In the few minutes that takes, I make use of Windows, Linux, SSH, internet protocols, API gateways, tokenizers, PyTorch or TensorRT, inference engines like vLLM, the Dart language, the Flutter framework, Android build tools, and the network stack underneath all of them. And that is just the software layer.

Not one single person on Earth can reproduce all of this alone.

In 1958, Leonard Read published a short essay called “I, Pencil.” Milton Friedman later made it famous, in a video in which he shows a simple lead pencil and says:

“There’s not a single person in the world who could make this pencil.”

The graphite comes from one country. The cedar from another. The rubber from another. The brass, the paint, the glue, the saws, the trucks, the factories, the machines that shape the wood: all of it comes from different people in different places, most of whom have never met, and many of whom do not even know what final product their work is creating. No central planner held the whole blueprint in their head. The pencil simply emerged from trade, specialization, and coordination, resting on a wider background of global institutions and infrastructure that make such coordination possible.

That is one of the great miracles of modern civilization. No one knows everything, but enough people know enough, and their fragments of knowledge combine into something useful.

Something analogous has been true of software for a long time.

Even before AI, programmers were already building on layers they did not fully understand. Most web developers do not know how CPUs schedule instructions. Most app developers do not know how compilers optimize machine code. They learn what they need to make something work, and abstract the rest.

This has been a major theme in the history of computing: abstraction. Every big leap made computers more usable by letting more people ignore more of the machinery underneath. We moved from wiring machines by hand to machine code, from machine code to assembly, from assembly to higher-level languages, then to frameworks, libraries, cloud platforms, and app stores.

So the jump from code to natural language is not some bizarre discontinuity. It is just the next abstraction layer.

When ChatGPT appeared in 2022, regular people got their first mass-market experience of talking to a machine in natural language and getting useful output back. Very quickly, many people found LLMs useful for coding tasks, especially prototyping.

In June 2024 I made my first AI-assisted game without writing code myself. In February 2025 Andrej Karpathy coined a name for it: vibe coding.

I am writing this in May 2026, and in my (admittedly very nerdy) social circle, vibe coding has become common. I use it to make ad hoc apps for small problems, to build agents for my own workflows, and to make instant games for my kids when they ask for them. I love the feeling of having a creative vision in my head, describing it in ordinary language, and watching it become real a few minutes later.

No, I do not know how to make an app. I cannot personally recreate the operating system, the network stack, the compiler, the inference engine, the phone runtime, the chip supply chain, and the rest of the ladder beneath the thing I just built. But neither can the old-school programmer. Neither can the mechanical engineer make a tire from a rubber tree. Neither can the pencil executive mine graphite, mill cedar, smelt brass, and run a global logistics network alone.

AI lets us access an enormous store of accumulated knowledge. It does not eliminate skill, judgment or taste, and the low-level understanding of the systems that power it is still valuable. Knowing how things work can make debugging and maintenance easier, as well as reduce safety risks. So AI doesn’t eliminate the value of knowledge but it definitely lowers the barrier to yield it.

And there is another novelty here. For the first time, the tool can also teach the user how to use it. If you do not know how to start vibe coding, you can ask the AI how to start! The same interface that lowers the barrier to building also lowers the barrier to learning.

Vibe coding is as close to practical magic as our species has ever gotten: you imagine something, describe it, and a civilization-sized leviathan of accumulated knowledge, infrastructure and processes moves at your command to make it real.

Lucky us, all it takes to summon that leviathan is a good prompt.

The strangest thing about this isn’t the outcome. Games fail. The strangest thing is that the warning signs were visible in advance, and Sony still invested a fifth of a billion dollars and shipped the game.

During Concord’s open beta, player feedback was overwhelmingly negative. The game felt generic in a crowded genre, with too many competing mechanics making encounters hard to read. Separately, the $40 price point created visible friction in a genre that had been free-to-play for a decade. Both signals were visible even to outsiders. Sony does not appear to have course-corrected in response.

This isn’t a story about Sony’s incompetence, they do ship hits. It’s a story about a failure mode that’s nearly universal wherever production costs are high, feedback loops are slow, and the people making investment decisions are emotionally attached to the outcome. But with the right framework, it’s largely preventable.

What Every Game Is Trying to Prove

Every commercial product has to prove one thing: will it make more than it costs to build and market it?

In games, that breaks down into three questions:

1. Will enough people enjoy playing this game?

2. Can we acquire them at a cost that makes the math work?

3. Will they pay enough (or watch enough ads) to make the unit economics work?

I’ve worked as a game designer and product manager on Tennis Clash, and as creator and game director of Sky Warriors through its full development cycle. And the ideal way to look at game development in my opinion is to see it as a sequence of hypothesis tests, each designed to help you answer the questions above.

The earliest tests focus on questions 1 and 2. First: do strangers playing an unpolished build come back the next day? That’s retention. Second: can we acquire those strangers at a viable cost? That’s CPI (cost per install). The core loop is the hypothesis; retention and CPI are the primary evidence. Question 3 is only tested later. LTV, monetization, long-term revenue, all depend on clearing these first. If retention and CPI fail, nothing downstream saves you.

I want to share two stories about this process. In one, the metric misled us. In the other, the signals were correct but a major publisher kept ignoring them for years.

Tennis Clash: When the Test Was Wrong

During the testing of Tennis Clash, our CPI numbers were subpar. Not terrible, but below what we initially projected, and below what was needed for the game to be profitable. The natural read was that the game wasn’t working, and we seriously discussed killing it. Before we did, we asked a different question: what is the hypothesis that has to be true for this game to succeed?

The answer was that it had to appeal to real tennis players, and we had been testing in markets with no tradition in tennis. So we ran a market analysis to identify where tennis culture was deepest, selected Spain and Italy as better testing grounds, and wrote down what we expected: CPI should drop substantially in these countries, and the organic install multiplier should rise meaningfully.
The test was designed to falsify those specific predictions. The results cleared them. With a very low marketing budget, the game rose to the top of the download charts in both countries. This was the confirmation we needed to continue the development of the game, and it went on to become one of the most successful sports titles in the industry.

The lesson isn’t “ignore bad KPIs”. The lesson is: before you act on a test result, ask whether the experiment was designed to test the actual relevant hypothesis. An invalid test produces misleading evidence, and can sometimes be worse than no data at all. Don’t let one bad test talk you into killing a game if the test itself wasn’t valid.

The Pre-Bet

This reasoning, however, is the most abused tool in game development. Every team eventually learns to say “the test wasn’t valid” when the numbers come back bad. So I try to hold the reasoning to a rule:

Before the next test runs, write down three things: the specific KPI you predict will change, the magnitude (a number or a range), and the kill threshold: the number at which you end the project if the prediction fails. All three are required.

For Tennis Clash, that might have looked like: “CPI is inflated by testing in markets with no tennis audience, for a better test we should run in Spain and Italy. We expect CPI to drop by X% and organic multiplier to increase by Y%, at least. If CPIs move less than Z%, we kill the project.” The exact numbers aren’t the point, as they’ll always be game-specific. The discipline is naming them in advance.

That’s a falsifiable bet with a prediction, a magnitude, and an exit. If you can’t name all three, you’re not flagging an invalid test, you’re protecting a prior you got attached to.

The Pre-Bet is the difference between updating and rationalizing. It’s also the part of this essay I most want you to steal.

One honest note about this rule: I wrote it because I need it. I like retelling the Tennis Clash story because it’s the case where I was the one who spotted the need for a different test, and we ran it with a proper Pre-Bet. That isn’t always how it goes. Like most designers, I see my own projects’ numbers through rose-tinted glasses, and I often need other people to question my conviction and scrutinize my numbers the way I’d scrutinize theirs.

