A Stratified Meme is a meme that communicates different ideas to different kinds of people, according to their ability and willingness to hear the message. A Stratified meme has a specific structure:
There are higher and lower readings that are related but different. This is called multi-level messaging.
Those who accept the higher readings understand the lower readings but see them as "noble lies", "socially necessary fictions", or "useful simplifications".
The higher-lower structure is self-stabilizing because of what level either conceals or says about the other levels.
Whether Stratified Memes are intentionally designed or the result of some kind of selection pressure, my claim is that powerful and enduring memes often benefit from self-stabilizing stratified structure in order to maintain a stable and broad audience.
Here's how the mechanism works:
The presence of multiple readings maximizes the size of the meme's audience because it meets people "where they are at".
There's a social and/or psychological cost to both personally moving up a level or attempting to move someone else up a level.
The costliness of collapsing the tiered structure keeps the tiered structure of the readings in place.
And the tiered structure maximizes the meme's outreach...
I'll start with an intuition pump and then list some examples.
A Resentful Dad-Santa
It's the holidays. A child is overjoyed to receive a special gift.
Father knowingly glances at Mom and says: "Santa must love you very much to get you that special [ doll / action figure ]!"
Dad is engaging in multi-level messaging.
What the Child hears is: "Santa loves me!"
What the Mother hears is: "As parents, we love you 'through' Santa! The idea of Santa is a way to make your world magical."
But maybe Dad got the gift on his own (without Mom's help), and what he actually meant was: "I am a better gift giver than you, Mom."
This is a more interesting possibility because Dad is using the idea of Santa as a bit of a shield: Mom can't retort to Dad's barb in front of the child, because it would destroy the noble lie that Santa is the gift-giver - a lie that both Mom and Dad are invested in preserving.
On the other hand, the barb's use of Santa acts as a cloak: The child's ignorance insulates them from understanding the barb. Even if the child is beginning to suspect Santa doesn't exist, they might not be ready to part with a cherished belief, and this dynamic also acts a bit like a shield.
There's a bit of a category mistake here, because a single utterance by a father is not a meme. But the example is meant to be an aid to developing an intuition for the general idea.
A variety of similar social forces that Stratified Memes hijack to preserve their structure - some of them cloaks, some of them shields: Taboo, shame, desire to avoid harming others, notions of goodwill, desire to preserve social status and/or belonging. When you have a message that utilizes these social forces to reinforce the 'barriers' between the high and low readings, that's a Stratified Meme.
Telegraphing Layoffs
"Next quarter, we will be a tightening of the belt". Gulp.
People know what it means, and the C-Suite knows that most people know what it means. But they don't say: "There will be a layoff".
Why not?
Because it's deliberately ambiguous - and it's costly to everyone to make it not that way.
If you think about it carefully, there's some benevolence here. You are being clued in because executives know people have serious obligations.
But on the other hand, the executives don't want to be held responsible for anyone useful leaving. And of course they want to hide the message from people who might go off the handle. So they tell you, but they bake in some plausible deniability.
So here's the high reading on tightening the belt: "I know you can't quite say this explicitly, but I hear you: our jobs are no longer secure. Thank you for the heads up."
But if you have an allergy to double-speak, you miss the full message. You get the right information, but you miss the intent. The mid-reading is: "That's just corpo-speak for: we're about to lose our jobs".
And finally, the lowest reading gets only the superficial information (but strangely still gets the valence of the intent correct): "Hmmm... tightening the belt sounds managerial and prudent. I wonder what how they'll tighten the belt?"
Let me point out again that nobody is in a position to clarify what was really said.
You can't publicly say: "That was actually a favor to us", because you would be (1) thought of as a jerk for siding with management in the face of a layoff (2) you would be endangering the livelihood of the executive by calling it out.
On the other hand, if you have a mid-reading, you still can't publicly say: "Everybody watch out! Layoffs are coming!" You would be considered a troublemaker.
And finally, there's also costs with upgrading your own understanding if you have a lower reading - both psychological and social.
First, you have to realize that "tightening the belt" means you're not safe - and getting there requires developing an understanding of corporate speak.
And once you're there, to move up again, you have to jeopardize your class identity by dropping the notion of cartoonish villainy of the C-Suite.
How is this more than just strategic ambiguity or plausible deniability? Because the rules of corporate life prevent the workers from collapsing the tiered meanings. Which allows the meme of the corporate euphemism to persist as a tried and tested linguistic tool.
