Nassim Nicholas Taleb on the Pandemic
What this covers
Nassim Nicholas Taleb speaks with Russ Roberts on the structure of pandemic risk and how it should reshape decision-making. The argument rests on a single pivot: pandemics are fat-tailed phenomena—multiplicative risks that can spiral out of control—and therefore demand a different logic than the thin-tailed risks (car accidents, falls) that conventional decision science handles. Taleb frames the core tension as one between individual-level and systemic-level thinking. A clinician's expertise in treating single patients does not transfer to understanding collective dynamics, just as the skill to build a roulette wheel does not make one expert at gambling. This scale transformation means that risk of ruin (where survival itself is the precondition) cannot be weighed against economic costs the way ordinary trade-offs can be.
The conversation traces implications across medicine, economics, history, and ethics. Taleb argues that pre-modern Eurasia embedded quarantine into its economic systems so effectively that the Marseilles plague of 1720 occurred precisely when a mayor abandoned those rules, while modern global connectivity has made super-spreader events far more extreme—a single case now reaching hundreds of countries in days. He contends that the right pandemic protocol is not full lockdown but targeted quarantine, super-spreader elimination, and universal masking, since when two people wear masks the infection probability compounds quadratically rather than adding linearly. On the contentious side, Taleb rejects utilitarian arguments for letting the elderly die to preserve economic value, treating such trade-offs as both culturally offensive and logically invalid under tail risk, where harm boundaries remain unknowable. He also claims that standard decision science literature is "largely bogus" for pandemics because it ignores fat-tailedness and scaling, and that forecasters using averages and single-point predictions are practicing something other than science when dealing with distributions where one observation dominates the mean.
Taleb argues that pandemics are the fattest-tailed of all risks, so they must be treated under a precautionary protocol (kill them in the egg via masks, quarantine, and eliminating super-spreaders) rather than judged by single-point forecasts or naive cost-benefit trade-offs, because individual risk and systemic risk do not scale the same way.
- Contagious risk is multiplicative, so individual caution has nonlinear systemic effects that conventional decision science misses
- Under fat tails, single-point forecasts and averages are meaningless; science is about understanding properties and asymmetry, not predicting events
- Some risks (ruin, geronticide) admit no trade-off because survival is a precondition for all other gains
Decision science and epidemiology are incompetent because they ignore fat-tailedness, scaling, and nonlinearity, mislabeling rational precaution as irrational.
- The decision science and rationality literature is largely bogus for pandemics because it omits fat-tailedness and scaling, so that people who do the right thing (acting cautiously against multiplicative risk) are mislabeled as irrational.
“the fat-tailedness and the scaling are absent from the Decision Science literature, which effectively makes people doing the right thing look like they're irrational”
- Understanding medicine at the level of a single human does not let you understand the collective, because that is a scale transformation; just as the carpenter who built the roulette wheel is not the expert on how to gamble, a clinician's individual-level expertise does not confer authority on population-level epidemiological dynamics.
“people think that understanding medicine, which is understanding a human, single human, allows you to understand the collective. That's another skill transformation.”
- Many claims attributed to statistical models are invalid because the models' assumptions are violated under fat tails: linear regression requires the Gauss-Markov theorem to hold, and psychologists' use of correlation fails when effects are nonlinear, so violating these foundational mathematical rules means the work is no longer science.
“you cannot do linear regression if you don't satisfy the Gauss-Markov theorem”
- People's built-in instinct (e.g. not drinking water of uncertain quality) is vastly superior to the formal reasoning of the CDC, decision-science professors, and the UK Nudge group, who are incompetent at basic things an ordinary grandmother gets right.
“people's built-in instinct is vastly superior to what the CDC had... All these people are completely incompetent when it comes to basic things that your grandmother gets.”
- For fat-tailed variables, single-point forecasts are unscientific because the average is observed only rarely: in a skewed log-normal distribution with mean just above 7, 85% of observations fall below the average and 50% fall below 1, so you cannot work with averages or judge a forecaster by whether the realized outcome matched the predicted mean.
“The title is "On Single-Point Forecasts for Fat-Tailed Variables."... you only observe the average 15% of the time... 85% of the observation fall below the average.”
Pandemics are multiplicative fat-tailed risks fundamentally different from thin-tailed accidents, spiraling exponentially and demanding early suppression.
- Pandemics are the fattest-tailed of all risks because they are multiplicative and can spiral out of control, representing an existential risk; therefore they must be stopped in the egg while it is still cheap to do so.
“in my opinion, they are the fattest tails, and fat tails means that these things could really go out of control”
- Multiplicative risks like pandemics cannot be compared to thin-tailed risks like drowning in a bathtub or falling from a ladder, because one person's death in a pandemic causes others' deaths whereas independent accidents follow a Chernoff bound and do not compound; if a billion people died, it would almost certainly be a pandemic, not isolated accidents.
“If I tell you a billion people died, you know they did not die drowning in their bathtub in any given year”
- Pandemics implode naturally—like the Plague or Spanish Flu—not primarily through herd immunity but through reduced connectivity; the goal of measures is not to bring cases to zero but to keep them from overwhelming and collapsing the healthcare infrastructure.