So on top of applying the Pre-Bet, there’s a second layer: a culture where “the data says stop” can be said to the person most attached to the outcome, and heard by them.

EA’s Mobile Shooters: When the Test Was Right

For the opposite shape, look at a public case: EA’s mobile shooter portfolio.

EA made two parallel bets on mobile shooters. In 2018, it acquired Industrial Toys, a studio led by former Bungie CEO Alex Seropian, specifically to build Battlefield Mobile. Separately, it partnered with Tencent’s Lightspeed Studios to build Apex Legends Mobile, which launched globally in May 2022 and won Best Mobile Game at The Game Awards 2022.

In January 2023, EA cancelled both projects in the same earnings call. Apex Legends Mobile shut down in May 2023, twelve months after its global launch. Battlefield Mobile, which had run a closed beta in Indonesia and the Philippines starting in autumn 2021 and expanded to an open beta across five Southeast Asian markets in November 2022, was killed before reaching global launch. Industrial Toys was closed in the same announcement, four and a half years after EA acquired it.

What’s notable is why both projects died. Alex Seropian later attributed the decision to three structural conditions:

1. Battlefield 2042’s troubled launch triggered broader introspection about extending the franchise to mobile

2. Apple’s IDFA changes drove mobile shooter CPI up substantially across the category

3. Apex Legends Mobile’s monetization couldn’t sustain the content treadmill the genre demanded

Notice how the three reasons compound. The IDFA changes raised user-acquisition costs across the entire mobile shooter category. Apex Mobile’s monetization failure revised revenue expectations downward. BF2042’s bad reception eroded the brand’s organic pull on mobile. None of these were product issues that could be solved by the product teams, but they could have been spotted beforehand.

This is the failure mode the Pre-Bet framework is designed to catch. And here it could have been applied at the category level, not just the product level. IDFA was announced in June 2020 and rolled out in April 2021, well before Battlefield Mobile’s open beta and only a year after Apex Mobile entered serious development. A studio with an explicit category-level Pre-Bet, such as “shooter is a good genre because marketing costs are low, so if mobile shooter CPI rises by more than X%, we re-evaluate our entire mobile shooter portfolio”, would have caught and reacted to the structural shift earlier instead of clinging to an outdated belief about the genre.

EA’s eventual decision to kill both projects was disciplined; but the two-plus years between IDFA’s full rollout and the cancellation suggests the discipline arrived later than it should have. This is a form of escalation of commitment bias, in which the team kept working on an entire category because no one named the condition that would invalidate the thesis behind the bets, not just the bets themselves.

What You’re Actually Estimating

Zoom out and look at the industry.

The top 1% of mobile publishers captured 92.5% of all in-app purchase revenue in 2025. In the same year, only 828 mobile games globally earned more than $10 million in IAP. For context, roughly 200,000 new games launched across the App Store and Google Play that year, to give you the scale of the funnel. The entertainment industry is tough, tech is tough, and being in the intersection of both the games industry is particularly cruel. Hits are rare relative to releases, and rarer still within any peer group you actually belong to.

That’s the shape of the distribution before you have any game-specific information at all. And yet most teams plan and work as if they are very likely (or certain!) to make a hit. That’s an epistemic error, and it happens before a single mechanic has been designed.

Your starting confidence should be anchored to that outside view. Starting at high odds before you have evidence isn’t ambition, it’s miscalibration.

It’s also not a fixed number. I don’t get to set my confidence once and defend it; every test forces me to start from a new number. A game that survives alpha carries a different starting confidence into beta than one that hasn’t been tested at all. Confidence is a function of how much evidence you’ve passed through, not an intrinsic quality of the idea.

A rough, illustrative model of how games move through development at a healthy mobile studio: out of 100 ideas, maybe 30 reach alpha, 10 enter soft launch, 3 make it to global launch, and 1 becomes a meaningful success. The exact numbers vary widely by studio, genre, and risk appetite, this is not a benchmark, it’s a sketch of the shape. The takeaway is that the idea-stage prior is somewhere around 1%. By soft launch it’s risen considerably. Not because you’ve talked yourself into confidence, but because you’ve passed through rounds of evidence that most games don’t survive.

The right comparison isn’t “all mobile games,” though. Most of that denominator is asset flips and first-time developers. Your reference class should be games of your genre, budget, and studio track record, what Flyvbjerg calls reference class forecasting: identify the most relevant peer group, use its distribution as your starting point, then adjust for what makes your project materially different. This is a big epistemic advantage of incumbents: a larger graveyard of past failures to calibrate against. Established studios usually have dozens of launched games and hundreds of killed ones to use as a analytical starting point. But this doesn’t change a hard fact: Most games will fail, even when being created by amazing teams. Accepting that is the first step.

And “will this game succeed?” is the wrong question anyway, it’s too binary to be useful. In an ideal world, we would try to predict the exact likelihood distribution of outcomes. But since that is not feasible, most teams will approximate that by looking at tiers of success: probability of breaking even? Of $10K in daily revenue? $100K? $1M? Different questions with different answers, and the evidence that moves one doesn’t necessarily move the others.

There’s a counterintuitive effect here worth naming: a game that’s very likely to break even at the prototype stage is often less likely to reach $1M a day. A safe bet to break even is often a bad bet for a hit. Most hits require bold innovations, which are inherently risky. So if you are aiming for a hit, you should be comfortable with a low probability of breaking even. And if your studio can’t afford to go through multiple failed games, you should probably aim for a safer bet, and accept that it will likely not be a genre-defining hit (but might give you the resources for the next game to be one).

Finally, there are two different jobs inside “developing the game,” and teams conflate them constantly: measuring and improving. Measuring the game accurately is creating a better model of what the game is. Improving the game is actually changing the product. A test that reveals 35% D1 retention sharpens your estimate of where you stand. A tutorial redesign that lifts that D1 to 45% changes the underlying reality.

Teams that try to improve a game without good player data are trying to change reality with a flawed model. Teams that keep testing the game in different settings with no proper hypotheses are just trying to find a map that fits their hopes, not reality. For a game to succeed you need to know both what to change, and how to change it. The two questions you need to answer, recurrently during development, are:

1. What experiment would improve our confidence at a specific revenue threshold? That’s the measurement question.

2. What change to the product would shift those odds upward? That’s the improvement question.

The Pipeline

Different studios have different pipelines, but they all follow similar principles. Here’s a simplified, illustrative version of how a healthy mobile free-to-play studio might approach the funnel. Actual ratios, market choices, and gate criteria vary widely by genre, budget, and risk appetite, what matters is the shape of the process, not any specific number or country:

Idea to prototype. Every game starts as a conviction: a mechanic that could be fun, a genre that feels underexplored, a better version of an existing idea. Most studios define it in some form of short Game Design Document or Investment Thesis so other people can criticize and improve it before a single line of code is written. Then teams prototype it. Sometimes through hackathons, compressed sprints where concepts compete on merit before anyone gets attached, sometimes through teams building prototypes in series. Both approaches ask the same question: is this idea fun?

Alpha. The first external test, typically in lower-CPI markets. Southeast Asia and parts of Latin America are common choices in mobile because they offer enough scale to read a retention curve without burning a global UA budget. The KPIs measured are usually retention and CPI. The core loop is the hypothesis: If strangers playing an unpolished build aren’t retained by the core loop, polish will not save it.

Beta and soft launch. If alpha passes, the product expands and moves to markets with high disposable income and mature mobile habits. Markets like Australia, Canada, and Scandinavia are common for a first read on LTV. The question evolves from “is this fun?” to “how much money can this make?” By the end of beta, the game should be able to sustain itself, running further experiments without burning additional capital. That’s when a studio typically considers it in soft launch, and you should be able to estimate probabilities across all your revenue thresholds (or at least the probability to break even) with observed data, not conviction. If UA efficiency, retention, and monetization hold under this pressure, you’ve earned the right to spend a global launch budget.