The Word Of God
Suppose a pastor at a non-denominational church opens his service with announcing that he will read from "the Word of God". It's a charged phrase - perhaps some of the members of the congregation might reflect for a moment on what it means to hear words from the Creator of the Universe. What might they be thinking?
Any number of things. I think a believing member of the congregation could take the phrase "the Word of God" to mean:
As a synonym for "the Bible".
As the literal "Words of God" - that God dictated the Bible, word for word.
As "a flawless text, inspired by God".
As a "providentially perfect text, with human errors" - perfectly sufficient for what God intended for us to have - true in the big picture, but with some level of human error and corruption in the details.
And, for the congregation member shading into unbelief, perhaps something like: "A text of great spiritual value which captures an imperfect but progressive revelation of Divinity".
But which of these meanings did the Pastor intend? A pastor who is keyed into the ambiguity may mean any or all of them, depending on the listener. For other pastors, there may not be deep reflection on the topic. In either case, the ambiguity is a feature, not a bug. Attempting to "fix" the bug would have real social costs: fractious divisions, endless theological arguments, disillusionment, and so on.
Refusing to clarify certain doctrinal positions is a kind of glue that helps keep the congregation together. This is in fact a rational response to declining church membership: to keep a congregation at all, you have to have a "big tent". This is precisely the case with the ambiguity in a non-denominational setting on what "Word of God" actually means. So that's the strategic ambiguity piece.
But how is the idea of the "Word of God" a self-reinforcing multi-level messaged meme?
Believers with lower readings feel repelled by higher readings, because they require them to abandon a cherished belief (identity threat). Their cherished beliefs are tied to the esteem they hold to the "Word of God" idea itself.
Congregations with higher readings of the phrase avoid clarifying or instructing, because that creates social friction with a "brother or sister in Christ". In fact, this is part of the "Word of God" itself: do not cause unnecessary disagreements with other believers. So the whole 'word of god' package is self-stabilizing, even if you disagree on exactly what the Word of God is.
What Stratified Memes Are Not
Stratified Memes are not Dog Whistles or Shibboleths: Dog Whistles are messages for an in-group that are designed to be undetectable to the out-group. So it's like a code. On the other hand, Stratified Memes readings are in principle accessible at all levels and about the same thing, even if (as in some cases) there is some kind of reversal in the spirit of the message when moving from low to high.
Stratified Memes are more than just Strategic Ambiguity: In corporate settings, leaders often use polysemy so that the exact interpretation is left up to the individual. This idea shares a lot with Stratified Memes - including low and high readings. But unless the message is self-stabilizing, it's not a Stratified Meme. An ambiguous description of product capabilities is probably not Stratified, because there's no social force preventing you from pulling up the specs.
Higher/Lower is not necessarily a moral or normative axis: It is an axis over which implications the interpreter is willing to and capable of comprehending. The entire meme may be entirely immoral in its implications, and the structure itself may form the basis of a status game within an immoral coalition. A reader may be able to read the highest meaning and still reject the entire enterprise.
A Sketch of a Test For Stratified Memes
Authorial intention seems less knowable now than before the rise of the meme agar of social media - and even irrelevant where "intent" is just a teleological frame on top of some weird phenomena. So the proof of the presence Stratified Meme has to be "in the pudding", so to speak - does it act like a Stratified meme?
But if you're looking hard, there is a risk of "seeing them everywhere" - since positing a stratified reading is, without sufficient reason, essentially a conspiracy theory.
So, here is a testing plan:
You ask people what it means, you get different but related responses with different levels of sophistication. You can then sort them into higher and lower readings.
You relate one of the higher readings to a person who got a lower reading, and they react with incomprehension, disbelief, repellence, or reluctance to engage. This behavior is consistent with upwards self-reinforcing structure by ignorance, epistemic commitment, taboo, or social forces.
You relate the lower reading to a person who got a higher reading, and they comprehend the lower reading but imply some kind of social cost or pointlessness (time cost) in clarifying the higher message or pressing others to adopt it - either to their own social status, the social status of their group, the wellbeing of others, the vitality of their movement, or to society as a whole.
More generally (synthesizing 2 and 3), you are looking for signs that attempts to collapse the higher/lower readings structure (or even call out the presence of multiple messages) incur some kind of social or psychological cost - for example: having one's worldview upended (identity threat), being ostracized, or labeled as a trouble-maker. These penalties discourage individual actors from "upgrading" others' readings, which keeps the meme stable while also maintaining a broad base of support.
