Sunk Cost Sophistication

The sunk cost fallacy — throwing good money after bad because you've already invested so much — is one of the first things anyone learns about cognitive biases. It's in every intro economics course, every pop psychology book, every listicle of "thinking errors." And the standard advice is simple: ignore past costs, evaluate only future returns, walk away when the math says to walk away.

Gwern's contrarian examination of this consensus is one of those pieces that makes a familiar idea feel much less settled than you thought. His argument isn't that sunk cost bias doesn't exist — it's that the real picture is far more complicated than "humans are irrational about past investments."¹

Animals Don't Do It

The first surprise: sunk cost bias appears to be uniquely human, or at most uniquely primate. When researchers tested lower animals — digger wasps, cichlid fish, savannah sparrows, albino mice — they consistently found that animals optimise based on future costs and benefits, not past investments. Mother mice defend their litters based on how many pups they currently have, not how many they've been raising. Birds defend nests more fiercely as hatching approaches — but that's because the future investment required drops, not because the past investment increases.¹

This matters because the standard evolutionary story for cognitive biases is that they're adaptations that were useful in ancestral environments but misfire in modern ones. If sunk cost bias doesn't appear in other animals, it's unlikely to be an ancient adaptation. It seems to be something humans specifically learn — and children are less susceptible to it than adults, which is backwards from what you'd expect if it were a deep evolutionary feature.

Why It Might Be Rational

The clever move in Gwern's argument is to challenge the soundness, not the validity, of the sunk cost model. Yes, in a simple formal model with perfect information, sunk cost reasoning is clearly fallacious. But the real world isn't a simple model. Several factors make "ignore sunk costs" much less obviously correct:

Learning and persistence. If you're trying to acquire a skill, quitting at the first sign of difficulty is a terrible strategy. The "sunk cost" of prior effort is also an investment in a partially completed learning curve. Persistence through difficulty is how you get good at things. A perfectly sunk-cost-rational agent would sample many activities shallowly and master none — which is approximately what children do, and we specifically train them out of it.

Reputation and signalling. Abandoning projects easily makes you look unreliable. In a social species, being known as the person who sticks with commitments has real value, even when the specific commitment was suboptimal. The "don't waste" norm that Arkes and Ayton identify as the driver of adult sunk cost bias may be a culturally transmitted strategy that's beneficial on average, even if it occasionally produces bad decisions.

Switching costs. Dropping one project and starting another isn't free. The costs of transition, learning new contexts, and rebuilding momentum are real. A model that says "always choose the option with higher marginal return" ignores that switching to the higher-return option has its own costs. Under some assumptions about how expensive switching is, what looks like sunk cost bias is actually optimal persistence.

Organisational vs individual. The most compelling examples of sunk cost disaster — the Concorde, the Vietnam War, corporate escalation — happen at the organisational level, where the dynamics are completely different from individual psychology. A politician who withdraws from a war looks weak; a coalition of bureaucrats whose project gets cancelled loses influence. The sunk cost "fallacy" in organisations may be less a cognitive bias than a rational response to the incentive structures facing the decision-makers, even though it's irrational for the organisation as a whole.

The Meta-Lesson

The real lesson of Gwern's analysis isn't "sunk cost bias is fine, actually." It's that the intellectual move of learning a named bias and then pattern-matching it everywhere is itself a bias — the bias toward feeling clever by deploying frameworks. When someone says "that's just the sunk cost fallacy," they're often compressing a complicated situation into a simple label and missing important features of the actual decision.

This connects to a broader pattern in calibration and measurement: knowing the name of a bias is much easier than correctly identifying when it applies. A list of cognitive biases is a powerful tool, but like any powerful tool, it's dangerous in the hands of someone who applies it indiscriminately. The sunk cost fallacy is real. But so is premature quitting, and so is the Goodhart's law problem of optimising for "not-sunk-cost-biased" at the expense of actually making good decisions.