Global launch and live development. You’ve now tested every link, separately, in rationally increasing order of cost. The confidence you bring to launch isn’t intuition, it’s a posterior that survived multiple rounds of evidence. Not a high posterior in absolute terms: in the illustrative funnel above, only 1 of 3 global launches becomes a meaningful hit. But a 33% shot at a hit, backed by observed data, is a different investment from a 50% shot backed by nothing. You don’t need high absolute confidence before committing real capital. You need confidence that’s earned, and that is calibrated so that the math works out with multiple launches. After launch, the game itself becomes a continuing experiment: each new content drop, feature, or offer updates your estimate of how far it can go, and how much money you can keep investing in it.

The critical feature of this pipeline isn’t any individual gate.

It’s that investment scales with confidence.

You don’t spend global UA budgets before you have global data. You don’t assign a full development team before you have evidence that the core loop is fun. Each gate is cheap enough to kill, and consequential enough to count.

Why People Walk Through Closed Gates

Most mobile studios have some version of this stage-gate. What separates studios that survive from studios that pray isn’t whether they have the gates, it’s whether they honor them when the results are inconvenient.

Three cognitive failure modes reliably cause developers to walk through closed gates:

Uniqueness Bias. In game development it sounds like: “Our core loop is innovative enough to overcome average retention numbers.” This may be true. It doesn’t change the prior. You can believe your game has a 30% chance of hitting $50K a day and still acknowledge that the base rate is under 2% and that you need evidence, not conviction, before acting on the 30%.

To steelman the bias: some products do create their own markets. Early Minecraft spread through YouTube before its first marketing dollar. The original Clash of Clans had organic retention curves you couldn’t explain from the core loop alone. If a game is going to create its own category, you’ll see anomalous early signals: unusually high organic share, surprising retention in unexpected demographics, word-of-mouth that outpaces paid acquisition. But name them before you invest in it. If you can’t specify in advance which non-standard metric you’ll track, like k-factor, organic-install ratio, creator-driven reach, you’re not tracking anomalous signals. You’re leaving yourself room to rationalize after the fact. Flat metrics are not the signature of a misunderstood masterpiece. They are the signature of a game that isn’t working.

Illusion of Validity. Bad signals arrive with plausible alternative explanations. D1 retention below threshold? The tutorial is unpolished. CPI 3x above target? The creatives aren’t optimized yet. Every one of these might be correct. The problem is that teams apply them selectively, to negative evidence, never to positive. When D1 exceeds threshold, nobody says “well, the tutorial was unusually polished.” This asymmetric skepticism isn’t analysis. It’s narrative laundering: running a pre-existing conviction through a sieve of plausible excuses until it comes out looking like reasoning. The fix is the Pre-Bet: name the prediction, the magnitude, and the kill threshold before the next test. Convert the excuse into a bet.

Escalation of Commitment. EA’s mobile shooter portfolio earlier in this essay is this bias at scale: years of compounding investment across two flagship projects, with no visible category-level kill threshold until external conditions forced them to kill both games in a single earnings call. The pattern isn’t unique to mobile, either. Anthem at BioWare is the most thoroughly documented version outside F2P, with years of internal recognition that core problems existed (engine incompatibility, unclear direction, burnout severe enough to be tracked internally as “stress casualties”) met by leadership decisions to keep pushing rather than rescope. It’s the sunk cost fallacy in action: when you’ve spent six months, six more feels cheap. When you’ve spent sixteen, you’re trying to justify the first sixteen rather than evaluating the next sixteen. This is where uncalibrated studios die, not at a single bad decision, but across a chain of them that individually looked like conviction and collectively looked like denial in hindsight.

The Limits of the Pipeline

The pipeline has one failure mode worth naming: false negatives. It’s designed to catch games that look promising but aren’t, and it does that well. It’s less good at catching games that would have defined a category if they’d been given more time to find their audience. The anomalous signals I described above are real, but they’re often delayed, noisy, or invisible to paid-UA tests. In mobile free-to-play, where feedback loops are fast and soft-launch metrics predict global outcomes well, this is rarely the binding constraint. In other genres and platforms, it’s a considerable concern. I’m confident that the discipline I’m describing is the right default, but that doesn’t mean that it is the right tool for every bet. And even when it is the right tool, it isn’t infallible.

In 2022, Supercell killed Clash Quest after roughly sixteen months in soft launch, despite running it on the Clash IP, a franchise worth over $10 billion in lifetime revenue. “We have launched five hit games but we have killed 30+, by my latest count,” their CEO Ilkka Paananen wrote in 2023. At Supercell, when a team kills a game, they have a tradition to open champagne bottles. Not to celebrate failure, but to mark the learning. The discipline has made them one of the most profitable studios in the industry: $3 billion in revenue in 2024 with fewer than 700 employees.

But discipline isn’t infallibility. Squad Busters cost roughly $150 million across development, UA, and marketing, and posted D1 retention of 61% and D7 of 38% in its second beta, strong numbers by any standard. It was reasonable to expect that it would be a hit. But the good retention numbers distracted the team from the long-term issues and the monetization shortcomings. It failed to retain and monetize players after global launch, becoming the first Supercell game ever killed post-launch.

Passing the gates is necessary. It isn’t sufficient.

Supercell doesn’t kill games because they are bad developers. They kill games so that they can afford to fund the ones that are really worth it.

Back to Concord

Sony’s mistake wasn’t making the game. It was spending $200 million before the data justified it. A studio running a proper stage-gate, one that treated the negative beta feedback and pricing friction as evidence, might have killed the project after a $5M prototype, or pivoted before full production. Either outcome is better than eleven days of live service.

The goal isn’t to never fail. The goal is to fail cheaply.

A calibrated studio kills nine games and funds one hit. An uncalibrated studio funds one game and prays. Both have the same number of global launches. Their economics look nothing alike.

I’m writing about games because that’s what I know. But the failure mode is the same wherever production costs are high, feedback loops are slow, and the people making the investment decisions are emotionally attached to the outcome. Pharmaceutical pipelines, film studios, and venture-backed startups all face variants of this problem: the base rates are brutal, the narrative is seductive, and the sunk costs feel like arguments. Don’t fall in the same trap.

Epistemic Status

I draw on a decade of experience inside free-to-play mobile studios and on conversations with dozens of others across mobile, console, and PC. No studio has a failure-proof process. This essay is a summary of good practices, not a mandatory checklist.

Concord’s outcome is documented; its internal decision-making is not. Visible pre-launch signals existed. Whether Sony ignored them, underweighted them, or judged that pivoting was more expensive than shipping, I don’t know. A similar caveat applies to EA’s mobile shooter portfolio: I’m working from public reporting and from Alex Seropian’s post-cancellation account, not from inside knowledge of EA’s internal decisions.

The Tennis Clash example is partly a success-survivorship story. The Pre-Bet is the rule I wrote because I noticed, looking back, that other projects with the same shape of reasoning got the opposite outcome. The rule exists to force the prediction before the bet, not after.