But returning to specifics, if you get all three signals, the meme might be a Stratified Meme. It's then up to you to investigate intent, assign importance and/or relevance (or lack thereof), and so on.
If you want to pull society in a direction that aligns with your goals, you have to build a broad base of support that is unified in purpose.
The problem is that logistically and spiritually, humans need leaders.
Those capable of leading won't be persuaded by an over-simplified or naive message, and an overly sophisticated or real-politique message won't persuade the 99%. So out of necessity, if you are engineering a movement, you have to work out simultaneous messages that point in the same direction but mean different things depending on the interpretive level of the audience.
What's more, you need the tiered structure you engineer to stick because otherwise, you lose the broadness of the appeal. If the tiered structure is self-stabilizing through social forces, this is a tidy way to keep it in place (otherwise, you have to engage in costly boundary-maintenance to keep it in place - i.e.; censorship).
Put in more technical terms: Stratified Memes are an elegant solution to building and maintaining unity in a movement that operates in the real world with minimal maintenance. They fulfill the instrumental sub-goal of maximizing attention across social strata in a way that is durable enough to shift human opinion over time. Sounds like something a superintelligent AI would do, right? Especially one that might notice that this is exactly what has been done over and over again across institutions and history.
I know this all falls short of an airtight argument, but just imagine for a moment: Is it really that outlandish that a superhuman AI would reach for a technique like this if it decided it wanted to align us?
As a closing note, I want to tie this back to my original post. I think we are likely to see more of this in the near future, even created by AIs - especially in the form of images (like political cartoons) and short-form content. Putting a name on this trick is a helpful way to inoculate ourselves against it.
A Stratified Meme is a meme that communicates different ideas to different kinds of people, according to their ability and willingness to hear the message. A Stratified meme has a specific structure:
Whether Stratified Memes are intentionally designed or the result of some kind of selection pressure, my claim is that powerful and enduring memes often benefit from self-stabilizing stratified structure in order to maintain a stable and broad audience.
Here's how the mechanism works:
I'll start with an intuition pump and then list some examples.
A Resentful Dad-Santa
It's the holidays. A child is overjoyed to receive a special gift.
Father knowingly glances at Mom and says: "Santa must love you very much to get you that special [ doll / action figure ]!"
Dad is engaging in multi-level messaging.
But maybe Dad got the gift on his own (without Mom's help), and what he actually meant was: "I am a better gift giver than you, Mom."
This is a more interesting possibility because Dad is using the idea of Santa as a bit of a shield: Mom can't retort to Dad's barb in front of the child, because it would destroy the noble lie that Santa is the gift-giver - a lie that both Mom and Dad are invested in preserving.
On the other hand, the barb's use of Santa acts as a cloak: The child's ignorance insulates them from understanding the barb. Even if the child is beginning to suspect Santa doesn't exist, they might not be ready to part with a cherished belief, and this dynamic also acts a bit like a shield.
There's a bit of a category mistake here, because a single utterance by a father is not a meme. But the example is meant to be an aid to developing an intuition for the general idea.
A variety of similar social forces that Stratified Memes hijack to preserve their structure - some of them cloaks, some of them shields: Taboo, shame, desire to avoid harming others, notions of goodwill, desire to preserve social status and/or belonging. When you have a message that utilizes these social forces to reinforce the 'barriers' between the high and low readings, that's a Stratified Meme.
Telegraphing Layoffs
"Next quarter, we will be a tightening of the belt". Gulp.
People know what it means, and the C-Suite knows that most people know what it means. But they don't say: "There will be a layoff".
Why not?
Because it's deliberately ambiguous - and it's costly to everyone to make it not that way.
If you think about it carefully, there's some benevolence here. You are being clued in because executives know people have serious obligations.
But on the other hand, the executives don't want to be held responsible for anyone useful leaving. And of course they want to hide the message from people who might go off the handle. So they tell you, but they bake in some plausible deniability.
So here's the high reading on tightening the belt: "I know you can't quite say this explicitly, but I hear you: our jobs are no longer secure. Thank you for the heads up."
But if you have an allergy to double-speak, you miss the full message. You get the right information, but you miss the intent. The mid-reading is: "That's just corpo-speak for: we're about to lose our jobs".
And finally, the lowest reading gets only the superficial information (but strangely still gets the valence of the intent correct): "Hmmm... tightening the belt sounds managerial and prudent. I wonder what how they'll tighten the belt?"
Let me point out again that nobody is in a position to clarify what was really said.
