(Note: After reading this post, check out A Theory of Laughter—Follow-Up for some additional ideas and elaborations.)

1. tl;dr

There should be parallel explanations for laughter at two levels.

• At the brain level, there should be some mechanism / algorithm that produces laughter, and it should fit the data of when people laugh in practice.
• At the evolution level, there should be some explanation for why this mechanism exists in the first place. Why was it adaptive in our ancestors? And where did it come from—are there homologues in other animals?

I’ll summarize my proposals for both of these, in the opposite order:

1.1 First half of the tl;dr: Laughter in terms of evolution

I endorse the popular theory that laughter is an indicator of “play”, homologous to the play-related vocalizations and body language in other animals (e.g. the dog’s “play bow”).

• The evolutionary purpose of play is “practice for future dangerous situations”. For example, a wolf pup that engages in play-fighting and play-chasing would presumably be more skilled in its future real-life fights and chases.
• The evolutionary purpose of innate communicative play signals, like laughter in humans and play-bows in dogs, is to reduce the probability of accidental escalation from practice to serious. For example, if a play-fight between two wolf-pups escalates into a real fight between the pups, that’s dangerous for both pups. If the pups are emitting and responding to communicative play signals, then that kind of escalation is much less likely to happen. It’s kinda the same idea as “safewords” in fight-related sports (among other places).

1.2 Second half of the tl;dr: Laughter in terms of brain algorithms

My (oversimplified) pseudocode brain “business logic” for laughter is something like:

Indeed, I expect that there is some genetically-specified neuron group in the hypothalamus or brainstem (or more generally, what I call the Steering Subsystem), and that when future scientists look at its various connections and their functional properties, it will be straightforwardly obvious that this neuron group and its connections are implementing the pseudocode above.

(Side note: These scientists will also find that this neuron group has various other inputs that make laughing more or less likely on the margin—inputs related to mood etc.—which I omitted from the box for simplicity.)

Note that nothing in this box is particularly tied to humans. If we’re talking about 50kHz rat laughter instead of human laughter, I wouldn’t change a single word in the box above. However, later in the post, I will talk about human laughter in particular, including humor, and I’ll argue that this pseudocode box is a plausible match to the circumstances in which people laugh.

Also, the path by which I initially came to guess this pseudocode box (namely, introspection) was independent of how I came to believe the evolutionary story (namely, I read it in a book and it seemed obviously right). But I claim that the two stories match up beautifully—that the pseudocode box above is the natural, straightforward way to implement the “spec” associated with the evolution story, given background constraints about how I think brain algorithms work in general (see Section 3.3.2 below). That reassures me that I’m on the right track.

OK, that was the tl;dr. The rest of the article will elaborate on that picture, why I currently believe it, and broader implications.

1.3 Table of contents with section summaries

• Section 2 will explain the evolution story in more detail.
• Section 3 will explain the brain algorithms story in more detail—in particular, what exactly do I mean by that pseudocode above? And then I’ll go through three main reasons that I think it's right: (1) the pseudocode matches the evolutionary “spec”; (2) the pseudocode is highly plausible on the neuroscience side; and (3) the pseudocode seems to match the situations in which humans laugh (and in which animals emit analogous play-signals).
• Section 4 elaborates on the latter by fleshing out how to reconcile the pseudocode with everyday experience, in three domains:
  • Section 4.1 asks: How does this pseudocode shed light on laughter in physical play—tickling, chasing, peek-a-boo, water balloon fights, etc.? For example, why can’t you tickle yourself?
  • Section 4.2 asks: How does this pseudocode shed light on conversational laughter? For example, how is it that laughter can communicate so many different things in different contexts (e.g. friendliness versus aggression, or sincerity versus insincerity)?
  • Section 4.3 asks: How does this pseudocode shed light on humor and jokes?
• Section 5 asks: How exactly is this pseudocode implemented in the brain? I hypothesize that there are innate connections between neuron groups in the hypothalamus and/or brainstem that directly correspond to the pseudocode above. Alas, I cannot tell you exactly which neuron group it is—I think it’s a neuron group that nobody has studied yet. But I think I know what we should be looking for and where, and I think that a neuroscience lab could figure this out in the near future using standard experimental methods.
• Section 6 is the conclusion, in which I will discuss how this post relates to my job as an Artificial General Intelligence safety / alignment researcher.

2. The evolutionary story

2.1 What is play, and why do animals have an innate play drive?

A central example of play in my mind is two young animals play-fighting or play-chasing each other.

Why are the young animals doing this? There seems to be an obvious primary explanation:

• If a squirrel has spent many hours play-fighting and play-chasing other squirrels as a pup, it will presumably do better at actual-fighting and actual-chasing other squirrels as an adult.
• And likewise, if a squirrel has spent many hours running away from its sibling during play-chase, it will presumably also do better when running away from a predator.