Sources and Further Reading

Data and industry

• Sensor Tower, State of Gaming 2026 — 2025 release counts (~55K App Store, ~150K Google Play) and revenue concentration; key figures also reported by Game World Observer

• SuperScale, “Good Games Don’t Die” — whitepaper behind the 83% failure rate (PDF, free)

Cognitive science

• Kahneman & Lovallo (1993), “Timid Choices and Bold Forecasts: A Cognitive Perspective on Risk Taking” — inside view vs. outside view

• Arkes & Blumer (1985), “The Psychology of Sunk Cost” — escalation of commitment

Megaproject research

• Flyvbjerg, “What You Should Know About Megaprojects and Why: An Overview” — reference class forecasting and uniqueness bias (arXiv, free)

• Flyvbjerg, Budzier & Christodoulou (2024), “The Uniqueness Bias” — SSRN, free

• Flyvbjerg & Gardner, How Big Things Get Done (2023) — book-length treatment

AAA escalation case

• Jason Schreier, “How BioWare’s Anthem Went Wrong” — Kotaku, April 2019 — public post-mortem of escalation of commitment outside mobile F2P

EA’s mobile shooter cancellations

• GamesBeat, “EA cancels mobile Apex Legends and Battlefield games, shutters Industrial Toys studio” — coverage of the January 2023 announcement

• Game World Observer, “Why EA canceled Battlefield Mobile: Apple’s IDFA changes and Battlefield 2042’s botched launch” — Industrial Toys founder Alex Seropian’s post-mortem

• Mobilegamer.biz, “Battlefield Mobile has soft launched in five territories” — coverage of the November 2022 open beta

Supercell

• Paananen, “The Next Chapter of Supercell” (Feb 2023) — “five hit games, killed 30+”

• Paananen, “Comfortable Feeling Uncomfortable” (Feb 2024) — “greenlighting teams, not games”

• Paananen, “The Best Days of Supercell” (Feb 2022) — Clash franchise $10B+ lifetime

• Dower, “Quality is Worth Killing For” — GDC 2016 talk (free)

• Deconstructor of Fun, “Inside Squad Busters’ $150 Million Experiment” — the authoritative breakdown

Piccini.app

I have just published The Epistemological Zombie, an essay about how people can easily stop reasoning, reaching conclusions through tribal affiliation and then construct the reasoning afterward. I described several mechanisms that produce the failure. One of them is perfectly exemplified here. What I called “Reasoning Unchecked” combines Haidt’s elephant-rider model of moral judgment, and Mercier and Sperber’s argumentative theory of reasoning, and it predicts exactly what happened in the red-vs-blue dilemma. The discourse is the cleanest natural experiment in tribal epistemology I’ve watched in years.

I’ll tell you which button I pressed at the end. First, let’s see how the question devolved into tribalism.

The Game

Tim’s setup, simple version. Everyone on earth has to take a private vote, choosing between two buttons.

• Red: nothing happens. You walk away.

• Blue: everyone on Earth — all eight billion — presses whichever button they pressed. If more than 50% press blue, everyone lives. If less than 50% press blue, only the people who pressed red survive.

Different accounts reposted or reframed the question. The reposts were pretty consistent with blue winning with around 60% of the vote! But even small reframing tweaks moved the result by 30+ percentage points. When the colors became “yellow and brown,” 90%+ pressed yellow. When blue was reworded as “press blue, you die unless 50%+ also press blue,” red won 60%.

That last set of anecdotes is enough to see many respondents weren’t thinking that hard. The same logic problem changed answers by half a population when the framing moved. Of course framing matters, and if the question changes you can expect that other people will change their action even if you recognize that it is the same question, and adjust your own action in response. Some of this is rational focal-point updating: when the colors carry no implicit valence, players gravitate to whichever equilibrium the framing makes salient. But the magnitude, 90% in one direction on neutral colors like yellow/brown, with no underlying logic change, is too large to be focal-point updating alone. Pure coordination updating wouldn’t produce 90/10 splits on neutral colorings. We are not measuring beliefs about pivotality. We are measuring something else.

Two Questions, Not One

Here is the core observation of this essay. Tim posed one question. The discourse processed two.

Below is the discourse-defining thread. User @gfodor posts what he calls “the right analysis”: @Aves_Clovis’s contour plot showing red vs blue areas depending on the prior space. User @tmsvrb replies with a single line, “still blue:”, followed by a quote-tweet of psychometric data. I ask, in the next reply, how does this data make you vote blue? User @EditionA3 watches the exchange and gives the cleanest diagnostic of the entire week.

“He is not thinking ‘Which button would spare the most lives?’ He’s thinking ‘Which team DESERVES to win?’

He’s picked a tribe and is mostly arguing why his tribe deserves to win.”

Once you notice this, you can’t unsee it. The two tribes aren’t disagreeing about the same question. They are answering different questions. Tim asked one thing; each tribe heard their own.

Below is my attempt at a structural dissection. Each row is one axis along which the two frames diverge.

These are not disagreements about facts. They are disagreements about which question is being asked. Once a tribe is picked, the question gets locked, and every argument afterward is derived from a frame the other side never agreed to.

A few hundred posts of this. From both directions. For five days.

What the Mechanism Looks Like

Three psychological mechanisms predict this shape.

The elephant and the rider. Jonathan Haidt describes moral judgment as a small rider (conscious thought) perched on a massive elephant of intuition. In identity-salient domains, the elephant has already chosen its direction before the rider has noticed. The rider’s job is not to steer. It is to construct a convincing story about why the chosen direction was correct all along.

The buttons do not look identity-salient. There is no political team named “blue” or “red” (although political teams in the US are color-coded…). But Henri Tajfel’s minimal group paradigm, running for fifty years, established that arbitrary categorical labels are enough. Within minutes of being assigned to one of two groups based on which abstract painter they preferred, schoolchildren began discriminating against the other group in resource allocation. No history, no shared interest, no real stakes. The category alone generated tribes.

Red and Blue are minimal groups at Twitter scale. The buttons created two tribes in real time. By day two, each side was diagnosing the other’s moral character. The categories generated the tribalism on their own.

Argument as advocacy. Hugo Mercier and Dan Sperber argue that reasoning evolved as a tool for winning arguments in social contexts. Its job was social, with truth-tracking as a frequent but secondary outcome. In small face-to-face groups this works: each person argues their position, the group’s collective skepticism filters the weakest claims, social friction produces reasonable consensus. Confirmation bias makes you a better advocate, and the group corrects you.

The bug emerges when you remove the social friction. Scroll a curated feed alone, and the argumentative module runs in a closed loop, generating justifications for wherever the elephant wanted to go, with no skeptical interlocutor pushing back. The Twitter discourse on the buttons is the Mercier–Sperber prediction in real time: every participant entered a sub-feed of others who had pressed the same button, and produced an arms race of justifications none of which had to survive contact with the other side.

Sacred-value protection, but mutually directed. Philip Tetlock’s taxonomy of tradeoffs was built for exactly this kind of confusion. He distinguishes three types: routine (secular vs. secular, like time vs. money, processed without drama), taboo (sacred vs. secular, like organs vs. cash: produces moral outrage at the framer), and tragic (sacred vs. sacred, like saving one child or another: produces grief and difficulty rather than outrage).

The button game is structurally a sacred-vs-sacred tradeoff. Aggregate life weighed against individual life. So by Tetlock’s framework we’d expect a tragic-tradeoff response. Sadness, denial, careful deliberation. What we got was mutual outrage. The resolution: each tribe perceives the other as committing a taboo tradeoff, sacred-vs-secular. Red voters see blue as trading individual life (sacred) for performative virtue (secular). Blue voters see red as trading four billion lives (sacred) for personal safety (secular). Each side is, in its own frame, watching the other commit a moral pollution. The math, in either direction, is “manipulative” because accepting it means conceding that the speaker’s sacred has a price.

These three mechanisms — minimal groups, argument as advocacy, sacred-value protection — predict, jointly, the bulk of what we saw.

The diagnosis is symmetric. Each tribe assigns the other a moral pathology. The pathologies are constructed to be unanswerable: a coward cannot rebut a charge of cowardice without sounding cowardly; a virtue-signaler cannot defend their position without confirming the diagnosis. The categorical accusation is the perfect terminal move.

