A new paper by Alexandre Morozov and Alexander Feigel in PNAS offers a genuinely interesting evolutionary result. If each player in a population of Prisoner’s Dilemma agents carries an opponent-indexed cooperation probability — \(p_{c,i\to j}\), one entry per partner identity — then selection and mutation routinely produce cooperative configurations. The press release framed this as resolving the Prisoner’s Dilemma after seventy-five years of pessimism. The paper itself is more careful, and the careful version is the one worth looking at, because it makes visible a distinction formal theorists rarely state out loud.
The Prisoner’s Dilemma is not its payoff matrix. The dilemma is a property of the game form.
Four numbers — the payoffs to \((C,C), (C,D), (D,C), (D,D)\), satisfying \(e_{dc} > e_{cc} > e_{dd} > e_{cd}\) and \(2 e_{cc} > e_{cd} + e_{dc}\) — are necessary for a dilemma but nowhere near sufficient. What makes it a dilemma is the form: simultaneous moves, no commitment, no observability of the opponent’s plan, no contingent responses. Keep the four numbers and relax any one of those constraints and you no longer have a dilemma. You have a different game wearing the PD’s matrix as a costume. So if a paper reports cooperation in something it calls the Prisoner’s Dilemma, the first question is whether what was solved is the dilemma or the costume.1
What the paper actually shows
Morozov and Feigel earn the generous reading. Their model is not an Axelrod-style tournament of pre-specified strategies — they do not hand-code tit-for-tat into the strategy set and watch it win. Reciprocity-like configurations have to arise from selection and mutation on an opponent-indexed matrix. That is a different exercise from the iterated-PD literature, and the central result deserves to be stated cleanly.
In the standard opponent-blind formulation, where every player has one cooperation probability used against everyone, average fitness evolves as \(d\langle F\rangle / dt = -r\,\mathrm{Var}(p_c)\). The variance is nonnegative, the PD strength \(r\) is positive, so the derivative is nonpositive. Cooperation can only decay. This is Fisher’s fundamental theorem applied to a frequency-dependent fitness, and it is the formal heart of the standard “cheaters win” verdict.
Their move is to ask what happens once fitness becomes opponent-specific. The payoff tensor \(\gamma_{ij}\) decomposes into a symmetric part \(\alpha_{ij} = \frac{1}{2}(\gamma_{ij} + \gamma_{ji})\) and an antisymmetric part \(\beta_{ij} = -\beta_{ji}\), and average fitness now satisfies
\(\frac{d\langle F\rangle}{dt} = 2\bigl[\mathrm{Var}(F^s) + \mathrm{Cov}(F^s, F^a)\bigr].\)
This is, I suspect, the most undersold contribution of the paper. It is a generalized Fisher theorem for asymmetric games, and the structural echo is precise: the symmetric component is a potential game, the antisymmetric component is a tournament — a pairwise net-flow of who exploits whom — and the rate of change of average fitness now depends on the covariance between the two. Fitness has become relational rather than intrinsic. The Hodge-decomposition-of-games crowd will see the carpentry immediately. That is real mathematical infrastructure and deserves a wider audience than the “hope for cooperation” headline gave it.
The hinge is commitment
Once one has the infrastructure, the question is whether the model is studying the Prisoner’s Dilemma or something else wearing the matrix. I think it is something else, and the most precise way to say so is in terms of commitment.
In an evolutionary model, each player is its strategy. The genotype is heritable, fixed within the lifetime of the individual, and — once recognition is granted — exposed to the opponent. No agent best-responds to anything in real time; the strategy is simply executed. Deleting the within-encounter decision node is not a neutral convenience of the replicator formalism. It is the grant of commitment power.