In other words, play is practice for dealing with future dangerous situations.

And what is the evolutionary advantage of having an innate play drive? My answer is: Obviously, squirrel pups do not have the foresight and knowledge to deduce from first principles that it’s a good idea to spend some of one’s free time practicing for future dangerous situations. So instead, play is an innate drive.

In other words: Evolutionarily, play is a means to an end, but from the squirrel’s within-lifetime perspective, play is its own reward—it’s just intrinsically enjoyable.

2.2 What is an innate communicative play signal, and why do animals have them?

Play signals are any vocalizations or body language that happen primarily or exclusively when the animal is playing.

One example is the “play bow” in dogs (which often comes with a high-pitched “play bark”):

Another is rat laughter, a 50 kHz (ultrasonic) chirp emitted by rats under various circumstances including play-fighting among young rat pups. Jaak Panksepp and his then-student Jeffrey Burgdorf first hypothesized that these chirps are the rat version of laughter in the late 1990s; see for example this 2000 paper by them, which involved manually tickling rats.

(I know what you’re thinking—but don’t worry! Today’s scientists no longer have to suffer the indignity of tickling rats by hand. Instead they can use an “automated protocol for tickling in which … rats are forced to move and interact with a constantly rotating rod”.)

There has been a good deal of follow-up work investigating rat-laughter, some of which I will be referencing below.

So anyway, other animals have play signals. Why should humans be different? And more specifically, our chimp cousins emit laughter-like panting sounds, and “laugh most when tickled, during rough-and-tumble play, and during chasing games (the chimp being chased laughs most)” (that quote and much more in chapter 5 of Provine’s book).

(See: YouTube of a young chimp getting tickled.)

So we arrive at the theory that laughter is a play-signal for humans, homologous to the play signals in other animals. I think this theory goes back to at least Darwin; SEP mentions Max Eastman in 1936. For my part, I originally heard this theory from a chapter in the book Elephant In The Brain by Kevin Simler and Robin Hanson. (I found that book chapter very enlightening, although I don’t agree with everything they said.^[1])

Why do animals have innate communicative play signals? Imagine two young squirrels play-fighting. This is a mutually-beneficial activity, for reasons mentioned above. However, it is also possible for two young squirrels to actually fight, an activity which is very dangerous for both of the squirrels, and sometimes also for kin bystanders.

And now we see the problem: there is an obvious resemblance between play-fighting and actual-fighting—such that a play-fight could accidentally escalate to an actual fight. Therefore, each squirrel benefits from communicating to the other that it is play-fighting.

Thus, we expect animals to have evolved innate mechanisms to emit play signals, along with corresponding innate mechanisms to notice those signals and react to them.^[2]

3. The brain story

3.1 The pseudocode

I’ll re-copy the box at the top for easy reference—I think it’s something like this:

There are presumably various other things that modulate this circuit (a.k.a. shift the thresholds)—person-to-person variation, and variation with age (and certainly with species), and dependence on other innate signals (e.g. angry people tend to laugh less). But I think the box above is the main story.

3.2 What exactly do I mean by “evidence that I’m in danger” etc.? How are these things operationalized?

To be clear, I am not talking about consciously believing that I’m in danger, or safe, etc. For example, my fight-or-flight reaction can easily activate even when I consciously believe I have nothing to worry about, like during a scary movie.

Instead I am talking about innate signals in the hypothalamus and/or brainstem—what I call the “Steering Subsystem”. These signals (I claim) have particular innate “meanings” / algorithmic purposes with legible relation to ecological / homeostatic requirements. For example, see my discussion here of a particular group of neurons in the hypothalamus that is activated by a physiological need for food, and causes various downstream effects (like energy conservation and hunger sensations) that are appropriate to that state. These kinds of innate signals can exist even if we have no conscious (interoceptive) access to them, and no common English-language concept that perfectly aligns with them. (See my recent post on where I disagree with Lisa Feldman Barrett.)

Anyway, I suggested in the box above that “danger” might be operationalized via some of the same innate signals that activate the sympathetic nervous system (directly or indirectly), and “safe” might be operationalized via some of the same innate signals that activate the parasympathetic nervous system (directly or indirectly). Which signals exactly? I don’t know.

What about ingredient (C), the “social” innate signals? I’m not sure about that either. But there do seem to be innate signals in the hypothalamus and brainstem that are intended (by evolution) to correlate with being in a social situation. One obvious way for evolution to calculate such signals is via heuristics calculated in brainstem sensory processing systems. For example, a recent preprint I like—Liu et al. (2023)—identified two groups of cells in the hypothalamus (medial preoptic nucleus). The activity of one group correlated with being socially isolated, and the activity of the other group correlated with the end of that isolation. Interestingly, they found that, as far as these two cell groups were concerned, “social” was being operationalized via touch sensations—not sound, not pheromones, it had to be touch. (The mice were blind so they’re not sure about vision.) But that finding of the centrality of touch does not generalize to other social instincts. For example, another rodent social instinct is mating, and here, the innate circuits are triggered at least in part by pheromones. Anyway, in sum, I don’t know exactly how “social” is calculated for the purpose of Ingredient (C) of laughter in rodents, let alone in humans, but such a calculation is definitely a thing that the brain can do.