The actual game: A Modified Stag Hunt

The button game has a name in game theory. It’s not exactly a stag hunt (Rousseau’s parable of two hunters who must cooperate to take down the bigger animal, or settle for a hare each on their own) but it shares the structural pathology that’s been studied for two centuries since.

The pathology is this. There is a cooperative equilibrium where everyone wins (all blue or all red both produce zero deaths) and a risk-dominant equilibrium where defectors are protected (all red, or anything that lands clearly below the 50% threshold; only red voters are guaranteed safe). Between these equilibria sit the catastrophic mixed states near 50/50, where roughly half the population dies. The structural problem is that cooperating exposes you while defecting doesn’t. If everyone successfully coordinates on blue, no one dies and no one was exposed. If coordination fails, blue voters bear the entire cost. Red voters are insured against coordination failure by construction.

Two results from the game-theory literature press hard on a coordination game shaped this way.

Risk dominance beats payoff dominance under uncertainty. Harsanyi and Selten worked this out formally in the late ‘80s. When a coordination game has multiple equilibria and players are meaningfully uncertain about each other, the risk-dominant equilibrium wins, even when a payoff-dominant equilibrium exists. Cooper, DeJong, Forsythe, and Ross confirmed it experimentally a couple years later: subjects placed in coordination games with explicitly higher cooperative payoffs still defected to safe equilibria, especially when they couldn’t communicate beforehand.

A non-communication constraint compounds the effect. Eight billion voters cannot signal to one another before voting. Schelling, in The Strategy of Conflict, observed that without communication, players gravitate to whichever equilibrium is most salient given the framing — the focal point. The framing-effect data above is exactly this. When the game was reworded as “press blue, you die unless 50%+ also press blue,” the focal point was red, and red won 60.7%. When the colors became “yellow and brown” with no killing implied, the focal point was yellow, and 90%+ pressed yellow. The underlying logic problem was unchanged. The focal point moved, and votes moved with it.

This is the structural bedrock under the math. Even a perfectly rational player, given the actual constraints of the game, has reasons to expect that betting on coordination is unsafe. That’s the game theory. Whether you should still press blue depends on whether your prior on the population disagrees with the game-theoretic default — which is what the math comes down to.

When Is Each Button Rational?

The decision is governed by one expression. The expected-value advantage of voting blue over voting red:

ΔEV(blue) = λ · p(your vote is pivotal) − w · p(red wins)

Where:

• λ is the number of lives saved if blue wins because you voted blue. At the pivotal scenario, blue voters number about half the population, so λ is on the order of four billion.

• p(your vote is pivotal) is the probability your single vote is the deciding one: the probability the rest of the world lands at exactly 50/50 without you. This depends entirely on your prior about how the world will vote.

• w is how much you weight your own life against a random stranger’s. At w = 1, every life counts equally.

Almost all the disagreement between rational red and rational blue voters reduces to one term: p(your vote is pivotal). And p(pivotal) reduces, in turn, to two summaries of your prior on the population vote share: its mean (μ = what fraction do you expect to press blue?) and its variance (σ² = how confident are you in that fraction?). Plug them into a Beta distribution and you get a clean diagnostic of where each button is rational1.

The picture below plots ΔEV across the (μ, σ²) plane.

The blue island is the only region where blue is rational. It requires that you believe two things at once: that your prior mean is near 50/50, and that your prior variance is small. If you believe humans cluster on tribal questions (high variance), or if you think the mean is anywhere away from a near-tie, red has the better expected value.

Why this specific region as my plausible prior? The choice of bounds is qualitative, not derived. My read is that on questions of personal survival, humans have strong evolutionary wiring — toward self-preservation, toward in-group conformity, away from clean coordination on arbitrary thresholds — and the realized vote on a population large enough to average over individual quirks is unlikely to land at exactly 50/50 with low spread. That pushes the prior toward higher variance: the world is more likely to fall well above or well below 50% blue than to land on a tight bell at the threshold. Which way it falls — red-leaning or blue-leaning aggregate — I’m uncertain about, hence μ ∈ [0.3, 0.7]. This is a qualitative read, and if there is empirical evidence that humans coordinate cleanly on survival questions — that the realized vote share would be a tight bell near 50% with low variance, putting a defensible prior in the small blue island — I would update toward blue. I have not seen that evidence; the empirical patterns I know about (in-group polarization, voter clustering, conformity research, the Twitter discourse itself) point the other way.

Both Rational. Both Irrational.

Both buttons can be rational. The disagreement between a rational red voter and a rational blue voter is mostly a difference in priors, not values — both want to maximize expected lives saved; they differ on what they expect other people to do. A rational red voter has high prior variance: they expect humans to cluster on tribal questions, they expect coordination to fail, they expect the realized vote to land somewhere far from a clean tie. A rational blue voter has lower prior variance: they expect coordination on blue to succeed, or they expect the realized vote to land near the threshold, and they are pretty confident in it. Both are doing the same math. They differ on the inputs.

The model says something stronger, though: the range of priors that make blue rational is narrow. The blue island in the figure is a small region of (μ, σ²) space where you’d need both high confidence in coordination and high confidence the realized vote lands near 50%. Reasonable people can hold those commitments, but the conditions are specific, and most of what we know about how humans behave on survival questions points away from them.

The other side of the coin: both red and blue can also be chosen for irrational reasons. The same act of pressing red can come from either a rational reading of the prior or a tribal allegiance to whichever side codes as individualistic, anti-virtue-signaling, smart, strong. The same act of pressing blue can come from either a rational reading of the prior or a tribal allegiance to whichever side codes as cooperative, pro-social, kind, protective, selfless.

An operational test for “tribal.” This essay’s central claim about tribality should be falsifiable, so here is the test. Of the replies I read across the threads documented above, fewer than 1% engaged with pivotality, specified a prior, or processed reframes substantively. The rest diagnosed character, denied premises, or refused the framing. I would happily revise this essay if as few as 10% of replies cleared that bar: the claim is not that nobody reasons; it is that the modal reply doesn’t. Fewer than one in a hundred is a vanishing minority. If a future tally produced 10%+, I’d update toward “this discourse was more substantive than I claim, and tribalism is one explanation among several.”

Which Button I Pressed

Red.

I posted the formula in the thread. The reactions were predictable in shape. Some users engaged with the math. More called me a coward. One asked whether babies count, was told yes (per Tim Urban’s clarification), and pivoted to “blue voters are retards then.”

My prior region is the dashed amber rectangle in the figure, μ ∈ [0.3, 0.7] and σ² ∈ [0.10, 0.20], because every empirical pattern I know about how humans cluster on survival questions points to high variance and uncertainty about which way the aggregate leans. Under any prior in that region, ΔEV is negative at w = 1, which is to say: red saves more expected lives than blue. Combine that with the game-theoretic argument from the previous section (risk dominance, no communication, focal-point fragility under reframing) and the case for red is overdetermined: both the math and the structure of the game point the same way under any prior I’d defend.

If you have evidence for a different prior — that humans coordinate cleanly on survival questions, that the realized vote share would be a tight bell near 50%, putting your prior inside the blue island — you’d land on blue. We’d both be doing the same math. We’d have different priors about coordination. That disagreement is real and reasonable, and it’s narrower than it looks: the blue island is small, so a defensible blue prior requires specific empirical commitments most evidence doesn’t strongly support.

The reactions to the post were the demonstration. Each one was Mercier and Sperber exactly as written, a tribe defending its prior with whatever rhetoric was at hand. The math was visible. The tribality was louder.

