In 2004, a professional poker player named Annie Duke made what was, by the rules of her profession, a very good decision at a very big moment. She had pocket aces in a tournament with life-changing money on the line. She bet aggressively. Her opponent called. Her aces lost to a backdoor straight. She walked away without the prize, watched the winner hoisted on people’s shoulders, and went home.
Most people hearing that story would say she made a bad decision. She lost. Good decisions produce good outcomes. Bad decisions produce bad outcomes. That’s how we learn.
Except, as Duke would later argue in a career she built after poker as a decision researcher and writer, that entire logic is wrong. Not slightly wrong. Fundamentally wrong. And learning to see why it’s wrong is probably the single most important shift an adult can make in how they reason about their own choices.
Decision quality vs. outcome quality
The core insight — Duke calls it resulting — is that good decisions don’t always produce good outcomes, and bad decisions don’t always produce bad ones. The world is noisy. Luck plays a large role. Most meaningful decisions involve uncertainty, which means their outcomes are, in statistical terms, samples from a distribution. You can make an excellent bet and lose. You can make a terrible bet and win. Judging the quality of the decision by the quality of the outcome confuses two different things.
This sounds abstract until you apply it. A student who cheats on an exam and passes hasn’t made a good decision; they’ve gotten away with a bad one. An entrepreneur who borrows everything and bets it on an improbable venture that happens to succeed hasn’t made a good decision; they got lucky. A driver who drinks heavily and makes it home safely hasn’t made a good decision; they rolled dice and the dice came up the easy way. Judging by outcomes alone would tell you all three had done well. The moment you introduce probability into your reasoning, the picture changes completely.
Conversely: a careful, well-reasoned investment that happens to lose money in a market crash wasn’t a bad decision. A sensible operation that happens to produce a rare complication wasn’t a bad call by the surgeon. A thoughtful relationship choice that ends painfully wasn’t, necessarily, a mistake. The outcome tells you something, but not the thing most people think it tells you.
What expected-value reasoning looks like
The mathematical version of this principle is called expected value. For any decision with uncertain outcomes, you multiply each possible outcome’s magnitude by its probability, add them up, and compare the options. The option with the highest expected value is, across many similar decisions, the one that produces the best results on average.
Consider a simple example. You’re offered a coin flip. If heads, you win one hundred dollars. If tails, you lose fifty. Should you take it? Outcome-based reasoning says well, you might lose, so probably not. Expected-value reasoning says: the expected value is (0.5 × 100) + (0.5 × –50) = +25 dollars. Over many flips, you’ll come out ahead by an average of twenty-five dollars per flip. You should take this bet every time it’s offered, even though any individual flip might produce a loss.
This seems obvious once laid out, but most people don’t reason this way in practice. They imagine the specific loss, feel its sting, and decline — even when the probabilities make the offer an excellent one. They also take bets where the expected value is negative — lottery tickets, casino games, certain speculative investments — because the rare large payout dominates their imagination.
The research behind the principle
The idea of expected-value reasoning is ancient; it was effectively invented by the seventeenth-century mathematicians Blaise Pascal and Pierre de Fermat. But the research on why humans don’t use it naturally, and how they can learn to, is modern.
Daniel Kahneman’s decades of work on what he called System 2 thinking — the slow, deliberate, effortful mode of cognition that can override our intuitive System 1 responses — underpins much of this. Expected-value reasoning is a classic System 2 operation. It requires you to slow down, to explicitly enumerate possibilities, to think about probabilities rather than specific outcomes, and to weight rare events in ways your intuition won’t. It’s available to everyone, but almost nobody does it automatically.
A separate body of research, led by Jonathan Baron and John Hershey in the 1980s, documented what they called outcome bias: when participants were shown identical decisions that had produced different outcomes (by random chance), they consistently rated the decision that had produced the better outcome as higher quality — even though the decision itself was the same. The outcome, not the reasoning, was driving their judgement of quality. This bias runs in all of us, and it’s worse in hindsight than in the moment.
