Internal Family Systems (IFS) is a psychotherapy school/technique/model which lends itself particularly well for being used alone or with a peer. For years, I had noticed that many of the kinds of people who put in a lot of work into developing their emotional and communication skills, some within the rationalist community and some outside it, kept mentioning IFS.
So I looked at the Wikipedia page about the IFS model, and bounced off, since it sounded like nonsense to me. Then someone brought it up again, and I thought that maybe I should reconsider. So I looked at the WP page again, thought “nah, still nonsense”, and continued to ignore it.
This continued until I participated in CFAR mentorship training last September, and we had a class on CFAR’s Internal Double Crux (IDC) technique. IDC clicked really well for me, so I started using it a lot and also facilitating it to some friends. However, once we started using it on more emotional issues (as opposed to just things with empirical facts pointing in different directions), we started running into some weird things, which it felt like IDC couldn’t quite handle… things which reminded me of how people had been describing IFS. So I finally read up on it, and have been successfully applying it ever since.
In this post, I’ll try to describe and motivate IFS in terms which are less likely to give people in this audience the same kind of a “no, that’s nonsense” reaction as I initially had.
Epistemic status
This post is intended to give an argument for why something like the IFS model could be true and a thing that works. It’s not really an argument that IFS is correct. My reason for thinking in terms of IFS is simply that I was initially super-skeptical of it (more on the reasons of my skepticism later), but then started encountering things which it turned out IFS predicted—and I only found out about IFS predicting those things after I familiarized myself with it.
Additionally, I now feel that IFS gives me significantly more gears for understanding the behavior of both other people and myself, and it has been significantly transformative in addressing my own emotional issues. Several other people who I know report it having been similarly powerful for them. On the other hand, aside for a few isolated papers with titles like “proof-of-concept” or “pilot study”, there seems to be conspicuously little peer-reviewed evidence in favor of IFS, meaning that we should probably exercise some caution.
I think that, even if not completely correct, IFS is currently the best model that I have for explaining the observations that it’s pointing at. I encourage you to read this post in the style of learning soft skills—trying on this perspective, and seeing if there’s anything in the description which feels like it resonates with your experiences.
But before we talk about IFS, let’s first talk about building robots. It turns out that if we put together some existing ideas from machine learning and neuroscience, we can end up with a robot design that pretty closely resembles IFS’s model of the human mind.
What follows is an intentionally simplified story, which is simpler than either the full IFS model or a full account that would incorporate everything that I know about human brains. Its intent is to demonstrate that an agent architecture with IFS-style subagents might easily emerge from basic machine learning principles, without claiming that all the details of that toy model would exactly match human brains. A discussion of what exactly IFS does claim in the context of human brains follows after the robot story.
Wanted: a robot which avoids catastrophes
Suppose that we’re building a robot that we want to be generally intelligent. The hot thing these days seems to be deep reinforcement learning, so we decide to use that. The robot will explore its environment, try out various things, and gradually develop habits and preferences as it accumulates experience. (Just like those human babies.)
Now, there are some problems we need to address. For one, deep reinforcement learning works fine in simulated environments where you’re safe to explore for an indefinite duration. However, it runs into problems if the robot is supposed to learn in a real life environment. Some actions which the robot might take will result in catastrophic consequences, such as it being damaged. If the robot is just doing things at random, it might end up damaging itself. Even worse, if the robot does something which could have been catastrophic but narrowly avoids harm, it might then forget about it and end up doing the same thing again!
How could we deal with this? Well, let’s look at the existing literature. Lipton et al. (2016) proposed what seems like a promising idea for addressing the part about forgetting. Their approach is to explicitly maintain a memory of danger states—situations which are not the catastrophic outcome itself, but from which the learner has previously ended up in a catastrophe. For instance, if “being burned by a hot stove” is a catastrophe, then “being about to poke your finger in the stove” is a danger state. Depending on how cautious we want to be and how many preceding states we want to include in our list of danger states, “going near the stove” and “seeing the stove” can also be danger states, though then we might end up with a seriously stove-phobic robot.