You can press blue sincerely and correctly, by a process. You can press red sincerely and correctly, by a process. The epistemological zombie is the person who presses either button because that’s what their team does, and constructs the reasoning afterward. The people who refused to engage with p(pivotal), or with the game-theoretic structure, are the base rate. They are most people, in most decisions, most of the time.

The button game made the mechanism visible. It did not make it new.

Epistemic Status

Confidence in the tribal-vs-mathematical claim: high. The tweet evidence is direct. The framing-effect data (yellow/brown 90%+, blue-as-suicide 60.7% red) shows magnitudes too large for pure focal-point updating; less than 1% of replies engaged with pivotality substantively (operational test stated above; would update at 10%+).

Confidence in the academic mechanisms: high for Haidt, Tajfel, Mercier and Sperber. Moderate for Tetlock, the sacred-vs-sacred application requires the mutual-taboo-perception framing I describe; if respondents experience the dilemma as a tragic tradeoff, the framework predicts grief rather than outrage, and we’d need a different mechanism for the observed behavior. Moderate for Taber & Lodge applied to this specific context, the original findings are in political reasoning, and the replication literature is mixed on numeracy tasks.

Confidence in the game-theoretic framing: high. Stag-hunt dynamics, risk-dominance vs payoff-dominance, and focal-point analysis under non-communication are well-established. The button game maps onto these results cleanly, with the additional asymmetry that defection is personally insured.

Confidence in the mathematical analysis: medium-high for the derivation; the prior region claim (μ ∈ [0.3, 0.7], σ² ∈ [0.10, 0.20]) is qualitative, defended on evolutionary-wiring and human-clustering grounds. The shape of the conclusion (red dominates outside the narrow blue island) is robust to (α, β) parameterization within the region and to weightings 0.5 ≤ w ≤ 2.

What would change my mind: a large-scale analysis of the discourse showing 10%+ of replies engaged with pivotality, specified a prior, or processed reframes substantively. I’d also update on evidence that the blue/red split tracked pre-existing political identity (which would suggest the tribalism was imported rather than generated by the game), or empirical evidence on coordination behavior in survival-relevant collective decisions that supports a low-variance prior near the threshold.

Sources

Psychology & moral judgment

• Haidt, J. (2012). The Righteous Mind. Pantheon. — Elephant and rider; social intuitionist model.

• Mercier, H., & Sperber, D. (2017). The Enigma of Reason. Harvard University Press. — Argumentative theory of reasoning.

• Tajfel, H. (1971). Social categorization and intergroup behaviour. European Journal of Social Psychology, 1(2), 149–178. — Minimal group paradigm.

• Tetlock, P. E. (2003). Thinking the unthinkable: sacred values and taboo cognitions. Trends in Cognitive Sciences, 7(7), 320–324.

• Taber, C. S., & Lodge, M. (2006). Motivated skepticism in the evaluation of political beliefs. American Journal of Political Science, 50(3), 755–769.

Game theory & coordination

• Skyrms, B. (2004). The Stag Hunt and the Evolution of Social Structure. Cambridge University Press. — Modern treatment of the stag hunt.

• Harsanyi, J. C., & Selten, R. (1988). A General Theory of Equilibrium Selection in Games. MIT Press. — Risk dominance vs payoff dominance.

• Cooper, R. W., DeJong, D. V., Forsythe, R., & Ross, T. W. (1990). Selection criteria in coordination games: Some experimental results. American Economic Review, 80(1), 218–233.

• Schelling, T. C. (1960). The Strategy of Conflict. Harvard University Press. — Focal points; coordination without communication.

Pivotal voting & the math

• Good, I. J., & Mayer, L. S. (1975). Estimating the efficacy of a vote. Behavioral Science, 20(1), 25–33.

• Margolis, H. (1977). Probability of a tie election. Public Choice, 31(1), 135–138.

Empirical sources

• @Aves_Clovis, contour plot of blue advantage by prior, April 2026 — origin of the figure inspiration.

I call this an epistemological zombie. The philosophical zombie is a famous thought experiment: a being physically indistinguishable from a human but with no inner experience, no qualia, the lights are on but nobody is home. The epistemological zombie, or epizombie, is similar, but it exists in real life: biologically human, holds degrees, manages companies, votes in elections, argues with conviction online. Internally, though, the machinery required to reliably arrive at the truth is either missing, broken, or entirely switched off. They possess a library of beliefs, but they lack a reliable way to go from those beliefs to the truth.

The concept needs a careful definition, because it’s easy to weaponize as a sneer. An epizombie is not someone who holds beliefs I think wrong; it’s a property of epistemic process, not belief content. The behavioral criteria are concrete: an epizombie cannot state the conditions under which they would update their view; they interpret identical evidence asymmetrically depending on whether it confirms or threatens a prior; and their conclusions, over time, do not track the truth better than chance in their domain of supposed expertise.

Those criteria collapse into a single operational test: the bet. A person who genuinely engages with “what odds would you take, and what resolves the wager?” and then follows through on the resolution is, by this definition, not an epizombie on that topic. The refusal to put an explicit probability on a belief and stake something visible on it is the cleanest tell. Everything else can be performed; the bet cannot.

Some people are epistemological zombies all the time, but it’s a condition any person can enter, in some domains, at some times. The lawyer who applies rigorous logic to every contract is the same person who interprets economic data through a tribal lens. The scientist who demands rigor at work forwards conspiracy theories about nutrition. The engineer who calculates structural loads precisely defends a budget projection she never stress-tested because it confirms what she already wanted to do. You have seen this, I am sure — and both you and I are susceptible to it as well. The question is not whether you can enter this state, it’s in what domains, with what triggers, and how to avoid it.

A Taxonomy of Epistemic Failure

Distinct mechanisms produce identical-looking failures. Two people can both refuse to update on the same evidence: one because they lack the bandwidth to follow the inference, the other because they’re protecting a tribal identity. Same surface behavior; opposite interventions. The taxonomy that follows organizes the mechanisms along three causal layers: hardware (the raw capacity to do the inference), software (the methods you were or weren’t taught), and environment (the incentives and information context the methods run in). The point isn’t to be exhaustive. It’s to make sure you can tell which layer is doing the work in any given case, because the fix you owe someone depends entirely on the answer.

Hardware

Some epistemic failure is genuinely cognitive capacity: children, advancing dementia, severe impairment. The clearest cases sit at biological extremes. They’re also irrelevant to almost everyone reading this. If you’re following the argument so far, hardware is not your bottleneck; it almost certainly isn’t the bottleneck for the people whose epistemic failures concern you most either. Blaming hardware is convenient because it implies the problem is irreversible and located in someone else. But it’s rarely the real story, unless you are spending too much time on the wrong parts of the internet.

Software

This is where most epistemic failure lives. The good news is that it’s often fixable. The bad news is that it takes time and effort to fix, and it’s not something that comes naturally to most people.

The Mindware Gap. Keith Stanovich coined the term dysrationalia for a finding that should unsettle anyone who conflates education with rationality: high IQ and broken epistemic software coexist routinely. Standard intelligence tests measure raw processing power: how fast the machinery runs, how much it can hold in memory at once. They do not measure what Stanovich calls rationality: the disposition to update beliefs as evidence accumulates. His Comprehensive Assessment of Rational Thinking (CART) shows the dissociation is partial but real: myside bias and base-rate neglect are nearly orthogonal to IQ; cognitive reflection is weakly correlated. On the topics that matter most, intelligence is not protective. The reason is that probabilistic thinking, falsification, and the scientific method are not factory settings. They are acquired technologies of thought, software that has to be installed deliberately.