“I don't believe in herd immunity as much as in reduced connectivity.”
- Modern global connectivity makes multiplicative phenomena vastly more extreme, producing winner-take-all effects in both economics (Google, Pavarotti) and disease, where a single virus can take over the planet via super-spreader events that redistribute infection to hundreds of countries, unlike the slow 30-km-a-day spread of the Silk Road.
“we have more connectivity today than we did before, so therefore things that are multiplicated will be vastly more multiplicated today”
Mask-wearing compounds nonlinearly; halving viral load cuts infection risk far more than half, and dual masks reduce risk exponentially not additively.
- Reducing viral load by half does not necessarily reduce the probability of infection by half; because of the S-curve (convex at low doses), halving viral load can reduce infection probability by as much as 99%, a nonlinearity that mask analysts who only looked at viral load reduction missed.
“if I reduce the viral load by half, I don't necessarily decrease the probability of infection by half. I may decrease it by 99%”
- When two people both wear masks, the risk reduction is not the single-mask protection P but P-squared, because the protections compound: if each mask leaves a 50% infection probability, two masks together reduce it to 25%.
“if I wear a mask and you wear a mask, we're not reducing the risk by P, the protection, by a mask, but by P-squared”
- The statistical observation that Taiwan and Hong Kong were largely spared despite heavy traffic from Wuhan provided an early hint that masks work to control transmission.
“why is it that Taiwan, that has a lot of traffic from Wubai-- Wuhan--was completely spared?”
You take precautionary action when harm is small and benefit massive, without waiting for clinical trials, following payoff asymmetry not probabilistic evidence.
- Rigorous scientific decision-making must incorporate payoff asymmetry: when the potential harm from an action (wearing a parachute or a mask) is vastly smaller than its potential benefit, you take the action regardless of the absence of clinical studies, just as you'd wear a parachute jumping from 35,000 feet without needing a controlled trial.
“you understand asymmetry... the harm you're going to get from wearing a parachute is vastly inferior, potentially, than the benefits you may get for it”
- The central idea of the Incerto is that uncertainty in a system makes decision-making much easier, not harder: if you are uncertain about water quality you simply don't drink it, if uncertain about a pilot's competence you get off the plane, and if you don't know whether masks work you wear them anyway.
“when you have uncertainty in a system, it makes your decision-making much, much easier”
- The entire point of the precautionary principle is captured by an old trader's rule—take all the risks you want, but make sure you're in tomorrow—which is how Goldman Sachs survived 161 years; standard economics is flawed because it ignores this absorbing barrier and the ergodicity it implies.
“'Take all the risks you want, but make sure you're in tomorrow.'”
- Science is not about evidence or forecasting single events but about understanding the properties of a process; it can be scientific to describe a process and recommend an action (like wearing masks) even in the absence of direct empirical evidence, because absence of evidence is not evidence of absence.
“Science is not about evidence. Science is about properties.”
Individual precaution is a systemic duty because contagious risk scales nonlinearly from person to population, making personal restraint a collective moral imperative.
- Wearing a mask is a moral imperative because you wear it not primarily to protect yourself but to prevent infecting others who could each infect many more; the small personal cost yields a large systemic benefit.
“I wear a mask not for myself, but because a person I'm going to infect will infect maybe 10 more”
- Contagious disease risk does not scale linearly from individual to collective risk, so failing to behave conservatively dramatically increases risk for everyone else; this makes individual precaution a systemic duty even when it seems selfishly unnecessary.
“if you don't behave conservatively, you will increase that risk dramatically because the collective risk doesn't scale linearly from individual risk”
- There are trade-offs you simply do not make: utilitarian calculations about letting the elderly die to save economic value are offensive and impermissible in classical Mediterranean societies that honored elders, and even on purely utilitarian grounds geronticide is not a real trade-off because under tail risk you don't know where the harm will stop.
“There are trade-offs you don't make.”
Systematic quarantine of all international arrivals, not origin-specific bans, and decentralized local responses outperform centralized pandemic control.
- Restricting quarantine only to people coming from China was foolish because the disease could and did arrive from elsewhere; the correct response is systematic quarantine of international arrivals regardless of origin.
“the fact of having quarantine for people coming from China is foolish, not knowing that they can come from somewhere else”
- The correct pandemic protocol is not full lockdown but a mechanism for quarantine plus identifying and eliminating super-spreader settings (subways, elevators, large gatherings, political rallies); once masks are added and we learn that open air is safer, activities like games can resume with universal masking.
“You don't have to lock down things... Super spreaders are subways, elevators, and big gatherings like the Trump convention.”
- Although historical quarantine was forty days (the root of the word), the modern two-week quarantine works well enough, because any meaningful reduction of connectivity is very effective at controlling spread.
“two weeks works. I mean, two weeks is good enough... whatever reduction of connectivity is very good.”
- Pandemic response is a victory of localism: just as some medieval towns fared far better than others during the Plague by enforcing their own quarantines, US states and cities can block or check travelers from hotspots, and decentralized responses outperform centralized ones because outcomes vary sharply by locality.
