337 | Kevin Zollman on Game Theory, Signals, and Meaning
What this covers
Sean Carroll interviews philosopher Kevin Zollman about game theory and its reach across biology, human behavior, and language. Zollman argues that game theory functions as a mathematical tool—like calculus or linear algebra—rather than an empirical science that can be proven false. This framing explains why it proves useful far beyond economics: it can model any scenario where agents with goals interact strategically, whether those agents are humans negotiating, animals competing for mates, or genes spreading through populations. The conversation traces how this formal apparatus illuminates concrete cases, from peacock displays to poker hands to the origins of linguistic meaning.
The discussion moves through several domains. In biology, evolution acts as the rational decision-maker, allowing evolutionary game theory to explain why animals behave as if they calculate odds, even when no calculation occurs—side-blotched lizards cycling through three mating strategies, or peacocks evolving costly ornaments. In human interaction, the prisoner's dilemma and repeated games like tit-for-tat show how cooperation emerges or breaks down depending on payoff structures and shadow of the future. Zollman then makes a more provocative move: defining meaning and signaling entirely through game-theoretic terms—a peacock's tail means fitness not because the bird intends it, but because costly signaling creates an equilibrium where only genuinely fit males can afford the display. This naturalizes how language coordinates behavior without requiring intentionality, and how conversational implicature emerges from pure strategic reasoning. The episode also touches on premature convergence in science, the role of signaling in science denial, and why the handicap principle solved a biological puzzle so thoroughly that rival theories languished for decades.
Zollman argues that game theory is a domain-general mathematical tool—not a falsifiable empirical theory—that can model strategic interaction across humans, animals, and even genes, and that it offers a naturalistic account of meaning, signaling, and language without requiring intentionality.
- Game theory is like math (a tool that can be useful or not) rather than science (right or wrong), so it can model nearly any goal-directed interaction.
- In biology, evolution acts as the 'rational actor' selecting strategies, letting game theory explain animal signaling, mating displays, and fights.
- Defining meaning, lying, and deception in game-theoretic terms shows their core can be captured naturalistically without higher cognitive intentions.
Repeated games sustain cooperation through credible future consequences, while one-shot games do not, unless players actually care about others' outcomes.
- In the repeated prisoner's dilemma, the simple tit-for-tat strategy—cooperate first, then copy the opponent's last move—does very well and sustains cooperation that is impossible in the one-shot game, because the threat of future retaliation (defecting tomorrow if you defect today) makes staying silent rational.
“what was impossible before, the rationality of staying silent, becomes possible now. Because the threat of punishment, if you confess on me today, then I'll confess on you tomorrow... The threat of that is sufficient to keep the cooperation sustained.”
- Many situations described as prisoner's dilemmas aren't, because if you actually care about the other person's outcome that must be built into the payoffs, which often converts the interaction into a different game such as a stag hunt; game theorists don't prescribe selfishness—they say given what you care about, here is what you should do.
“if I don't want you to go to jail either. We're not playing a prisoner's dilemma, we're playing a different Game... It's called a stag hunt, probably, depending on how much I care.”
- Attempts to study one-shot anonymous games in small-scale societies break down because secrets are hard to keep and, more deeply, religious beliefs in divine punishment or karma mean no action is ever truly isolated from future consequences—effectively turning all games into repeated games.
“in a certain sense, God means that all games are repeated, right? There's no way that you can ensure that somebody's action is completely isolated from its consequences way down the road.”
- The prisoner's dilemma is just a math problem with no inherent moral implications: whether cooperation (staying silent) is good depends on the case—two colluding gas stations keeping prices high is bad for consumers, so sometimes defecting is the moral outcome—and people wrongly moralize criminals staying silent and escaping punishment.
“it's a math problem. It doesn't really necessarily have any moral implications.”
Utility functions mathematically represent preferences through choices, but observed behavior alone cannot reveal preferences without importing psychology.
- In the classic economic sense, a utility is a mathematical way of representing preferences—saying you prefer tea to coffee just means tea has higher utility—and the specific numbers assigned are arbitrary unless you use richer methods to make differences meaningful.
“a utility is just a mathematical way of representing what your preferences are. So I like tea more than coffee. So we would say that my utility for tea is higher than my utility for coffee because I prefer it. And that's very weak. You know, you could use any two numbers there.”
- Economists' hope of inferring preferences purely from observed choices fails because the same observed action (e.g., eating a glob of wasabi) is ambiguous between very different underlying mental states, so you cannot interpret choices without letting at least some psychology back in.
