– by Victor Vikram Odouard
Driving my Fiat 500 convertible through the Swiss countryside surrounding Geneva, a sign caught my eye. In large hand written letters, it read “CAiSSE” – “CHeCKOUT” in French. Behind the sign, a heap of pumpkins lay on a threadbare canvas blanket. But unlike most checkouts, no one was there. And unlike most self-checkouts, there were no cameras in sight. No, this little Hofladen relied on the honor system.
Why wasn’t there a horde of pumpkin bandits raiding our humble farmstand?
Scientists have proposed many mechanisms to explain the evolution of altruistic behavior, such as kin altruism, punishment, and direct & indirect reciprocity. These all share a common characteristic: they provide a second-order benefit to altruists, that compensates for the initial cost.
Most of these mechanisms require knowledge of who is cooperating (I will sometimes call altruistic behavior cooperation in this post). Sometimes that knowledge can come from observation, as is the case with reciprocity (I know who did me a favor in the past). But other mechanisms, like third party punishment and indirect reciprocity, require that knowledge of who has cooperated spread to those who were unlikely to observe the interactions in question. This requires a faithful communication system (read: gossip!) that accurately transmits the relevant information.
And here we arrive at the question we pose in this paper: why should we even have such a communication system, when it might be in people’s interests to lie?
The question echoes the original cooperation question – “why should people benefit society by acting altruistically when it can come at a personal cost?” Here, we simply replace “acting altruistically” with “communicating truthfully”. The parallel points to a solution – perhaps the same mechanism that keeps people cooperative can also keep people communicative.
In our paper, we analyze a model of indirect reciprocity to determine whether such a mechanism can exist. Imagine an agent, Ariel. She meets someone else, Blake, and she can make a choice: she can cooperate or defect. If Ariel cooperates, she pays a cost in order to produce a benefit for Blake. Let’s say she does cooperate. Afterwards, Blake can “gossip” about Ariel’s behavior, in this case, “Ariel good”, since she cooperated. In the next interaction, when Ariel meets Corina, and Corina is deciding whether to cooperate with Ariel, Corina is aware of Blake’s gossip – in this case, that Ariel cooperated. Blake’s gossiping thus produces Ariel’s reputation. Corina can then, for example, decide to cooperate with Ariel since she cooperated with Blake. This is indirect reciprocity.
The twist in our model is that the indirect reciprocity “twists back” on itself: instead of reputations being based solely on how agents act, they can also be based on how agents gossip. How does this work? Returning to our previous scenario, let’s say Ariel cooperates with Blake, but Blake nonetheless says that “Ariel bad”. Perhaps someone, Diana, saw the whole thing go down, and she communicates “Blake bad”, because he lied. In Blake’s next interaction, with Ed, Ed could therefore decide not to cooperate with Blake.
In this scenario, the machinery being used to maintain cooperation (accurate communication) is being co-opted to maintain itself! In our paper, we ask whether such a self-referential system can actually lead to a stable, high-cooperation state, and we found that yes, it could, in the case where (1) agents made occasional (but not too frequent) errors and (2) the norm punishes defection while nonetheless being self-consistent. Both of these conditions will require further explanation.
The first condition reflects an important fact about many complex systems, which is that heterogeneity can enhance stability: for instance, a virus often has a harder time sweeping through a population with many diverse immune responses to its infection, as it must circumvent many different types of defenses. Taking this one step further, if a population has been infected by many different sorts of viruses in the past, it can mount a diversity of different immune responses to any new pathogen – thus, a diversity of past “disturbances” (viruses, in this case), can make a population more robust by diversifying its immune defenses. This is exactly what we see in our model, where error introduces diversity by causing, for instance, some defections, or some untruthful gossip, where they wouldn’t usually occur. We found that this diversity of small-scale disturbances inoculates the system against more serious destabilizing forces.
As for the second condition, a norm in our setting is a determination of whether a given action is good or bad. A gossip system can encode such a norm: for instance, it might call cooperation and truthfulness “good” and defection and untruthfulness “bad”. Agents can then respond to those judgments: for instance, in one of the above examples, Blake was deemed “bad” for having lied about Ariel, and Diana therefore refused to cooperate with him. But a question arises: does the norm judge Diana’s action, which is in effect enforcing the norm (cooperating with those who followed the truthfulness norm, defecting with those who didn’t), as good? In this simple case, no – defection is considered bad. So Diana is judged as “bad” under the very norm that she is enforcing! This norm is therefore not self-consistent – a norm is self-consistent if it does reward the actions enforcing it. A self-consistent norm instead might deem cooperating with “good” agents good, and defecting with “bad” agents good, and everything else “bad”. This norm, called stern judging, is self-consistent, because it rewards the behaviors that enforce it. For instance, there are some societies that consider theft to be permissible if committed against people in bad standing in the community (for instance, people who unjustifiably stole from someone else). This is a self-consistent norm, as the punishment of individuals in bad standing is permissible under the norm’s own rules.
In our analysis, we find that maintenance of cooperation and communication requires such a self-consistent norm. This highlights that for decentralized norm enforcement, self-consistency is critical. If norm enforcement is centralized, self-consistency need not apply, as different rules may exist for judging the centralized body of enforcers and everyone else. For instance, most people judge killing to be bad, but in some countries, this judgment does not apply to a government doling out the death penalty. However, with decentralized norm enforcement, the people who enforce the norms are the people, and therefore no such distinction can exist. Self-consistency is thus needed.
Our study is fundamentally about whether a self-contained norm-enforcement system, not buttressed by anything external, can maintain high levels of cooperation. The external buttress that provoked our study was communication – most indirect reciprocity models assume an already-truthful communication system. But by noticing here that communicating truthfully is not so different from cooperating (in that both provide a public good), we hypothesized that the same indirect reciprocity mechanism enforcing cooperation could enforce truthful communication, without relying on a preexisting communication. The two conditions under which such a state was stable emphasize other important properties of a self-contained norm enforcement system. First, that it can “create its own diversity”, through error, allowing the system to function without being propped up by some external source of diversity. And second, that the norm is able to “reward its own enforcers”, by being self-consistent – this allows enforcement to occur without requiring a separate body of enforcers, operating under different rules.
Read the original article: Odouard, V.V., & Price, M.H. (2023). Tit for tattling: cooperation, communication, and how each could stabilize the other. Evolution & Human Behavior, 44(4), 359-372.