“This is the victory of localism because it showed that, just like the Plague. The Plague was a victory of localism. Some towns fared a lot better than others.”
Long-term COVID harms remain unknown; young people face unquantifiable risks, and systemic survival requires treating identified fat-tailed risks distinctly.
- Risk of variation (medium-sized fluctuations like car accidents) is fundamentally different from risk of ruin at a systemic level: society must survive, so once you identify which risks are systemic and existential, you can avoid those and take many more risks elsewhere.
“the central idea that risks of variation is extremely different from risk of ruin, particularly at a systemic level”
- The rationale for devoting resources to the elderly is ergodic: today's young will themselves become the elderly, so caring for the old is implicitly a promise to the young that they too will be cared for, embodying an intergenerational gratitude (as when the Greeks honored the mules that built the Parthenon) that crude utilitarianism misses.
“you got to think ergodically--will be the elderly... by taking care of the elderly, 'We're going to take care of you.'”
- Five months in, we know almost nothing about COVID-19, including its morbidity: survivors may emerge with damaged lungs or autoimmune effects, and even young people may suffer lasting harm, so its long-term consequences cannot be ruled out and it should not be dismissed as harmless to the young.
“We know nothing about morbidity... A lot of young people have damaged lungs. We don't know.”
- As a precautionary heuristic, you ignore any disease until it has killed more than a thousand people (in today's population-equivalent), but the moment it crosses that threshold you treat it as a fat-tailed monster that could kill a billion and must be killed in the egg—because all such pandemics should be treated identically regardless of their apparent current lethality.
“It means that you don't have to worry about anything that hasn't killed yet a thousand people... The minute it kills--if it's in the category that it has killed more than a thousand people, then you have to worry about that.”
Most economic damage flows from rational fear, not shutdowns; thin-margin businesses collapse on modest revenue loss, weakening the shutdown-harm trade-off argument.
- Based on accumulating data—such as cases where infected hairdressers exposed clients for 15 minutes with zero contagion—wearing two masks (with an N95) plus other measures makes air travel acceptably safe even for visiting elderly relatives, and brief unmasking to drink while seated far from others poses negligible risk because exposure is nonlinear in duration.
“if you have two masks... you should be okay traveling”
- Most of the economic contraction is caused by the pandemic itself, not by government shutdowns: rational, risk-averse people stop going to restaurants, bars, and planes out of fear and uncertainty, and since many businesses operate on thin margins a mere 20% drop in revenue bankrupts them—so the trade-off between government shutdown and economic harm is far smaller than claimed.
“It's not fighting the pandemic that's costing money. It's the pandemic.”
- Society is antifragile: like the fact that another plane crash becomes less likely after one crashes, we learn from this pandemic and reduce the odds of suffering the truly big one (e.g. antibiotic-resistant bacteria), while the economy benefits from stressors as remote work, online conferences, and cheaper global education emerge as durable adaptations.
“We're antifragile. I mean, we learned from this. It's like when a plane crashes, the probability another plane crashes drops.”
- In sufficiently fat-tailed distributions the average is dominated by a single observation—out of a million observations, one will determine the mean and only that one will lie above the average—just as 1% of Americans own 50% of wealth and Jeff Bezos outweighs hundreds of millions of people, which is why averages and forecasting cannot represent such phenomena.
“the average for this distribution--if it exists, assuming it exists--will come from one observation. So, you can have a million observation. One will determine the average.”
- A flaneur is someone not guided by any constraint—engaging in aimlessness, whether intellectual or physical—and any attempt to make such wandering productive will fail.
“aimlessness is the idea, whether intellectual or physical”
- Practicing authenticity even where it is unobservable—like Visconti putting real jewels in a prop box no viewer would see, or dressing well below the waist on a video call—builds a habit that protects against being an imposter, and fakeness (especially among psychologists who write about probability without knowing it) is always detectable because a fake person is fake everywhere.
“You should never be an imposter... There's something about a fake person. You notice because they will be fake elsewhere”
Historical societies built quarantine into economics effectively; modern failure reflects incompetence and misallocation relative to the actual threat.
- Pre-modern Eurasia handled recurring plagues effectively by building quarantine into its economic system: cities refused suspect ships, returning residents went to lazarettos, and crossing the Ottoman-Habsburg border required quarantine; the 1720 Marseilles plague happened precisely because its mayor bypassed Genoa-style quarantine rules.
“in 1720 a ship left Smyrna with merchandise coming from the East... The Genovese didn't let it in because they had something called lazaretto, which was a quarantine for all ships”
- The WHO and CDC's anti-mask stance stemmed not merely from lying about a supply shortage but from genuine incompetence: they failed to understand the compounding nonlinearity and wrongly assumed only governments could produce masks, when in fact a market exists and people could improvise effective barriers from household items.
“It's not just that they lied. They're incompetent. They didn't understand the compounding.”
- America spends roughly a trillion dollars a year on defense because it refuses to gamble on national security, yet pandemics are vastly more dangerous than wars and require far less spending to guard against, so the failure to invest in proper testing and quarantine reflects a gross misallocation relative to the risk.
“we don't have to spend as much for pandemics. And, pandemics are vastly more dangerous than wars.”