long-time reader, wrote 21 posts rebutting the ivermectin piece point by point. In February 2023 Scott published a long reply working through them.

What he concedes is serious. Of Biber et al, an Israeli study, he had written in the original that the primary endpoint appeared to have been buried because it was negative, and sneered that the authors' regression was "the kind of thing you do when your primary outcome is boring and you're angry". In fact the primary outcome was positive and listed exactly where it should be, in the abstract and results. He writes: "On the one where I was wrong, I was egregiously wrong, and I apologize to the study authors and to you." He also concedes that his informal meta-analysis used the wrong statistical test — a DerSimonian-Laird test was correct, and using it makes the numbers more favourable to ivermectin; that he got a claim about Cadegiani wrong; and that he over-weighted worms and under-weighted publication bias, dropping that confidence from 50% to 35%.

Nobody can be right all the time, but I try not to be simultaneously wrong and smug. I failed at that here and need to do better at Principle of Charity […] Realistically, this project was outside my expertise and competence level. […] I give myself a C+ for results but an A for effort. Response To Alexandros Contra Me On Ivermectin, 2023

But the same exchange contains an uglier fact, also in his own words. In the second paragraph he admits: "I'll be honest - I also didn't want to read a 21 part argument. I would say I have read about half of his posts, and am mostly responding to the summary." A man whose entire method is built on reading all 29 studies one by one read half of the most systematic critique ever aimed at him. That is not my inference; it is his sentence. The persuasive force of his method rests on exhaustiveness, and exhaustiveness is the first thing to fail when he is the one under scrutiny.

The second weakness he marked inside the theory himself. The trapped priors model has an admitted hole: under most reasonable weighting functions, even a strong prior plus a little contrary evidence should soften the perception slightly, and enough iterations should shift it. It doesn't. "I don't know why this doesn't happen in real life… I realize this is a weakness of this model and something that needs further study." This is not a detail — the entire explanatory weight of the model rests on that unfilled mechanism.

And the sharpest criticism of it is the one he wrote himself, in the last line of the same essay:

Ironically, my prior on this theory is trapped - everything I read makes me more and more convinced it is true and important. I look forward to getting outside perspectives. Trapped Priors As A Basic Problem Of Rationality, 2021

You can read that as charming self-deprecation. You can also read it as a real problem: a theory that can explain every counterexample as the other person's prior being trapped is very hard to falsify.

The third is structural, in bounded distrust, and he presses on the bruise without solving it. The whole method depends on a skill whose rules cannot be written down — knowing where a source will bend the truth and how far. He says "I think I have a good sense of the dynamics here", and immediately grants that this is playing with fire. The trouble is that "I have developed judgement" and "my prior is trapped, so I can always explain why the experts are trustworthy this time" are indistinguishable from the inside. His own theory of trapped priors predicts he would not be able to tell them apart. And in the one case where Marinos really caught him, the failure was at exactly this layer: he believed an accusation relayed by someone he trusted (Meyerowitz-Katz), who had not actually made it.

Put the three together and a shape emerges worth naming: his method is sharpest pointed outward and bluntest pointed inward. Twenty-nine studies, dismantled one at a time; twenty-one posts about him, half read. That is not a moral charge — he did publish the reply, and he did enter the error on a list that now runs to 61 items — but it is the direction in which this method systematically tilts, and worth carrying with you as you read him.

Finally, what this page could not read, stated plainly. Sharper external criticism exists: Noah Smith published a response to Scott's argument about Jewish achievement on his own blog, and RationalWiki hosts a dedicated critical entry on him. This run's network policy permitted only a small set of domains; both of those, and his own old site slatestarcodex.com, were unreachable. I therefore report only the bibliographic fact that these criticisms exist. I did not read them, and I will not paraphrase arguments on their behalf. By the same rule, his most famous old-site essays — Meditations On Moloch, I Can Tolerate Anything Except The Outgroup — appear here only as posts he himself keeps linking back to. Not one word of them is quoted.

Sources

• Trapped Priors As A Basic Problem Of Rationality (2021)The trapped prior. From why flooding fails as a phobia treatment to why more scientifically literate people are more partisan. The most back-linked post on the site.
• Heuristics That Almost Always Work (2022)The guard, the doctor, the futurist, the skeptic, and the Cult Of The Rock on its volcanic island. Why expert opinion is not independent evidence.
• Bounded Distrust (2022)FOX and the shooting, the Washington Post and Lincoln, Swedish criminology, and the Soviet harvest.
• Ivermectin: Much More Than You Wanted To Know (2021)Twenty-nine studies taken apart one by one, ending in worms. Also the source of the temporarily-embarrassed immunologist and the hostile aliens fable.
• Response To Alexandros Contra Me On Ivermectin (2023)His reply to Alexandros Marinos's 21-post rebuttal: what he conceded, what he didn't, and the admission that he read half of it.
• Still Alive (2021)The relaunch post after deleting 1,557 essays. Why he used a pseudonym, why he burned the blog, and why he ended up printing his own name.
• My Left Kidney (2023)The whole donation process: the arithmetic showing the screening CT was 15× deadlier than the surgery, and the call rejecting him over childhood OCD.

On sourcing: this page draws on the public text of Astral Codex Ten — 870 posts scraped and parsed (~3.4M words), with internal back-link centrality computed over that corpus; 12 posts read in full, spanning January 2021 to June 2025. The 11-post introductory list he recommends on his own About page lives entirely on the old site, slatestarcodex.com, which was unreachable under this run's network policy; no text from the old site is quoted here.