This shows up most sharply in medicine, where statistical reasoning carries life-or-death weight and years of intensive training are required to qualify. In a much-cited Gigerenzer et al. study (2007), 160 practicing gynecologists were given a conditional-probability problem in natural-frequency form:

Of every 1,000 women, 10 have breast cancer. Of those 10, 9 will test positive on screening. Of the 990 without cancer, 89 will also test positive. A woman tests positive. What are the odds she has breast cancer?

The correct answer is 9 in 98, or about 9%. The most popular answer among the gynecologists was 9 out of 10, or about 90%. Off by an order of magnitude. Only about 21% got it right. After a brief training session in natural-frequency reasoning, 87% did.

This is not an isolated finding. The earlier Casscells, Schoenberger & Graboys (1978) NEJM study at four Harvard teaching hospitals asked 60 physicians, residents, and medical students an analogous problem (1/1,000 prevalence, 5% false-positive rate). The most popular answer was 95%. The correct answer was 2%. Roughly 18% got it right.

These are not failures of intelligence. They are failures of installed software. Logic, probabilistic thinking, and the scientific method are not factory settings, they are acquired technologies of thought. A degree is domain training, not epistemological training. The credential certifies access to a field, it does not, by itself, teach students how to think rationally. Like reading, like algebra, the technology of proper reasoning has to be explicitly taught. Most people are reasoning without it.

Cognitive Miserliness: The Rational Irrationality Problem. Daniel Kahneman formalized cognition as the conflict between System 1 (fast, automatic, cheap) and System 2 (slow, deliberate, expensive). The cognitive miser uses System 1 even in cases where we would expect System 2 to be used. The cost of using System 2 is that it is tiring and hard. In most domains, this cost is not worth it.

Bryan Caplan gave this its sharpest economic framing in The Myth of the Rational Voter. In markets, being wrong is costly: your business fails, your investment collapses, the surgery goes badly. In politics (as in most beliefs about distant, abstract questions), being wrong is costless. One vote has never decided a national election. Your belief about immigration policy has no effect on immigration policy. You can hold whatever belief feels good — socially rewarding, tribally signaling, emotionally satisfying — with zero personal downside.

Caplan calls this rational irrationality: it’s instrumentally rational to hold epistemically irrational beliefs when the private cost of being wrong is zero and the social rewards are positive. In the past this would be mostly nullified by the wisdom-of-crowds, but this correction only works when individual errors are independent. Once shared media environments and tribal epistemics produce highly correlated errors, or in topics that are truly counter-intuitive to our 100,000-year-old machinery, that aggregation breaks. The fix that worked in small groups does not work in a modern information ecosystem dealing with very complex issues.

The implication: you cannot fix epistemic miserliness by telling people to try harder. You have to change the incentive structure. Two distinct levers do this, and they’re worth distinguishing.

The individual-discipline version is what Annie Duke writes about in Thinking in Bets. The question “would you bet on it?” forces explicit probability assignment most contexts let you skip; it discovers, in seconds, that your confidence is more theatrical than real. Caplan himself is famously aggressive about putting money on his beliefs for exactly this reason. As private discipline, the bet works.

The institutional version is harder and bigger. Caplan’s diagnosis — that voters optimize for tribal capital because the social rewards exceed any private epistemic cost — is not solved by personal bets. It’s solved by making being wrong publicly visible at scale. Philip Tetlock’s Superforecasting research shows public track records dramatically improve calibration, because being wrong suddenly has a visible price. Platforms like Manifold, Polymarket, Metaculus, and Kalshi are early-stage real-world implementations of the same logic. For the even harder problem of elections, two proposed solutions are Epistocracy, championed mostly by political philosopher Jason Brennan, and Futarchy, proposed by economist Robin Hanson. Both are quite radical changes and unlikely to be adopted nationally anytime soon. But are worth reading about and considering when thinking about solutions.

Environment

Even when hardware is sufficient and software is installed, the environment can corrupt the output.

Reasoning Running Unchecked. Jonathan Haidt famously describes moral reasoning as a small rider — conscious thought — perched atop a massive elephant of intuition. In identity-salient domains, the elephant has already chosen its direction before the rider has noticed. The rider’s job is not to steer; it is to construct a convincing story about why the chosen direction was correct all along.

Mercier and Sperber arrive at the same place from evolutionary cognitive science. Reasoning, they argue, evolved primarily to win arguments in social contexts — not to find truth. In small face-to-face groups this worked: each person argued their position, the group’s collective skepticism filtered the weakest claims, social friction produced reasonable consensus. Confirmation bias makes you a better advocate, and the group corrects you.

The bug emerges when you remove the social friction. Scroll through a curated feed alone, and the argumentative module runs in a closed loop, generating justifications for wherever the elephant wanted to go, with no skeptical interlocutors pushing back. Construct an information bubble of people who already agree with you, and you get validation instead of correction. Same direction, forever, at accelerating speed.

The Elite Zombie: Intelligence as a Weapon. The most dangerous epizombie is not the person who lacks cognitive capacity. It’s the person who has plenty of it, and deploys every bit of it to defend conclusions arrived at tribally.

Dan Kahan’s motivated numeracy experiment makes this concrete. Participants were given a statistical table and asked to interpret it. On a neutral topic (whether a skin cream reduces rash), the high-numeracy group substantially outperformed the low-numeracy group, and the gap was clean and large. Reframe the identical data structure as evidence on gun control, however, and the numeracy advantage collapses or inverts. Highly numerate partisans become more polarized than less numerate ones, applying their quantitative skills not to track the truth, but to find reasons to reject results that threaten their priors.

The effect is domain-specific. High intelligence is not a warning sign in engineering, medicine, logistics, or most technical work. There, capacity correlates strongly with accuracy, since errors are costly and feedback is swift. The liability emerges in identity-loaded, politically salient domains where the correct answer conflicts with tribal priors. There, additional intelligence makes you a better advocate for the wrong conclusion.

The institutional danger is obvious: the educated pundit whose credentials are mistaken for epistemic credibility, defending tribal conclusions with sophisticated arguments. In identity-loaded domains, the sophistication is not the reassurance. And it may even be a warning sign.

The Firehose and the Fog: Epistemic Learned Helplessness. A third environmental cause does not corrupt your reasoning so much as exhaust it.

When institutions fail repeatedly, when experts contradict each other on the same week’s question, when official guidance reverses without acknowledgment, when every fact arrives wrapped in a counter-narrative... a rational response is to stop trusting fact-evaluation as a project. Epistemic learned helplessness is the conclusion that truth is systematically unknowable. It is not stupidity. It is the adaptive response to an environment that has punished truth-seeking often enough that the cost of trying exceeds the expected gain.

The pattern can be deliberately weaponized. The RAND Corporation’s 2016 analysis of Russian disinformation describes the strategy as a firehose of falsehood — high-volume, multi-channel, contradictory messaging whose goal is not necessarily to convince anyone of any specific claim, but to flood the information environment until coherent inference becomes impossible. The mechanism doesn’t require deliberate sponsorship to operate, though. Routine partisan media, algorithmic amplification, and even good-faith attempts to correct misinformation by regulating discourse produce the same result for most consumers. Whether the fog is intentional or emergent, the destination is the same: the reader who concludes that everything is a psyop and retreats into tribal priors as the only stable signal.

Unlike the reasoning-running-unchecked or elite-zombie cases, where thinking is active but distorted, learned helplessness is the case where evaluative thinking has stopped happening at all. The “nothing is true” reader is not wrong that they have been lied to; institutions did lie, repeatedly. They drew the wrong conclusion from a correct observation. That distinction matters for how you intervene.

The Deliberate Zombie: Epistemic Relativism. The final cause is the only one that’s deliberate, though its practitioners would frame it as a philosophical achievement.