A prisoner who could commit, observably, to a conditional strategy — a reaction function the opponent could read before moving — would no longer face the dilemma. Commitment to an unconditional cooperative action would just get the prisoner exploited; that is not the trick. Observable committed conditional strategies are Schelling’s solution to commitment problems, and \(p_{c,i\to j}\) is exactly such an object. It is a reaction function (conditional on identity), it is committed (genotype-fixed), and it is observable (the recognition assumption — and in the neural-network version of the model, the opponent literally reads the other agent’s network weights \(W_j\) before deciding). The model hands the prisoners Schelling’s entire toolkit and then reports that cooperation emerges.
That is the move, and it is worth being precise about. Not because anything in the paper is wrong, but because nothing in the paper marks the transition.
Where the history lives
A useful way to see what has actually happened is to ask where the information lives. Every model of cooperation that survives the existence question — every one of Nowak’s five rules, every Folk-Theorem construction, every reciprocity story — solves the problem by carrying information about past play forward in time.2 The interesting comparative question is where the carrying happens.
In a Folk-Theorem construction, the information lives in the within-pair history, which forward-looking agents condition their best responses on. Direct reciprocity, indirect reciprocity, reputation: variations of the same theme — history-as-input-to-strategy. In Morozov and Feigel, the information lives in the strategy itself. The opponent-indexed genotype is an addressed, opponent-keyed store of size \(O(N)\) per individual: not a literal record of past play, but a compressed representation of whatever the opponent’s identity has come to predict. Identity, in this model, is not a primitive. It is a sufficient statistic for the histories the agents would otherwise have had to track. No organism would be selected to maintain a per-opponent ledger if the histories with those opponents did not matter, and the histories matter only because the future does. The compressed representation has had its origin story erased, but the compression is downstream of exactly the thing the paper sets out to do without.
Maggie’s 2009 American Journal of Political Science paper, “A Model of Farsighted Voting,” makes the contrasting move available. She is studying a different cooperation problem — coalition formation over continuing policies, not pairwise PD — but the structural question is the same: how does information about the future enter the present, without enforcement, without binding agreements, without commitment power? Her answer is to put the information in the state. The status quo is a payoff-relevant variable that persists until replaced; every voter confronts it; the equilibrium conditions on nothing but the current state and the agents’ types. No trigger, no memory of play, no punishment. The past reaches the future only through the standing variable.
State-borne information is enforcement-free, because the variable is a thing everyone confronts. Strategy-borne information is commitment, because the strategy is a thing the agent is. That is not a small distinction. It is the precise statement of “the game form changed.”
Silver spoons
The paper contains a figure that reads, to me, as the cleanest demonstration of what the commitment is hiding. In Figure 4A, a single individual starts the simulation as “popular” — every other agent cooperates with this individual at probability 0.98, while the popular individual reciprocates at 0.10. The popular agent’s lineage rapidly sweeps the population. Average fitness rises monotonically. Both the variance and the covariance terms in the generalized Fisher decomposition contribute positively. The trajectory is, by the paper’s own framing, a success of emergent cooperation.
It is also a story about a trust-fund heir. The popular individual won because the rest of the population was committed — committed by their fixed genotypes — to cooperating with him. He did not need to reciprocate. The thing that produced his lineage’s reproductive advantage was being cheated toward, not cooperating. By the time his descendants fix in the population they cooperate with each other (\(p_{c,1\to 1} = 0.99\)) because the diagonal entry is the only one that matters once everyone shares the genotype — but the diagonal entry was not under selection during the phase that decided who won. Cooperation emerged as the consequence of a cheater’s victory, not as its cause.
The deeper point is that nobody in the model can see this. The replicator equation has no node at which an agent asks “why am I winning?” The faculty required to inspect the source of one’s own success — counterfactual reasoning over the strategy axis — is the same faculty the model needs to deny in order to call the result emergence. The model grants counterfactual reasoning generously along the opponent axis (recognition is itself a counterfactual claim, that who you face changes what happens) and revokes it along the strategy axis. The selectivity has no principle behind it.