3.3 Three reasons I like this proposal

3.3.1 This pseudocode matches the evolutionary “spec” of Section 2

I think the correspondence here is strong and straightforward, and I see this as key evidence that I’m on the right track. Going through the items above:

• (A)—some evidence that I’m in danger. If I’m in a situation that is triggering some fight-or-flight bodily reactions—regardless of the exact reason—then I’m almost definitely in a situation that resembles dangerous situations that I might get into in the future. And therefore, being in this situation almost definitely constitutes good “practice”.
• (B)—stronger evidence that I’m safe. Without this ingredient, it’s not “practice”—it’s the real thing! I should be getting out of the situation, not rushing into it! I should find it aversive, not fun.
• (C)—evidence that I’m in a social situation. Without this ingredient, it still qualifies as “practice”, and thus it is still evolutionarily adaptive to stay in this situation, and therefore our evolutionary expectation should be that a situation would be “fun” even without ingredient (C). But without ingredient (C), I don’t expect to laugh, evolutionarily speaking. There is no point in emitting a communicative signal, if there is no one around to hear it.

I put in ingredient (C) based on various studies (in both humans and rats) that much more laughter occurs when conspecifics are around. For example, Provine found that laughter was 30× less frequent when people were alone.^[3] (I think watching TV by oneself is an intermediate case—the TV presumably tricks our brains into thinking we’re in a social situation, at least to some extent.)

So for example, my theory would be: If I go on a roller coaster with friends, we may well be laughing together during the scary parts; whereas if I go for a ride on an otherwise-empty roller coaster, with nobody in eyeshot or earshot, then I’m much less likely to laugh out loud, but I’ll still probably find it fun. Likewise, if I’m alone, I might want to read a humorous book, even if doing so will probably not make me laugh out loud.

3.3.2 This kind of pseudocode is highly plausible on the neuroscience side

Unfortunately, despite some effort, I cannot tell you the exact neurons that implement the pseudocode in the box above. See Section 5 below for more details.

However, the pseudocode above is fully compatible with everything I think I know about the brain. And that counts as strong evidence for me, because "the things I think I know about the brain" are highly constraining! (For some of those constraints, see my post “‘Learning from Scratch’ in the Brain”. And see my post on social instincts for an example of how those constraints rule out lots of possibilities in practice.)

In particular, I know that the right kind of signals are in fact present in the hypothalamus & brainstem (see Section 3.2 above), and I know that the genome is easily capable of building a little cluster of neurons in (probably) the hypothalamus that perform the requisite logical operations. I have seen lots of little clusters of neurons in the hypothalamus doing broadly this kind of genetically-specified business logic.

3.3.3 This pseudocode seems to fit the data of when humans laugh (and when animals emit analogous play-signals)

As far as I can tell, the pseudocode is compatible with all our experience of laughter in humans and corresponding play-signals in other animals, including all three of the categories elaborated in Section 4 below—physical play in Section 4.1, non-“humor” conversational laughing in Section 4.2, and humor in Section 4.3.

You can read Section 4 to better understand where I’m coming from. But then you should take a long hard look at the pseudocode box in Section 3.1 above, and then scroll down to the comments section, and complain about some situation where the predictions of that pseudocode box are wrong!

…And then I’ll respond with things like:

• “OK sure, but that’s because the pseudocode box of Section 3.1 is oversimplified, like I left out how the laughter-reaction is suppressed by other innate signals correlated with anger and mating and some other things like that.”
• “OK sure, but I was using words like ‘safe’ and ‘dangerous’ when obviously those are just the closest English-language words I can think of, not perfect descriptions, and they come apart from the actual innate signals in various ways, which by the way I cannot specify in detail”.

…And then you will roll your eyes at me and accuse me of special pleading and unfalsifiability.

Totally fair!

But I still currently think I’m getting much more out of this pseudocode box than I’m putting in—and that’s even without knowing exactly where the corresponding neurons are and what they’re connected to.

(But please do leave those kinds of skeptical comments! They will be very appreciated!)

4. Relating the pseudocode to everyday experience

There isn’t a sharp line between them, but I’ll separately discuss three categories: laughter in physical play, laughter in non-“humor” conversation, and humor. In that order:

4.1 Laughter in physical play (e.g. tickling, chasing, peek-a-boo, etc.)

(This is the most straightforward case, and also the case where humans are most similar to other animals.)

4.1.1 What are the sources of “Ingredient (A)” (i.e., evidence of danger / cause for physiological arousal) in physical-play laughter?

In physical-play laughter, I think there are many possible sources of Ingredient (A), and they’re mostly pretty obvious. In particular, we (like almost all animals) have a diverse suite of innate defensive reactions, including in our case:

• Startle and orienting responses to unexpected sensory inputs
• Various types of defensive flinches in anticipation of physical trauma
• Some innate reaction to the sense that you’re falling down
• Running away from threats (“escape response”)
• Etc.

All of these innate reactions trigger not only certain muscle behaviors, but also physiological arousal—hence Ingredient (A).

And all of these are situations where you’ll see little kids laughing uproariously during physical play.

4.1.2 More discussion of tickling in particular

It strikes me as extremely obvious that tickling is part of play-fighting. For example, the most ticklish parts of your body seem to coincide with the parts that are most vulnerable to serious injury, like the front of the neck.^[4] So the evolutionary story behind tickling is straightforward. What about the brain-level story?

I claim that the pseudocode of Section 3.1 above is perfectly adequate to explain everything about tickling at the brain-level. To flesh that out, here are some additional details and discussion:

The source of Ingredient (A) in tickling is basically the normal startle and orienting responses to unexpected sensory inputs, as mentioned above. But, for obvious evolutionary reasons, those innate reactions are presumably especially strong for touch sensations on vulnerable parts of your body. Think of the circumstances when those reactions would naturally trigger—maybe there’s a scorpion crawling on your neck, or maybe your enemy has successfully gotten his hands around your neck during a fight, etc.

You can’t tickle yourself for the same reason that you do not routinely have an involuntary startle reaction from the sound of your own voice when you start talking, and the same reason that you do not routinely have an involuntary orienting reaction when you wave your own hand in front of your own eyes. (Unless of course you have schizophrenia—see here—or if you are an infant surprising yourself by making a funny noise for the first time, etc.)

The source of Ingredient (B) in tickling is the knowledge that you’re safe among trusted friends. If you’re getting “tickled” in a situation where you’re genuinely terrified for your own safety, my impression is that you’re going to be screaming rather than laughing.

Related to this, even if someone enjoys getting tickled in general, they will still try to push your hand away from the most ticklish areas. The evolution-level reason for this behavior is obvious: that’s part of how play-fighting works. It would hardly be good defense practice otherwise! But what about the brain-level reason? I propose that when their sensitive areas are successfully getting tickled, Ingredient (A) gets so strong that it cuts off Ingredient (B), tipping the experience from fun into aversion.

4.2 Laughter in non-“humor” conversation

4.2.1 Background: Most conversational laughter is not “humor”

Even if we ignore physical play, the connection between “humor”^[5] and laughter is less tight than you might think. In Provine’s book he talks about his “ecological” studies of modern USA people hanging out in public, and says “only about 10%-20% of [comments immediately preceding somebody laughing] were estimated by my assistants to be even remotely humorous”. Then follows a helpful table of typical laugh-eliciting comments, including such zingers as “I’ll see you guys later!” and “Can I join you?”

Pay close attention next time you’re in (or overhearing) a normal face-to-face group conversation, and you’ll probably notice something similar.

Other cultures seem to be similar to the USA in that respect—for example this article talks about hunter-gatherers laughing in response to gentle teasing and such—not “why did the chicken cross the road”.

4.2.2 What are the sources of “Ingredient (A)” (i.e., evidence of danger / cause for physiological arousal) in conversational laughter?

Section 4.1.1 listed a bunch of obvious ways that arousal gets invoked in physical play, like if you think you’re alone but someone suddenly jumps out at you from hiding while screaming at you. But what about conversation—can mere words invoke physiological arousal too? Yes, obviously! Wandering into a stressful conversation topic can get your heart rate up just as surely as can wandering into a swarm of angry bees.

To be more specific, it seems to me that there are a great many natural sources of Ingredient (A) in conversations, including:

• Confusion
• Embarrassment / guilt / shame (possibly vicarious)
• Disgust (possibly vicarious)
• Surprise (possibly vicarious)
• Threats (both bodily threats and status threats) (possibly vicarious)
• Taboo-breaking

These are not mutually-exclusive, and there are others too.

4.2.3 Laughter communicates a pretty universal message about my transient “internal” state, but then the listener has to infer why I feel that way, and the latter inference is complicated, contextual, and widely-varying.

Remember from Section 2 above, I’m claiming that “avoiding accidental escalation during practice for future dangerous situations” is the evolutionary explanation of why the brain mechanism for laughter is there in the first place. But given the existence of that brain mechanism, it will be active in lots of situations, many of which may have nothing to do with “avoiding accidental escalation during practice for future dangerous situations”. (See “Adaptation-executors, not Fitness-maximizers”.) And I think this is much more true in humans than in other species, who seem to mostly just laugh during play-fighting, play-chasing, etc.

In all situations, I think the universally-shared “meaning” of laughter is related to certain signals in the head of whoever is laughing, as specified by the pseudocode of Section 3.1. But then the listener needs to infer from context why those signals are in the laugher’s head. And here things get very complicated and contingent. Some examples:

• Scenario 1: You are walking by me in middle school, when I just dropped my contact lens. I say “please help me” with a desperate tone, and you laugh as you say “sorry, I don’t help losers”.
• How I might take that: I know that you feel safe (ingredient (B)), and I might guess that it’s because I’m in your outgroup and that you regard my suffering as no threat to your own well-being. I also know that you feel a bit of ingredient (A), and maybe I’ll guess that you feel like it’s cringe that I even dared ask you for help in the first place. In short, your laughter communicated to me that you feel unsympathetic and superior.
• Scenario 2: You are my spouse. I say “Man, my company is really dysfunctional”, while slightly laughing, and in response you laugh as you say “yeah”.
• How I might take that: Well, for my part, I was laughing mainly because I felt kinda annoyed about the dysfunction (Ingredient (A)), but also not overly worried for my own sake (Ingredient (B)). Then when you laugh in your response, I might infer that you are invested in my well-being, and empathetically mirroring my feelings on both counts. In short, your laughter communicated to me that you feel empathy and camaraderie.

I could go on. I think that, in different contexts, laughter can signal friendship, or animosity, or superiority, or inferiority, or sincerity, or insincerity, etc. etc. There is no simple theory, because we can feel a certain way for many unrelated reasons. 

4.2.4 …And given that laughter is able to communicate stuff, people skillfully wield laughter as part of their communicative toolkit

In the previous section I was implicitly treating laughter as an incidental side-effect of the emotions that someone feels in the course of a conversation. But once humans have learned (consciously or unconsciously) that laughter communicates things, they will start wielding laughter to skillfully advance their communicative intentions. For example, in Scenario 2 just above, maybe you laughed in part because you wanted to communicate empathy and camaraderie.

This doesn’t have to be a conscious explicit decision or desire, and in fact it probably usually isn’t. It’s probably more often an unconscious habit—in lots of previous conversations, you’ve laughed or not-laughed in a certain way in a certain context, and it led to good results, so you unconsciously learned to repeat that behavior next time you’re in a similar situation.

While purposeful control of laughter is not directly part of the “business logic” I wrote down in Section 3.1, we obviously can in fact voluntarily laugh. Mechanistically, I think it typically (though not always^[6]) happens indirectly—if we want to laugh, we steer ourselves into a transient emotional state that has Ingredients (A-C), and if we want to not laugh, we steer ourselves into a transient emotional state that doesn’t have all three ingredients. How do we do that? Well, we have some control over our transient emotional state because we can attend to some aspects of our situation rather than others, choose which frames / analogies to mentally invoke, etc.

4.3 Humor

As mentioned in 4.2.1 above, humor does not necessarily cause people to laugh, and most laughter occurs in the absence of humor.

That said, humor can obviously lead to laughter, so I ought to briefly say something about how. I don’t have a grand theory of humor, nor do I think there is one, beyond what I’ve already said in this post. I’ll just mention a few considerations that I find helpful to keep in mind when thinking about humor and jokes.

4.3.1 What are the sources of “Ingredient (A)” (i.e., evidence of danger / cause for physiological arousal) in humor?

I think the list is pretty similar to Section 4.2.2 above for conversational laughter above. I won’t re-copy it—you can scroll up. Again, that list is not exhaustive, nor mutually exclusive.

4.3.2 There’s an inverted-U dynamic for “Ingredient (A)”

According to the pseudocode box above, if there is too little of Ingredient (A), there’s no laughter (e.g. a boring conversation), and if there’s too much (A), then there’s also no laughter, because it undermines Ingredient (B) (e.g. it might just feel stressful, painful, scary, confusing, etc.) Somewhere in between is optimal for laughter.

So I think we wind up with inverted-U dynamics like this:

(I was implicitly talking about this same inverted-U in Section 4.1.2 above, when discussing why someone might enjoy getting tickled a little bit, but find direct tickling of their most ticklish spots to be too much, and thus aggressively push the tickler away.)

4.3.3 The (somewhat arbitrary and drifting) cultural expectations / tropes / rituals surrounding “humor” can be a key component of the explanation of why something is funny.

In particular, if a listener has a general cultural expectation that “humor” should involve X (e.g. a punchline), then a joke-teller can either:

• Increase the amount of Ingredient (B) by conspicuously including X (e.g. an obvious punchline), OR
• Increase the amount of Ingredient (A) by conspicuously excluding X (e.g. punchline-free absurdist anti-humor)

Either of these can be helpful for the joke, depending on how much (A) and (B) are present from other sources.

And if the latter (exclusion of X) happens a lot, then we collectively stop expecting X in the first place, causing gradual (anti-inductive^[7]) drifts in what people find funny, or schisms between humor-subcultures.

4.3.4 I’m not generally impressed by “theories of humor” beyond their overlap with the discussion above.

For example, the Psychology of Humor textbook (Martin & Ford 2018) describes three “classic theories of humor” which I will comment on in turn:

• “Relief theory” seems to capture the kernel-of-truth that many instances of humor follow the following time-course: first, we have a bunch of Ingredient (A), and second, something changes and introduces a heap of Ingredient (B) (while the (A) has not yet fully faded from our brainstems). Those two steps can be described as “tension” and “relief of tension” respectively.
• “Superiority theory” seems to capture the kernel-of-truth that, in many instances of humor, Ingredient (A) is vicarious—from imagining someone in danger or duress—while Ingredient (B) comes from my comfort in the knowledge that I, the listener, am not that person, and am superior to that person, and have therefore nothing to fear for myself.
• “Incongruity theory” strikes me as a mishmosh of different things. For example, Ingredient (A) can be based on the feeling of confusion, and (B) by its resolution. Or Ingredients (A) and (B) can come from different (incongruous) ways to view the same situation, one of which is normal / safe and the other embarrassing / dangerous / etc. In still other cases, I wonder whether we’re all just following a cultural telling-a-joke script, and as a listener, Ingredient (A) is my concern that I don’t “get the joke” and will be embarrassed to admit it, and Ingredient (B) is my relief when I do.^[8] It can also be several of these things simultaneously, and more.

After that, the textbook moves on to discuss three “contemporary theories of humor” (“reversal theory”, “comprehension-elaboration theory”, and “benign violation theory”). My comments on those would largely overlap with the above bullet points, so I’ll just move on.

5. More detailed discussion of the neuroscience

5.1 Overview

The kind of “business logic” pseudocode of Section 3.1 is to be found, I claim, in what I call the “Steering Subsystem” (hypothalamus and brainstem—see definition and discussion here). My guess is more specifically as follows:

5.2 Where exactly in the brain is the “laugh behavior controller” (top box in that diagram) where I can read out the alleged pseudocode of Section 3.1?

Sadly, I have failed to deduce from existing literature where in (probably) the hypothalamus we would find the core “business logic” pseudocode of Section 3.1. (It could also be split among a couple places.)

Focusing on rats (although I expect the answer to be similar in humans), one candidate mentioned in the literature is the so-called “parvafox nucleus” of the lateral hypothalamus. Check out Alvarez-Bolado & Celio (2016) for an argument along those lines. As far as I can tell, the strongest evidence in favor of this hypothesis is that bilateral destruction of the parvafox nucleus dramatically (factor of >10) reduces rodent laughter, according to Roccaro-Waldmeyer et al. (2016). Relatedly, there is some evidence (including from laughter-inducing “gelastic seizures”) that “stimulations of [the] tuberal portion [of the lateral hypothalamus] provoke bursts of laughter” (Alvarez-Bolado & Celio (2016)), and that’s in the general vicinity of parvafox.

However, I think the current balance of evidence is that parvafox is not the controller for laughter, but rather is related to defense behavior—based on both direct stimulation of those cells (e.g. Cola et al. (2023)), and looking at where in the brain they project to (e.g. Celio et al. (2013), Bilella et al. (2016)). We still need to explain the Roccaro-Waldmeyer results from the previous paragraph, but there are a couple possibilities for that, including (1) that parvafox is essential for “Ingredient (A)” of the pseudocode of Section 3.1 (and thus upstream of laughter), or (2) that parvafox is physically proximate to the “real” laugh behavior controller and that Roccaro-Waldmeyer destroyed the latter accidentally in their experiments. (Lesion experiments are notorious for “collateral damage” of nearby neurons and fibers, if I understand correctly, but be warned that I’m not an expert.)

Well, if it’s not parvafox, then what is it?

I don’t know. If someone wanted to make progress on this question experimentally—to actually find that pseudocode implemented via innate brain signaling pathways—I think an obvious immediate next step would be a retrograde neural tracing experiment starting from the (lateral) part of periaqueductal gray (PAG) associated with laughter (as pinpointed in the very recent article Gloveli et al. (2023)), followed by further characterization of whatever upstream neuron-groups show up. Maybe such data already exists, and I missed it. (Parvafox does not seem to project to the correct part of PAG to match up with the Gloveli et al. neurons, as far as I can tell, although I’m not super-confident.)

Areas that I see as especially suspicious include:

1. neurons nearby but not part of the parvafox nucleus, in the tuberal region of lateral hypothalamus (for reasons stated above);
2. neurons somewhere in or around the medial preoptic area of the hypothalamus—for which stimulation can cause rat laughter (Wintick & Brudzynski (2001)), and which is known to be the home of the main behavior controllers for at least two other positive social behaviors that I know of, namely the behavior studied by Liu et al. (2023) mentioned above, plus aspects of mating as discussed in Bayless et al. (2023);
3. more generally, the rest of the hypothalamus too.

5.3 What about the “Learning Subsystem” (cortex, striatum, amygdala, hippocampus, etc.)?

As I described here, I claim that the brain should be divided into two subsystems, the “Learning Subsystem” which runs randomly-initialized learning algorithms, and the “Steering Subsystem” which runs “business logic”. Above I was only talking about the “Steering Subsystem”; but in fact laughter also interacts with the “Learning Subsystem”—i.e., the cortex, striatum, amygdala, hippocampus, thalamus, cerebellum, etc.

I claim that  the laughter-triggering signals coming from the Learning Subsystem are in the following two categories:

• (A) Outputs trained by reinforcement learning to maximize some signal from the Steering Subsystem (hypothalamus & brainstem) that acts as a “reward”, or
• (B) “Short-term predictors” of some signal from the Steering Subsystem.

For example:

• An example of (A) would be voluntary / learned control of laughter (Section 4.2.4).
• An example of (B) would be a “self-fulfilling prophecy”, where you laugh because you’re in a situation that pattern-matches to other situations that have caused you to laugh in the past.

I think there are probably other examples too, but that’s outside the scope of this post.

Direct evidence that the cortex is not necessary for play is provided by “decorticate” rats (= rats whose cortex has been surgically removed), whose play is “much the same [as] controls” (Whishaw 1990, Panksepp et al. 1994). I can’t immediately find direct measurements of the presence of the normal 50kHz “laughing” vocalizations in decorticate rats while they play, but if their play is similar in other respects, I would certainly guess that they are also vocalizing in a roughly normal way.

6. Conclusion

I remain adamant that the hypothalamus and brainstem are full of hundreds-to-low-thousands of specific neuron groups with specific connections that are all written directly into the genome, and which correspond to “business logic” that makes evolutionary sense—things like “if you’re malnourished, reduce your sex drive”. I think that understanding this tangle of “business logic” is annoying but possible, and that doing so seems possibly helpful for Artificial General Intelligence (AGI) safety and alignment, for reasons spelled out here.

I consider laughter an “easy” example of such business logic, as compared to something like the human innate status drive (assuming that there is a human innate status drive, which I currently believe but am not 100% sure). I think the human innate status drive would need substantially more convoluted pseudocode than the box in Section 3.1, and I don’t even have a plausible guess right now for how it works in detail. (See here for vague thoughts.) So maybe this post on laughter is a warm-up. In particular, I don’t think this post is directly important for AGI safety—if we iron out the details of the pseudocode of Section 3.1, and put it into the source code of future AGIs, and then we find that those AGIs are laughing in a vaguely-psychopathic-human-like way while they mercilessly murder me and my family and every other human … then that doesn’t really make me feel much better!

Still, even if understanding laughter is just a “warm-up” for more safety- and alignment-relevant things like compassion and friendship, I am still very interested in getting this right! I spent much longer on this blog post than usual (admittedly a low bar), including trying to be reasonably comprehensive in reading relevant sources, etc., and I am very eager for feedback. 

(Thanks Seth Herd, Linda Linsefors, Justis Mills, and Miguel De Guzman for critical comments on drafts.)

1. ^{^}

   Since the works of Robin Hanson are popular on this forum, I will say a bit more about where I differ from Elephant In The Brain. My biggest complaint is the part where they say:

   As we mentioned earlier, people are profoundly ignorant about laughter’s meaning and purpose (at least in our default state, before learning the science). But where does this ignorance come from? Why does introspection fail us so spectacularly here?

   It’s not simply because laughter is involuntary, outside our conscious control. Flinching, for example, is also involuntary, and yet we understand perfectly well why we do it: to protect ourselves from getting hit. Thus our ignorance about laughter needs further explanation.

   I disagree that it “needs further explanation”. I think we start out ignorant of literally everything, until we learn it / figure it out. And I think that figuring out the evolutionary purpose of laughter is just inherently much harder than figuring out the evolutionary purpose of flinching. It’s less obvious / salient, for various reasons that I claim are pretty obvious if you think about it. I don’t think there’s any more to it than that.

   I also don’t think there can be more to it than that. To explain what I mean by that, imagine if I said: “Here’s the source code for training an image-classifier ConvNet from random initialization using uncontrolled external training data. Can you please edit this source code so that the trained model winds up confused about the shape of Toyota Camry tires specifically?” The answer is: “Nope. Sorry. There is no possible edit I can make to this PyTorch source code such that that will happen.” By the same token, even if, as that book argues, there is a strong evolutionary pressure to make humans specifically confused about the evolutionary purpose of laughter, I don’t think there is any possible genetic change that would make that happen. Related discussion here.

2. ^{^}

   Whenever there’s a claim about communicative signals, one can ask a follow-up question of whether these signals are game-theoretically stable against “lies”. Like, what if Squirrel A laugh-squeaks to signal play, then Squirrel B lets down its guard, then Squirrel A attacks for real? If this happened a lot, wouldn’t squirrels eventually evolve to ignore play signals? And if that happened, then wouldn’t squirrels further evolve to stop emitting play signals in the first place? Good question! It’s worth thinking about those kinds of things. But I do think a good answer exists, even if I don’t know it in full detail. I think there are various ways that signals can be hard or costly to fake. And I also think that animals treat a play-signal by itself as insufficient reason to let down one’s guard, which reduces the benefit of lying. Animals are reacting to other cues too, like whether there is an existing trusting relationship. For example, if I’m a prisoner, and the cruel guard is pointing at me and laughing, that sure wouldn’t make me feel more relaxed.

3. ^{^}

   I think Provine’s claim that there’s 30× more laughter in social situations is less clear-cut than it sounds. In fact, it’s hard to do an apples-to-apples comparison of laughter in social versus nonsocial situations, because (1) social situations are drawn from a different distribution from nonsocial situations in many respects, (2) the presence of other people can change Ingredients (A) and (B) too, and (3) social situations can also affect laughter via voluntary control (Section 4.2.4). I still think (C) is probably a real ingredient in the algorithm, because Provine’s factor of 30 is so extreme that it seems difficult to fully explain by any of those indirect mechanisms.

4. ^{^}

    I’m extremely far from an expert on what parts of the body are vulnerable in combat today, let alone 100,000 years ago (How often were people punching each other, versus slashing with sharp rocks, versus getting bitten by wolves or spiders? What were the lived consequences of different types of injuries? Beats me!). But I still think this claim is probably true. For example, this article claims that ticklish areas correlate with areas protected by involuntary defensive reflexes.

5. ^{^}

   I am defining the word “humor” narrowly—something like “humor = the kind of stuff you find in the ‘humor’ section of a bookstore”, e.g. recognizable jokes and so on. On this definition, if someone walks up to my lunch table and says “can I join you?” and then laughs a bit, that’s conversational laughter but not “humor”. I believe Provine takes this definitional approach.

   Alternatively, one could define the word “humor” very broadly, as something like “humor = whatever makes someone spontaneously laugh”. On this definition, it’s perfectly possible that someone walking up to my lunch table and saying “can I join you?” is an instance of “humor”, when taken with its full context. I believe Hurley et al. take this definitional approach.

   Anyway, this is just semantics. “Humor” is just a word, and within limits, we can define it however we want. I find the narrow definition to be helpful for the purposes of this post, so that’s how I’m using it.

6. ^{^}

   Transient manipulation of our own emotional state is not the only method of learned / deliberate manipulation of laughter. Alternatively, you can laugh by pure voluntary motor control—just move your larynx, lungs, etc. and make laugh sounds, the same way you might voluntarily move your arm. I think such laughter can come across as fake sometimes (like maybe it can sound slightly different, and the eyes don’t squint in quite the same way, see Duchenne smile)—although people still do it plenty.

   Incidentally, I think the “summoning transient emotions” method of laughing and the “pure voluntary motor control” method of laughing in this footnote are probably two opposite ends of a spectrum, as opposed to a crisp binary.

7. ^{^}

   If you haven’t heard the term “anti-inductive”: "Inductive" reasoning is where you assume that if you’ve seen something a bunch of times in the past, then it’s likely to happen again in the future. “Anti-inductive” reasoning is the opposite—the more past evidence you have for something, the more likely it is to be false next time.

   As the old joke goes: “I recommend anti-inductive reasoning to anyone. After all, using anti-inductive reasoning has been a catastrophically bad idea every time I’ve tried it in the past. So it’s definitely a good idea going forward.”

   Here’s an example. If you’re up against a master at rocks-paper-scissors, you might want to use anti-inductive reasoning about the future: The more evidence you have for a pattern in your opponent’s behavior, the likelier it is that your opponent wants you to notice that pattern, and therefore the more you should expect that pattern to reverse on the next throw. (But this is happening at every level of abstraction simultaneously—which is kinda weird to think about.)

8. ^{^}

   I suspect that kids will laugh a comparable amount in practice when solving a riddle / brainteaser posed by a friend as when “getting” a pun told by the same friend, other things equal. Other things are not always equal though; puns often get an extra emotional “kick” from some other source, like from being sexual, or from the friend’s mood, or just from the very fact that it’s a pun, both because there’s a cultural trope that making puns is a shameful activity, and because puns are “supposed” to be funny and the expectation of laughter can be a self-fulfilling prophecy, see Section 5.3 below.