The most sustained research programme on how people can improve at probabilistic thinking came from the psychologist Philip Tetlock, whose Good Judgment Project ran for many years as a forecasting tournament open to the public. Tetlock wanted to know whether some people were genuinely better than others at predicting world events — and whether that skill could be learned. What he found was striking. A small proportion of participants — he called them superforecasters — consistently beat prediction markets, government analysts, and subject-matter experts. Their defining trait wasn’t intelligence (though they were intelligent), or subject-matter knowledge (often they had less than the experts). It was the habit of probabilistic thinking: assigning specific numerical probabilities to outcomes, updating them incrementally as new information arrived, and resisting the pull of overconfident predictions.
Tetlock’s work suggests that expected-value reasoning isn’t just a technique; it’s a trainable skill that, practised over years, produces dramatically better decisions than intuition alone.
The counter-view worth hearing
Not everyone is convinced that expected-value reasoning is the right frame for every decision.
The risk theorist Nassim Taleb has built much of his career arguing that expected-value reasoning works for what he calls thin-tailed distributions — domains where outcomes are clustered around an average, and rare extreme events are genuinely rare (dice, roulette, well-behaved financial markets in ordinary times). But it fails catastrophically in fat-tailed domains — where rare extreme events dominate the long-term outcome (pandemics, stock-market crashes, nuclear accidents, the success of start-ups). In fat-tailed domains, the average is a misleading guide because a single extreme event can overwhelm thousands of ordinary ones.
Taleb’s critique matters for decisions with possible catastrophic downsides. If you’re betting your life savings, or making a decision where ruin is one possible outcome, calculating the expected value as if ruin were just one more outcome on a probability-weighted ledger is deeply misleading. You shouldn’t take a bet with a positive expected value if losing would destroy you, because you’d never get to play enough times for the average to materialise.
For these high-stakes, one-shot decisions with catastrophic tails, Taleb argues for a different framework: survival-first reasoning. Never take a risk that could be terminal, even if the expected value is positive. Because expected value is only meaningful if you’re still around to collect it.
Two reasoning styles for two kinds of decisions
So what does this leave us with? The most useful synthesis, I think, is to recognise that expected-value reasoning is the correct frame for decisions you’ll make many times and can recover from. The repeated poker hand. The investment portfolio rebalanced over decades. The hiring decision made many times across a career. The small business operating in a stable market. For these, thinking in probabilities and averaging across many decisions produces dramatically better outcomes than outcome-based reasoning.
But for a small number of decisions — the ones where a single bad outcome could end the game — expected-value thinking needs to be overridden by survival thinking. Here, the relevant question isn’t does this have positive expected value, but can I survive the worst-case outcome. If the answer is no, the probabilities don’t matter. You don’t take the bet.
Most people’s lives contain many of the first kind and a small number of the second. Learning to tell which is which is nearly the whole game.
What this changes about how you reason about your own life
The practical implication of expected-value thinking runs deep. You start to judge decisions by the quality of the thinking that produced them, not by the outcome they produced. This is both a liberation and a kind of discipline.
It’s a liberation because when a thoughtful decision produces a bad outcome — a considered investment that loses money, a sensible relationship that ends, a careful career move that doesn’t pan out — you can stop interpreting the outcome as evidence that you’re bad at deciding. You made the best decision available with the information you had. The world was noisy. It broke the wrong way.
It’s a discipline because when a reckless decision produces a good outcome, you can stop taking it as vindication. The drunk driver who got home safely, the gambler on a lucky streak, the entrepreneur whose terrible plan succeeded — none of them learned the right thing from their own success. They made bad decisions and got away with it. Continuing to make bad decisions will, eventually, catch up with them.
The question that remains
Duke ends her book Thinking in Bets with a simple idea. At the end of each day — or week, or year — try to ask two questions about any significant decision you made.
First: was this a good decision, given what I knew when I made it?
Second: did it produce a good outcome?
Most people only ask the second. Learning to ask the first — and to separate your judgement of your own decisions from the noisy, luck-drenched results they happened to produce — is a skill with no ceiling.
The question for your own life, then, is perhaps this:
If you separated the quality of your recent decisions from their actual outcomes, how many of your wins were really wins, and how many of your losses were really losses?
Key research referenced: Annie Duke, Thinking in Bets (2018); Daniel Kahneman, Thinking, Fast and Slow (2011); Jonathan Baron and John Hershey’s outcome-bias research (1988); Philip Tetlock, Superforecasting (2015) and the Good Judgment Project; Nassim Taleb’s writing on fat-tailed distributions and survival-first reasoning.