In any case, we maintain a separate storage of danger states, in such a way that the learner never forgets about them. We use this storage of danger states to train a fear model: a model which is trying to predict the probability of ending up in a catastrophe from some given novel situation. For example, maybe our robot poked its robot finger at the stove in our kitchen, but poking its robot finger at stoves in other kitchens might be dangerous too. So we want the fear model to generalize from our stove to other stoves. On the other hand, we don’t want it to be stove-phobic and run away at the mere sight of a stove. The task of our fear model is to predict exactly how likely it is for the robot to end up in a catastrophe, given some situation it is in, and then make it increasingly disinclined to end up in the kinds of situations which might lead to a catastrophe.
This sounds nice in theory. On the other hand, Lipton et al. are still assuming that they can train their learner in a simulated environment, and that they can label catastrophic states ahead of time. We don’t know in advance every possible catastrophe our robot might end up in—it might walk off a cliff, shoot itself in the foot with a laser gun, be beaten up by activists protesting technological unemployment, or any number of other possibilities.
So let’s take inspiration from humans. We can’t know beforehand every bad thing that might happen to our robot, but we can identify some classes of things which are correlated with catastrophe. For instance, being beaten or shooting itself in the foot will cause physical damage, so we can install sensors which indicate when the robot has taken physical damage. If these sensors—let’s call them “pain” sensors—register a high amount of damage, we consider the situation to have been catastrophic. When they do, we save that situation and the situations preceding it to our list of dangerous situations. Assuming that our robot has managed to make it out of that situation intact and can do anything in the first place, we use that list of dangerous situations to train up a fear model.
At this point, we notice that this is starting to remind us about our experience with humans. For example, the infamous Little Albert experiment. A human baby was allowed to play with a laboratory rat, but each time that he saw the rat, a researcher made a loud scary sound behind his back. Soon Albert started getting scared whenever he saw the rat—and then he got scared of furry things in general.
Something like Albert’s behavior could be implemented very simply using something like Hebbian conditioning to get a learning algorithm which picks up on some features of the situation, and then triggers a panic reaction whenever it re-encounters those same features. For instance, it registers that the sight of fur and loud sounds tend to coincide, and then it triggers a fear reaction whenever it sees fur. This would be a basic fear model, and a “danger state” would be “seeing fur”.
Wanting to keep things simple, we decide to use this kind of an approach as the fear model of our robot. Also, having read Consciousness and the Brain, we remember a few basic principles about how those human brains work, which we decide to copy because we’re lazy and don’t want to come up with entirely new principles:
There’s a special network of neurons in the brain, called the global neuronal workspace. The contents of this workspace are roughly the same as the contents of consciousness.
We can thus consider consciousness a workspace which many different brain systems have access to. It can hold a single “chunk” of information at a time.
The brain has multiple different systems doing different things. When a mental object becomes conscious (that is, is projected into the workspace by a subsystem), many systems will synchronize their processing around analyzing and manipulating that mental object.
So here is our design:
The robot has a hardwired system scanning for signs of catastrophe. This system has several subcomponents. One of them scans the “pain” sensors for signs of physical damage. Another system watches the “hunger” sensors for signs of low battery.
Any of these “distress” systems can, alone or in combination, feed a negative reward signal into the global workspace. This tells the rest of the system that this is a bad state, from which the robot should escape.
If a certain threshold level of “distress” is reached, the current situation is designated as catastrophic. All other priorities are suspended and the robot will prioritize getting out of the situation. A memory of the situation and the situations preceding it are saved to a dedicated storage.
After the experience, the memory of the catastrophic situation is replayed in consciousness for analysis. This replay is used to train up a separate fear model which effectively acts as a new “distress” system.
As the robot walks around its environment, sensory information about the surroundings will enter its consciousness workspace. When it plans future actions, simulated sensory information about how those actions would unfold enters the workspace. