What Does “Signalling” Mean?
[Epistemic status: shortly after writing this post, I thought I’d regret it. It was not rated very highly by LW karma, and on outside view, it’s the sort of rant I often don’t endorse later on. But re-reading it after a few weeks, I think it holds up. I still endorse everything I’ve written here, with the exception of my suggestion that “virtue signalling” is an appropriate replacement for what most people mean by “signalling”.]
I still feel a strong empathy for the post You Can’t Signal to Rubes, which called out LessWrong for using the word “signalling” incorrectly. That post got heavily, and rightly, downvoted because it also got the definition wrong. :( But it had a point!
At the time of writing, the current definition of signalling on the LessWrong tag is:
Signaling is behavior whose main purpose is to demonstrate to others that you possess some desirable trait. For example, a bird performing an impressive mating display signals that it is healthy and has good genes.
I’m not even sure I should correct it, because this does seem to summarize the LessWrong consensus on what signalling means. But we already have a term for signalling desirable properties about yourself: virtue signalling! Maybe you’ll object that “virtue signalling” isn’t quite right. Ok. But, could you find another word? I would prefer for “signalling” to point to the subject of signalling theory, which I understand to be the game theory of communication (often focusing on evolutionary game theory).
Scott Alexander’s What Is Signaling, Really? seems to get most things right:
In conclusion, a signal is a method of conveying information among not-necessarily-trustworthy parties by performing an action which is more likely or less costly if the information is true than if it is not true. Because signals are often costly, they can sometimes lead to a depressing waste of resources, but in other cases they may be the only way to believably convey important information.
Although all of his examples are about signalling self-properties, he never stipulates that, instead always using the more general conveying-information definition. He also avoids the signalling is automatically bad pitfall. Instead, he explains that signalling is often unfortunately costly, but is nonetheless a very useful tool.
However, reading it, I’m not sure whether he means to contrast signalling with “mere assertion”, or whether he considers assertion to be a kind of signalling:
Life frequently throws us into situations where we want to convince other people of something. If we are employees, we want to convince bosses we are skillful, honest, and hard-working. If we run the company, we want to convince customers we have superior products. If we are on the dating scene, we want to show potential mates that we are charming, funny, wealthy, interesting, you name it.
In some of these cases, mere assertion goes a long way.[...]
In other cases, mere assertion doesn’t work.
[...]
I’ll charitably assume that he meant both cases to be types of signalling. But for anyone who was mislead by the wording: signalling is the theory of conveying information! Mere assertions, if they carry information, count as signalling!
So, to summarize the points I’ve raised so far:
Sometimes people talk like signalling is just the bad thing (the dishonest or not-maximally-honest practice of making yourself look good).
Relatedly, people tend to exclude “mere assertion” from signalling, making signaling and literal use of language mutually exclusive.
Often people restrict signalling to signalling facts about yourself. (In fact, often restricted to status signalling.)
To be honest, I’m not even sure academic uses of the term “signalling” avoid the “mistakes” I’m pointing at! The Wikipedia article Signalling (economics) currently begins with the following:
In contract theory, signalling (or signaling; see spelling differences) is the idea that one party (termed the agent) credibly conveys some information about itself to another party (the principal).
[Note that I’ve defaulted to the Wikipedia spelling of signalling; spelling on LessWrong seems mixed.]
On the other hand, the page on Signalling Theory (a page which is very biology-focused, despite the broader applicability of the theory) includes examples such as alarm calls (eg, birds warning each other that there is a snake in the grass). These signals cannot be interpreted as facts about the signaller.
Perhaps it is a quirk of economics which restricts the term “signalling” to hidden information about the agent, and LessWrong inherited this restricted sense via Robin Hanson?
I don’t understand the objection.
Virtue signalling is a subset of signalling. Specifically, it is signalling of moral virtues.
Therefore, a bird signalling health and good genes is not virtue signalling (but it is signalling in general). Because health and good genes are usually not considered to be moral virtues.
I think Scott got this right, but you misunderstood it.
X is a signal of Y if seeing X makes Y more likely. In some cases, mere assertions do that, in some cases, they don’t.
For example, saying “I read Less Wrong” is a signal of reading Less Wrong, because people who read Less Wrong are more likely to say that they read Less Wrong. However, saying “I am not a criminal” is not a signal of not being a criminal, because criminals also say it a lot.
It’s not about what the words mean, it’s about what they correlate with. Sometimes the act of speaking the words correlates with their literal meaning (not lying, or lying rarely). Sometimes the act of speaking the words has almost zero correlation with their literal meaning (lying almost always).
.
I agree with your third objection, that Less Wrong uses signalling in the narrower sense (about agent), because that is how Robin Hanson typically uses it, and most of us were probably introduced to the concept by him.
(I am not sure whether Robin never used signalling in the wider sense, or he did but we just didn’t notice.)
I agree that virtue signalling is specific to moral virtues rather than signalling virtues-in-general such as good genes, and I made a mistake by suggesting otherwise. My objection was really the part you agree with later, that people use signalling to refer exclusively to signalling properties of the self (and also use it predominantly to refer to duplicitous or otherwise negative cases, rather than to the general thing). But I don’t know what word to give people for what they’re trying to mean. “Virtue signalling” was my attempt.
Too bad I don’t have a good suggestion for a way to improve the terminology.
I also agree that mere assertions of X don’t necessarily signal X. I neglected to mention that because it wasn’t my focus in that part of the essay.
Oddly enough, the Signaling tag is currently awaiting a merge between the description you quoted, and a description imported from the old wiki, which links to and quotes the Scott Alexander post you referenced. (You can see both in the edit history; it looks a bit weird because there’s a revision written in the context of the tagging system before we imported stuff, which was written in isolation, but presented as though it were a diff relative to the imported, older wiki page.)
Signaling theory as a term in economics or game theory usually refers to the analysis of situations where an agent takes an action that transmits information that some other agent (or rather, the “principal”) does not have, and which influences the principal’s behavior. The agent is also often called the sender and the principal the receiver of the signal.
Often, this is information about the agent, but sometimes it is information about something else, so we can generally just say it is information about “the state of the world” or “the state of nature”. Usually, signaling theory is concerned with situations in which the information cannot be transmitted by “mere assertion” (or “cheap talk”, see below) but only by a costly action, and the cost of transmitting information about certain states of the world has to be different from transmitting information about other states of the world in certain ways; e.g. in Spence’s job-market signaling model, low-ability workers must have a higher cost of attaining education than high-ability workers, otherwise low-ability workers would also do it and the signal is worthless. (Note that in these models, the agent moves first and the principal second, but still the principal offers a contract based on the received information. If the principal moves first and offers a contract to the informed agent, we are in contract theory. Signaling theory and contract theory together are sometimes referred to as “information economics”, “economics of asymmetric information”, or sometimes the “theory of incentives”.)
Situations in which there are no such costly signals are usually called “cheap talk” models. Of course, if there is no conflict of interest, the informed party can always just transfer the information (and there would also be no need for costly signals then). But suppose there is a conflict of interest between the informed sender and the uninformed receiver. Then which kind of information is transmittable? The seminal paper is by Crawford and Sobel. They show that, basically, very fine-grained information transmission does not work when there is a conflict of interest.
Finally, if a sender can send costless credible signals but can strategically choose which ones, we are in the domain of “Bayesian persuasion” models.
(If you can send signals that are costless and there is no conflict of interest, then we are maybe back in basic statistical theory if the signals are noisy, but I guess there is no room for an economic analysis.)
These two domains seen to capture the twin aspects of my favorite signaling topic, flirtation. “Cheap talk” is where you try and display your attractive qualities. “Bayesian persuasion” is where you reveal information, often non-verbally, in order to get someone to escalate, back off, change their tactics, and so on.
I think some of the confusion with signaling arises from the fact that the labels aren’t very intuitive and that these two domains are easy to either conflate, or focus on one at the expense of the other.
Sorry, I think that’s a misunderstanding. I will edit the part about cheap talk.
Perhaps there should be a place illustrating the differences between LW terms and academic terms of the same name—besides this there’s also game theory and I feel like I’ve seen others as well (though I can’t remember what they were).
“Metaethics” is another example of this; sometimes it gets used around here to mean “high-level normative ethics” (and in fact the rationalist newsletter just uses it as the section header for anything ethics-related).
This is a hopelessly narrow definition and should be changed. An agent can signal anything about itself, including undesirable traits such as “I’m not edible” or “I’m batshit crazy violent, don’t mess with me”. So lets first lose the clause “desirable”.
And signals can be costly to the sender but not necessarily. The cost (or more precisely, lower cost to honest signallers and higher cost to pretenders) makes a signal more trustworthy in the cases where the sender may have a motive to cheat. In biology, low cost signalling is still signalling. Mothers’ signals to offspring tend to be pretty low cost for example. The offspring trust their mothers anyway and are incentivized to respond as intended; they have not much to gain from distrusting the mother. Babies’ signals to mothers may be costlier in cases when the mother is likely to abandon some offspring based on their fitness. E.g. possibly human babies pack on so much fat and look so plump so as to signal their high fitness to moms (human mothers are known to selectively abandon infants, unlike chimp mothers, whose infants are not so plump). The lower fitness infants are incentivized to look like they had high fitness, so mothers use a hard-to-fake trait such as fat content to assess fitness.
“I’m batshit crazy violent” obviously has a cost because being violent exposes you to aggression by the ones you attack, but I don’t think it involves the competition element like is present in the previous example. Consider a Northern lapwing singlehandedly attacking a hawk many times bigger than itself: she’s not intending to actually fight the hawk, she’s signalling to the hawk that she’s absolutely crazy and will fight to the death; that, of course, would be uncomfortable and painful to the hawk, so the hawk pretty reliably leaves, when it sees the lapwing coming. (Although sometimes lapwings are still killed by hawks). The signalling behavior here is approaching fast in a menacing manner, but lapwings are not competing with each other and the hawk is not trying to assess which lapwing is most violent; this signal is risky (costly) but it’s not subject to runaway competition and not therefore likely to accelerate into depressingly wasteful levels. It can still be perhaps modelled as a handicap signal like the plump infant example, because the costliness of the behavior is a feature not a bug: the hawk can trust the lapwing’s signal because a lapwing who’s not really willing to risk a fight would not dare to approach a hawk.
The alarm signals to one’s flock (“there’s a predator in the grass”) were already mentioned as not signalling anything about the sender; alarm signals are also costly if they attract the predator to the signaller, but here the cost is a bug that cannot be avoided, it’s not a feature of the signal. (Of course, one may combine alarm signals with status signalling, e.g. in some birds the ones who sound most alarms are the most dominant and respected birds in the group because alarm calling is dangerous and only the fittest birds can afford to do it a lot.)
In conclusion, signalling as used in ecology or economy is so much more diverse than in the quoted definition, I don’t see why it should be so narrow in the Wiki. The narrow meaning may be captured by some other term, such as “fitness signals”, “signals of high quality”, or suchlike.
Wait, this is not the example in the Wikipedia page, which is actually “When an alert bird deliberately gives a warning call to a stalking predator and the predator gives up the hunt, the sound is a signal.”
I found this page which gives a good definition of signaling:
So in the alarm example, the quality k is whether the bird has really detected the predator, and the “action” is for the predator to give up the hunt. Later in the Wikipedia article, it says “For example, if foraging birds are safer when they give a warning call, cheats could give false alarms at random, just in case a predator is nearby.”
Indeed, to me ‘signalling’ is doing some action which is differentially costly depending on whether some fact is or isn’t true—so mere assertion doesn’t count, even if it conveys information.
That’s not what “virtue signalling” means.
Communication under adversarial conditions? i.e. you expect noise with an incentive to ape signal.
But a bird warning other birds that there is a snake in the grass does reveal a fact about the bird: that it is a good bird who will risk its life for the other birds, and therefore is helpful for the other birds to keep around, perhaps even with bribes.
If ‘virtue signaling’ has the wrong connotations, maybe ‘Hansonian signaling’?
But the alarm call is a fact about the signaller—it conveys the fact that it is aware of the predator. For a bird to give the alarm call without being aware of a predator, it would have to do it a lot more often, wasting precious time and energy. Dishonest signal is more expensive than honest signal because a bird that sounded the alarm because it noticed a predator can minimize the alarm time and get back to whatever it was doing as soon as the predator is gone.
There are actually two types of alarm calls: those addressed to one’s group members and those addressed to the predator. The ones addressed to group members don’t necessarily convey meaningful information about the sender (at least this is not the immediate purpose, although these signals can in some cases have an additional benefit of demonstrating ones high quality to group members, see my other comment).
The alarm calls addressed to the predator are thought to signal the sender’s vigilance: “you’ve been spotted, we know you are here, you can’t catch us”. See mobbing. These calls often result in the predator politely taking a leave.
It does not follow for me, say “yelling” is a type of “talking” this we all agree.
but when I say “I am yelling” I am giving more/different information than just “I am talking”
sure mere “assertion” is a type of “signalling”
but for me when we use “signalling” I infered that we are doing things differently than “asserting”
is this what you mean?