“she argues that this idea that economics can completely divorce itself from the way people think about problems is just the wrong way to think about it. But she thinks you can kind of get away with just letting a little bit back in.”
- There exist representation theorems showing minimal conditions under which behavior can be modeled as maximizing a utility function, and people's behavior is broadly consistent with this, but there are famous failures, making expected utility a mathematical idealization whose error—like ignoring friction—is debated in behavioral economics.
“it's a mathematical idealization. So it's like, I mean, the debate is, is it like ignoring friction when you calculate the motion of the planets? Like, you know, it's wrong, but it's not that bad because there's not that much friction out there.”
Signaling requires unfakeable costs to convey hidden quality; costly handicaps guarantee honesty because only high-quality individuals afford the expense.
- For a signal to reliably communicate hidden quality it must be unfakeable or hard to fake, otherwise low-quality individuals would mimic it and the signal would lose meaning; the original handicap principle held that a costly signal (like an exhausting dance or an expensive Rolex) guarantees honesty because only high-quality individuals can afford the cost.
“in order to make it so that you're communicating something, that communication has to be unfakable or at least difficult to fake. Otherwise it would go away.”
- Zollman, Carl Bergstrom, and Simon Huttegger developed the 'hybrid equilibrium,' a coin-flipping signaling solution, and report what they believe is the first wild example in hummingbirds, where some females adopt male coloration—an outcome predicted by the theory whose biochemical mechanism remains unknown.
“called the hybrid equilibrium. It's one of these coin flipping examples... we actually just found what we think is the first wild example of this in hummingbirds, where hummingbirds are signaling whether they're male or female, but sometimes the females adopt the male's colors”
- When ecologists went out to measure the costs the handicap principle predicted, they didn't find them as much as the theory required, prompting biology to abandon the dominant single explanation and generate a large menu of alternative game-theoretic signaling theories that now need testing.
“when ecologists went out and started measuring costs, they didn't find them as much as the theory predicted that they should. And so biology is kind of an interesting moment right now where people have gone back to the drawing board”
Dominance, equilibrium, randomization, fair division, rock-paper-scissors cycles, and asymmetric fight games exemplify game theory's diverse strategic structures across nature.
- Von Neumann and Morgenstern's insight was to measure how much more someone prefers one thing over another by asking which lottery between a most-preferred and least-preferred option they treat as equivalent to a middle option, revealing how much risk of the disliked outcome they will accept to get the liked one.
“we can ask what lottery ticket that gives me tea with some probability and coffee with some other probability do I regard as equivalent to water? So it's like, how much risk of getting the thing I hate, coffee, am I willing to give up in order to get the thing I might, I love, tea.”
- The sociology of science went awry with signaling theory: after Zahavi proposed the handicap principle verbally and Grafen and Maynard Smith formalized it in game theory, everyone flocked to the first mathematically working solution to a long-standing mystery rather than generating alternatives, delaying competing theories by decades.
“everybody just flocked the first solution that anyone came up with that worked mathematically and just decided that's got to be the story... if people had been a little bit more open minded, they might have been able to come up with these a couple of decades earlier.”
- Game theory's early development was intertwined with poker, with von Neumann an avid poker player and famous game theorists contributing papers proving theoretically interesting results about it; the formal beginning came when Morgenstern brought von Neumann the problem of utilities, producing the book 'Games and Economic Behavior'.
“von Neumann said, oh, I know how to solve this. And hence this book, 'Games and Economic Behavior' was born, which is really the sort of formal beginning of Game Theory.”
- A dominant strategy is one that is best no matter what the other player does, but most games don't have one; instead the best action depends on what you think the other player will do, and equilibria are predictions about what a sufficiently sophisticated system—via reasoning, evolution, or learning—would settle on.
“best strategy is, in the language of Game Theory, called a dominant strategy. That is the thing that's like best no matter what the other person does. Some games have them, but not all games. And in fact there's a real sense in which I would say most games don't.”
- The game of chicken (also called hawk-dove) captures a contest where the best outcome for you is to go straight while the other swerves, but mutual aggression is catastrophic; Maynard Smith and Price used it in the 1970s to model animal fights, capturing the fundamental problem of whether to merely threaten or actually follow through.
“it was used in the 1970s by Maynard Smith and Price to model animal fights and has been used a lot in biology since then. Very simple game, but it captures this really fundamental problem, which is like, do you just threaten but eventually give up, or do you threaten and really follow through?”
