Cheap Talk, Signaling and Claude Code

Cheap Talk, Signaling and Claude Code

scott cunningham

Oct 26

In signaling games, there are usually thought to be three equilibria. There’s the pooling equilibrium, the separating equilibrium, and a hybrid equilibrium. I want to talk about about the first two in the context of Claude Code and the effect that — and all the other advances in AI agents for empirical research over the last year in the social sciences — we’re seeing and the implications that that will have for traditional signals.

In a job signaling game like Michael Spence had created, you start with two types of workers, sometimes represented as the “High” (H) type of worker and the “Low” (L) type of worker. So I will too. We will assume there are only two types of empirical researchers, though, and that one of them is a high type and the other is a low type and the entire population of quantitative researchers is one of the two.

High types produce high quality research that is above their wage. Low types do not. Plus, they look the same to any outsider, which is to say that when someone looks at their resume or papers or what have you, they cannot tell the difference between them.

The firm, though, has to hire one of them. And in this context, they make the wage offer, and the worker accepts it, and when they accept it, the firm pays them the wage, and the worker writes a paper, but the paper is either good or bad and that is not known until later. When the paper is good, the firm is happy and gets some utility measured in money with the letter V, but since they pay the wage with certainty, their payout is V-W. If they hire the high type worker, then V>W and therefore V-W>0 is some kind of “utility profit” that the firm enjoys. But if they are not a high type worker, then they must be a low type worker, and for simplicity we will just say that the paper is worth V=0 and the payout to the firm is -W because the wage must always be paid first.

Now, we all know how these equilibria work, and so I’m just going to jump straight to the punch line. It is a Nash equilibrium (“subgame perfect bayesian Nash equilibrium” specifically) for high and low type workers to both purchase a signal, S, and the firm to hire people with such signals insofar as the signal is sufficiently cheap. And I the case of the pooling equilibrium, that would mean W-S_H>0 and W-S_L>0. In these cases, both workers buy the signal because firms only hire if they have the signal.

But you can also have a separating equilibrium in which only the high type workers purchase the signal, whereas the low type workers abscond, and since the firm only hires people with the signal, it is a subgame perfect bayesian Nash equilibrium for high type workers to sign, low types to not signal, and firms to only hire the signalers (and thus the high types) if the cost of the signal is higher than the wage for the low type but less than the wage for the high type worker.

This is usually thought of as some kind of wasteful, costly action that is visible to the firm but which is relatively easy for the high type worker to do, but for any reason, not easy for the low type worker to do. The cost just has to be different for them, with them each on either side of the value of the wage, and the cost must be in the direction of their own productivity too.

Well, while I cannot easily think of just what I mean by this, I think it’s probably the case that something about writing papers was a costly signal in the past. Something about the writing quality. Something about the identification strategy the author wrote down. Something about the ability to find the data. Something that was visible but not easy to acquire was the cost of the signal. And it was hard for the low type to do it and as a result it did not get got by them, but it was easy for the high type and so they did get it. Firms hired the signalers, since the signalers were high type, the low types were not hired, but they also did not buy signals so it was not an overall loss. It just was a separation of workers into two categories around the information created by signals that helped reveal type to the firm.

I guess where I’m going with this is that the marginal cost of producing empirical research has collapsed. “Collapsed to zero” is what I usually tell people, and while I sometimes get push back, I just don’t see how anyone who has worked seriously with agents can possibly say otherwise. The marginal cost of creating a submission quality manuscript that one could send to a journal is zero. Which is not the same thing as saying it is the best manuscript — though my other guess is I bet you many of them are above the 60th percentile or higher in the distribution of empirical papers written at all anywhere a year go. Maybe even much higher than that.

Well, what exactly can the firm observe now to let them say “Oh, I’m interested in this person because this person did this thing ABC, and only high types can do ABC, so I’m going to interview them.” And then high types do ABC, and the low types don’t do ABC, and the market splits into those who do ABC and those who do not do ABC, and firms only hire the ones that do and simply ignore the ones that don’t.

And it’s just not at all obvious to me anymore what the signals going to be, but my hunch is that it is getting eroded whatever it had been. Interestingly, the signal does not even have to vanish for us to move from separating to pooling equilibrium too. You can really have the cost fall below the wage, everyone signal, firms only hire those who signal, and then it’s stuck in place as a Nash equilibrium.

But the thing — unless we move towards a purely automated reality where AI does all the cognitive work that had been previously done by the quant, data-intensive, researcher, and insofar as that output is highly valued by the firm, then I think it will get worked out. But why do I think that? It’s not as if that always happens. It is entirely dependent on the relative costs for the two types, but the cost must be in the direction of their underlying productivity, and if we are talking about things that AI can now do, then how does that happen? And I don’t really know.

© 2026 scott cunningham
n/a