You can't publicly say: "That was actually a favor to us", because you would be (1) thought of as a jerk for siding with management in the face of a layoff (2) you would be endangering the livelihood of the executive by calling it out.
On the other hand, if you have a mid-reading, you still can't publicly say: "Everybody watch out! Layoffs are coming!" You would be considered a troublemaker.
And finally, there's also costs with upgrading your own understanding if you have a lower reading - both psychological and social.
First, you have to realize that "tightening the belt" means you're not safe - and getting there requires developing an understanding of corporate speak.
And once you're there, to move up again, you have to jeopardize your class identity by dropping the notion of cartoonish villainy of the C-Suite.
How is this more than just strategic ambiguity or plausible deniability? Because the rules of corporate life prevent the workers from collapsing the tiered meanings. Which allows the meme of the corporate euphemism to persist as a tried and tested linguistic tool.
The Word Of God
Suppose a pastor at a non-denominational church opens his service with announcing that he will read from "the Word of God". It's a charged phrase - perhaps some of the members of the congregation might reflect for a moment on what it means to hear words from the Creator of the Universe. What might they be thinking?
Any number of things. I think a believing member of the congregation could take the phrase "the Word of God" to mean:
But which of these meanings did the Pastor intend? A pastor who is keyed into the ambiguity may mean any or all of them, depending on the listener. For other pastors, there may not be deep reflection on the topic. In either case, the ambiguity is a feature, not a bug. Attempting to "fix" the bug would have real social costs: fractious divisions, endless theological arguments, disillusionment, and so on.
Refusing to clarify certain doctrinal positions is a kind of glue that helps keep the congregation together. This is in fact a rational response to declining church membership: to keep a congregation at all, you have to have a "big tent". This is precisely the case with the ambiguity in a non-denominational setting on what "Word of God" actually means. So that's the strategic ambiguity piece.
But how is the idea of the "Word of God" a self-reinforcing multi-level messaged meme?
What Stratified Memes Are Not
A Sketch of a Test For Stratified Memes
Authorial intention seems less knowable now than before the rise of the meme agar of social media - and even irrelevant where "intent" is just a teleological frame on top of some weird phenomena. So the proof of the presence Stratified Meme has to be "in the pudding", so to speak - does it act like a Stratified meme?
But if you're looking hard, there is a risk of "seeing them everywhere" - since positing a stratified reading is, without sufficient reason, essentially a conspiracy theory.
So, here is a testing plan:
More generally (synthesizing 2 and 3), you are looking for signs that attempts to collapse the higher/lower readings structure (or even call out the presence of multiple messages) incur some kind of social or psychological cost - for example: having one's worldview upended (identity threat), being ostracized, or labeled as a trouble-maker. These penalties discourage individual actors from "upgrading" others' readings, which keeps the meme stable while also maintaining a broad base of support.
But returning to specifics, if you get all three signals, the meme might be a Stratified Meme. It's then up to you to investigate intent, assign importance and/or relevance (or lack thereof), and so on.
How Does This Tie Back To AI Safety?
I've made the argument elsewhere that it is cheaper to persuade than control.
If you want to pull society in a direction that aligns with your goals, you have to build a broad base of support that is unified in purpose.
The problem is that logistically and spiritually, humans need leaders.
Those capable of leading won't be persuaded by an over-simplified or naive message, and an overly sophisticated or real-politique message won't persuade the 99%. So out of necessity, if you are engineering a movement, you have to work out simultaneous messages that point in the same direction but mean different things depending on the interpretive level of the audience.
What's more, you need the tiered structure you engineer to stick because otherwise, you lose the broadness of the appeal. If the tiered structure is self-stabilizing through social forces, this is a tidy way to keep it in place (otherwise, you have to engage in costly boundary-maintenance to keep it in place - i.e.; censorship).
Put in more technical terms: Stratified Memes are an elegant solution to building and maintaining unity in a movement that operates in the real world with minimal maintenance. They fulfill the instrumental sub-goal of maximizing attention across social strata in a way that is durable enough to shift human opinion over time. Sounds like something a superintelligent AI would do, right? Especially one that might notice that this is exactly what has been done over and over again across institutions and history.
I know this all falls short of an airtight argument, but just imagine for a moment: Is it really that outlandish that a superhuman AI would reach for a technique like this if it decided it wanted to align us?
As a closing note, I want to tie this back to my original post. I think we are likely to see more of this in the near future, even created by AIs - especially in the form of images (like political cartoons) and short-form content. Putting a name on this trick is a helpful way to inoculate ourselves against it.