- Tit-for-tat is remarkable because it is extremely simple to implement yet performs very well, though it requires recognizing who you are playing against—a requirement that limits its applicability to organisms like bacteria that cannot identify partners.
“it's like this balance between immensely simple, like it's not complicated to implement... But on the other hand, it seems to do so very well... it does require that you know who you're playing against, which maybe doesn't always work in the animal kingdom.”
- The peacock signaling game gives the peacock four strategies (grow or don't grow a tail, conditional on being a good or bad mate) and the peahen four strategies (mate or not, conditional on seeing a long or short tail), with the choices being metaphorical because evolution, not the individual bird, determines the genetically programmed strategy.
“they could grow a big tail when they're a good mate and they could grow a big tail when they're a bad mate. So they... In a certain sense, they have four options... it's strictly speaking a four strategy game.”
- Two main theories explain mating displays: the sexy sons hypothesis, where ornaments like songs or bright feathers do nothing functional but are simply attractive (so attractive males have attractive sons), and the signaling hypothesis, where the display genuinely conveys unobservable qualities like strength or low parasite load—the latter being fully game-theoretic.
“One is what's called sexual selection or sometimes called the sexy sons hypothesis. And here the idea is that the song or the dance or the brightly colored feathers doesn't do anything. It's just, that's what it means to be attractive if you're a bird.”
- The chicken game has three equilibria: two asymmetric ones where one player always goes straight (hero) and the other always swerves (chicken), and a third symmetric equilibrium where both players randomize by flipping a coin, which is the best identical players can do and is predicted to arise in large populations of otherwise identical agents.
“there's a third which involves us flipping coins... that one is what's called symmetric... that's the sort of best you can do when we have to be identical”
- The 'I cut, you pick' procedure—proven by game theorists to produce a fair division because the cutter is incentivized to split evenly—generalizes far beyond cake to other parenting conflicts like dividing TV time or parental attention, provided the children are cognitively sophisticated enough (around six or seven, not two).
“Game Theorists have proven that this does what you want it to do. It makes them divide it fairly... if the kids are fighting over who gets control of the tv, you can say, okay, well, you both have an hour's worth of TV you can watch. One of you decides... and then the other gets to choose”
- Side-blotched lizards have three male mating strategies (harem-forming, monogamous, and roaming) that beat each other in a rock-paper-scissors cycle, so that whichever strategy dominates the population is undercut by the one that beats it, with evolution effectively keeping the proportions balanced so they appear randomized from any individual's perspective.
“if you're in a population of harem forming lizards, it's best to be the random one... If it's a population of people that just run around and mate with whomever they want to, it's better to be monogamous. And if you're in a population of monogamous, it's better to form a harem. So you have that same rock beats scissors, and scissors cuts paper, and paper beats rock structure.”
- Because language doesn't fossilize and no records exist from its emergence, computer simulations with agent-based models recreating the approximate circumstances of early language are essentially the only way to investigate how human language originated.
“a lot of our history of language, is buried in the deep past and we're never going to access it because language doesn't get fossilized... There's no other way to do it because it's just, the data is long gone.”
Game theory applies best when agents understand their strategic situation; evolutionary game theory works without agent awareness because evolution acts as decision-maker.
- Game theory is most applicable when agents are aware of their situation and the options available to themselves and others; it is least applicable when agents don't even understand they are in a strategic setting, though evolutionary game theory still works there because evolution acts as an aware decision-maker.
“the more aware agents are of their situation, the options they have available to them, the options that other people have available to them. So at the far extreme, if I'm engaged in a strategic interaction where I don't even know that I have options available to me or something like that, I think Game Theory is sort of least applicable in that setting.”
- In biological game theory, animals don't decide strategies; their strategies are genetically programmed, and evolution serves as the stand-in for the rational actor—selecting strategies that do better so that over the long run animals end up looking as if they reasoned rationally, even down to single-celled organisms.
“evolution is sort of the stand in for something like the rational actor model. So now it's just evolution is blindly, but still nonetheless guiding people, guiding animals towards smarter and smarter strategies... in the long run, they end up looking like they had decided all of this rationally”
- Game theory studies social interaction across the entire biological kingdom—scaling up from humans to corporations and countries, and down from humans to animals, plants, and single-celled organisms—whenever two parties each stand to gain or lose from the other's actions.
“it studies social interaction all across the biological kingdom, from us up and from us down.”
Word meaning functions as a coordination device when no conflict exists; conversational implicature follows game-theoretic reasoning about why utterances occur.