The vulgar version dismisses inconvenient facts with “that’s just your opinion” or “you can prove anything with statistics”, and can look awfully close to the previous case of learned helplessness. The sophisticated version, on the other hand, has specific claims that need to be unbundled — one with some merit, the other less so.

The global claim has three real components, and all three deserve to be conceded before they’re answered: observation IS theory-laden, scientific institutions ARE embedded in power and HAVE failed catastrophically (phrenology, eugenics, lobotomy), and the production of knowledge IS socially situated. Each of these is true. The mistake the global relativist makes isn’t in describing those conditions; it’s in concluding that the solution is to entirely reject the human project of knowledge-seeking instead of trying to improve it.

The Popperian move, beautifully articulated by David Deutsch in The Beginning of Infinity, concedes more than the relativist usually expects, and still defeats the conclusion. Theory-ladenness doesn’t imply all theories are equivalent; it means we compare them by what they predict, by what survives sustained criticism, by whether they self-correct under pressure. Historical abuse is real, and the relativist’s go-to examples are awkward for them: phrenology, eugenics, and lobotomy were discredited by the exact same epistemic apparatus that briefly certified them. That’s the project working, not failing. Power-embeddedness is real, but bridges built by power-seeking engineers still hold up cars. We need to see knowledge not as a possession guarded by an institution (which is what the relativist is correctly critiquing), but as an ever-evolving process of conjecture and criticism that produces predictions, which are testable in mind-independent reality. Every component of the global critique can be conceded; the process still produces progress that the relativist’s framework cannot explain or replicate.

The local claim is much harder to dismiss, and the reader has probably already agreed with it in a few cases. It says: in specific named fields where measurement is value-laden and replication is rare — much of social psychology before the replication crisis, parts of nutrition science, parts of macroeconomics — current institutional findings deserve a heavy discount, because the field’s incentive structure cannot in fact produce adequate epistemology now. The local skeptic is, in many of those domains, correct. The replication crisis was a vindication of local skepticism, not a refutation. The right response is to concede the ground they’re standing on, then note that this is a case for rebuilding specific fields, not for abandoning evidence-based knowledge as a project.

Coming back to the vulgar version: much of what looks like relativism today isn’t actually relativism, but just the learned helplessness from the previous section dressed in philosophical language. The intervention difference between the two cases is relevant: you can rebuild institutional legitimacy over time and re-engage the helpless person. You cannot philosophically defeat someone who has stopped trying to play the game.

The Antidote Is a Technology, Not a Virtue

The epistemological zombie exists on a spectrum, activates situationally, and usually operates through multiple simultaneous causes. No clean solution exists — anyone offering one is probably selling something. But the frame matters. Treat epistemic failure as a character flaw and you produce shame and defensiveness; treat it as a set of conditions, each with its own mechanism, and you produce something more useful: a diagnostic.

The mindware gap calls for explicit training in probabilistic thinking, falsification, and Bayesian updating. The skills exist; almost nobody installs them deliberately. Rational irrationality calls for making being wrong personally costly: at the individual level, the bet forces calibration in seconds; at the societal level, Philip Tetlock’s Superforecasting research shows that public track records dramatically improve calibration, because being wrong suddenly carries a price.

The unchecked argumentative module calls for deliberate exposure to genuine, skeptical disagreement. Find the smartest person who disagrees with you and read them carefully, truly trying to listen to the arguments instead of thinking of how to rebut them. Epistemic learned helplessness calls for the slowest fix in this list: rebuilding institutional credibility through demonstrated accuracy, transparent error correction, adversarial audits, and visible track records over time. Authority no longer pays. The only thing that re-engages someone who has retreated into “intellectual elites are untrustworthy” is the cumulative experience of seeing specific institutions get specific predictions right while admitting the ones they got wrong. And epistemic relativism calls for showing, case by case, that evidence-based methods produce predictions that land and structures that stand. That argument is slow. It’s the only one that works.

There’s a circular problem hiding inside that prescription, and it’s worth naming directly. The same forces causing the explosion of epizombies in modern times — institutional decay, weaponized disinformation, algorithmic amplification of correlated errors — also degrade the institutions needed to fix it. Who builds the prediction markets, runs the adversarial audits, demonstrates the track records? In practice, small networks of unusually rigorous actors do, and the surrounding civil society can reward or punish them. The current evidence is mixed: Manifold, Metaculus, GiveWell, Tetlock-style tournaments are working, real, and growing. They are also unimaginably small relative to the problem. And are seen with (often deserved) suspicion by the public and by regulators. The honest answer is that the fix is hard but not impossible. And it’s necessary if we want progress.

It’s also worth being honest about what this costs personally. Truth-seeking is painful. It often requires breaking with your tribe, admitting publicly that you were wrong, and accepting the social penalty for doing so. The zombie is comfortable. The truth-seeker is often not. That asymmetry is the reason most people default to the zombie path, and the reason institutional design matters more than individual virtue.

My prediction: rigorous epistemology will remain a very small minority technology for a long time. The structural diagnosis isn’t encouraging about rapid change at the population scale. But if we can improve epistemology a bit around key decisions in the next decade, that may be enough to tip the scales heavily in our favor.

The lights are off for most of us, most of the time. You can decide, today, to switch some of yours on.

Epistemic Status

Confidence in the core taxonomy: ~85%, these mechanisms are well-documented across multiple independent literatures.

Confidence that the prescribed antidotes work at the individual level: ~70% — well-evidenced for the bet/forecasting interventions; weaker for relativism antidotes.

Confidence that institutional design as prescribed will scale meaningfully in the next decade: ~35% — small working examples exist; the scaling story is not yet proven.

What would change my mind:

• Robust evidence that calibrated forecasting performance is not meaningfully dissociable from general intelligence — this would challenge the software/hardware distinction.

• Robust evidence that epistemic training does not transfer outside the training domain. There is some weak signal in this direction (debiasing-transfer studies often find modest, short-lived effects); a strong negative finding would gut the antidote case at the individual level and push everything onto institutional design.

Sources and Further Reading

Cognitive Science

• Kahan, Peters, Dawson & Slovic (2017), “Motivated Numeracy and Enlightened Self-Government”, Behavioural Public Policy

• Stanovich, What Intelligence Tests Miss (2009) — dysrationalia, IQ/RQ distinction

• Stanovich, West & Toplak, The Rationality Quotient (2016) — the CART battery

• Kahneman, Thinking, Fast and Slow (2011) — System 1/System 2, cognitive miserliness

• Mercier & Sperber, The Enigma of Reason (2017) — argumentative theory of reason

• Tetlock & Gardner, Superforecasting (2015) — forecasting tournaments, calibrated track records

• Gigerenzer, Gaissmaier, Kurz-Milcke, Schwartz & Woloshin (2007), “Helping Doctors and Patients Make Sense of Health Statistics”, Psychological Science in the Public Interest

• Casscells, Schoenberger & Graboys (1978), “Interpretation by Physicians of Clinical Laboratory Results”, NEJM

• Paul & Matthews (2016), The Russian “Firehose of Falsehood” Propaganda Model, RAND Corporation

Moral Psychology & Political Economy

• Haidt, The Righteous Mind (2012) — elephant and rider, social intuitionist model

• Caplan, The Myth of the Rational Voter (2007) — rational irrationality

• Brennan, Against Democracy (2016) — the case for epistocracy

• Hanson (2013), “Shall We Vote on Values, But Bet on Beliefs?” — futarchy

• Duke, Thinking in Bets (2018) — beliefs as wagers

Philosophy of Mind & Epistemology

• Deutsch, The Beginning of Infinity (2011) — good explanations, conjecture and criticism, knowledge as process

• Cassam, Vices of the Mind (2019) — vice epistemology, epistemic cowardice