This is, I want to stress, a coherent thing for a model to be. Coherent as dynamics, incoherent as explanation. Explanation requires the counterfactual the model deleted to get the result.
The audit
Here is where Morozov and Feigel’s work and Maggie’s sit side by side most productively. Both papers face the same explanandum: cooperation without classical enforcement, without binding agreements, without punishment. Both solve it by carrying information forward in time. They differ — sharply, and instructively — in their epistemic posture toward the assumption that does the work.
Maggie’s value function is, by construction, the faculty the silver-spoon agents are denied: each player’s evaluation of every policy in terms of what it is likely to produce over time, the counterfactual sweep over the whole alternative space. Her voters are constitutively farsighted. They run the very counterfactual the replicator equation has no node to pose. And then her Proposition 4 does the thing that makes the whole project legible: it constructs an explicit extensive-form game \(\Gamma\) and proves that the reduced-form equilibrium is a Bayesian Markov-perfect Nash equilibrium of that game. She builds the reduced form, then audits it against the noncooperative benchmark herself. The ESS-versus-BNE gap that Morozov and Feigel never close, she closes on purpose.
The most useful tool in her paper, for present purposes, is the sophisticated-voting passage. She distinguishes apparent concession — the kind sophisticated voters exhibit when they abandon what cannot win — from real concession, in which her farsighted voters take a genuine short-term loss for an expected long-term gain. Apparent concession is an artifact of representation; real concession requires the agent to value the future in a cardinal, counterfactual way that no purely ordinal solution concept can produce. That is the test Morozov and Feigel’s emergent cooperation needs and never receives: is the cooperation real, in the sense that it would survive an audit against the noncooperative foundations, or is it an artifact of writing strategies as objects rather than as choices?
One more thing about Maggie’s paper, because it is the load-bearing methodological move: she prices her assumption. \(Q\) — the belief over which alternatives will arrive in the future — is the formal analogue of “consistent opponent recognition” in the Morozov-Feigel setup. It is the single substantive thing the model rests on. And she names it, scopes it (“farsighted only with respect to \(Q\)”), flags it as the vulnerability, and identifies the perceived distribution of future considerations as a possibly omitted variable in empirical work. The bill is itemized. The audit is not a footnote; it is the architecture.
The right way to read Morozov and Feigel is in this light. The result is real, the mathematics is beautiful, and the generalized Fisher theorem deserves a wider audience than it has had. The thing they show is genuinely worth showing — opponent-conditioning produces cooperative outcomes under selection without explicit tit-for-tat designed in. But whether what they have studied is the Prisoner’s Dilemma, or a richer game in which the prisoners have been issued Schelling’s toolkit and asked to bargain, is a question worth asking with care.
The cooperative theorist in our household is the one who has spent her career insisting on the noncooperative foundation. I suspect that is not a coincidence.
With that, I leave you with this.
Notes
1 A pedantic note for formal-theory readers: I’m not invoking the Folk Theorem in its strict 1986 Fudenberg-Maskin form, which requires within-pair history, discounted best response, and forward-looking rationality — none of which Morozov and Feigel adopt. The Folk Theorem’s content, though, is what concerns me. Once players can condition behavior on partner identity or history, sustainable outcomes proliferate, and the scientific question shifts from existence to selection. Every iterated-PD result after 1971 has lived in that post-existence terrain. Morozov and Feigel arrive there from a different direction and report it as a discovery about existence. ↩
2 Nowak’s “Five Rules for the Evolution of Cooperation” (Science, 2006) catalogs kin selection, direct reciprocity, indirect reciprocity, network reciprocity, and group selection. Fletcher and Doebeli (Proceedings of the Royal Society B, 2009) showed that all five reduce to mechanisms for generating positive assortment between cooperators. Opponent-conditioning is a sixth route to assortment, not a removal of the assortment requirement — just a different generator of the same thing every cooperation mechanism has to generate. The cost did not vanish. It moved to a different counter. ↩