- The fundamental core of word meaning can be captured game-theoretically as a coordinating device: when there is no conflict of interest, words are arbitrary 'grease' that ensures our behavior coordinates (you fetch the tea I asked for), so meaning is defined by this coordinating function rather than by reference to possible worlds.
“the words are sort of superfluous. They're Just this grease that we use or this mechanism that we use to make sure that our behavior coordinates. And so if we can think about meaning as just like the... In that terminology”
- Conversational implicature—the gap between literal meaning and intended meaning, as in 'can you pass the salt?' literally asking about ability—can be explained by game-theoretic reasoning: the only reason you'd ask about my ability is if you wanted the salt, so I infer and comply, which is why the question comes to stand for a request.
“Well, why would you care whether or not I'm able to pass the salt? The only reason you would care is if you wanted me to pass it. So I should be able to kind of guess that ahead of time and go ahead and pass you the salt.”
- If meaning can be defined entirely in the mathematical language of signaling games—so that a peacock's tail 'means' he's a good mate—then you don't need higher cognitive capacities like intention or planning to mean something, suggesting human-language terms steeped in intentionality may have a core that doesn't require it.
“once you start defining meaning in a way that's entirely just mathematical in terms of the language of games, it's not so clear that you need a lot of these higher cognitive capacities... you don't really need to have an intention or a plan to be able to mean something, because the peacocks aren't planning in that way.”
Game-theoretic analysis reveals how incentive structures drive scientific progress and science denial; tribal signaling and expert credibility are modeled as strategic games.
- Scientists are driven by a mix of good motives (truth, justice) and pecuniary ones (citations, promotions, fame, Nobel prizes), and putting science itself in game-theoretic terms lets researchers ask counterfactuals—what if we removed Nobel prizes, grants, or peer review—to understand how incentive structures help or harm scientific progress.
“if you put this in the language of Game Theory, you can start to ask questions of how would science work if it were different? What if we got rid of Nobel prizes? Or what if we got rid of grants?”
- Much science denial is a form of tribal group-membership signaling—the same basic function as hairstyles, clothes, and language choices—so a key challenge is preventing science from becoming a signal of identity, while recognizing that science-loving people are also constantly signaling group membership.
“we're constantly signaling group membership to one another from the way we cut our hair to the clothes we wear, to the language that we choose to use... it's fine that I wear a certain kind of shirt to signal the kind of person I am, but maybe not good that I deny science to signal what kind of person I am.”
- Distrust in science connects to the old novice-to-expert problem—how a layperson decides which of two contradictory experts to trust—and game theory can model solutions like expert certification testing, as well as the motivations of misinformation spreaders, to design systems that lower their incentives.
“it actually connects back to a very old philosophical problem called the novice to expert problem, which is you've got this fundamental problem of, you don't know anything about a domain, you've got two people who both claim to be experts. They tell you opposite things. How do you decide who to trust?”
Zero-sum games have minimax equilibria protecting players from exploitation but also preventing them from exploiting opponents playing badly.
- In zero-sum games the equilibrium has a special minimax property: playing it guarantees that even a smarter opponent who can predict you cannot take advantage of you, but the flip side is that you also fail to exploit an opponent who is playing badly, which is why poker players sometimes deviate to exploitative strategies.
“if you play the minimax strategy, they... That guarantees they can't take advantage of you. So it's a kind of safety. But the down... The flip side of that is, they can't take advantage of you, but if they're playing badly, you're not really taking advantage of them either.”
- In many games—penalty kicks, tennis serves, run-versus-pass calls, and poker—the best strategy is to randomize, which is counterintuitive because flipping a coin feels irrational, yet was mathematically proven optimal in zero-sum games like poker by von Neumann and Nash.
“there are lots and lots of situations where randomization is the best thing... this is really interesting because it gives a role for something, randomization, which feels in a certain sense like irrational or at least outside of rational decision making.”
Behavioral phenomena like fairness preferences and revenge in ultimatum games quantify deviations from bloodless rationality that scale with stakes.
- The Ultimatum Game shows that the bloodless-rationality prediction—that a proposer offers a penny and the responder accepts anything—fails because people sacrifice money to punish unfairness, and the willingness to do so scales with stakes (you might reject 30 of 100 but accept 3 million of 10 million), letting researchers quantify preferences for fairness and revenge.
“the prediction that you just described is exactly the prediction that we would get to if we assumed that people only cared about their own monetary payoffs... But of course, that's not how people actually feel about these things.”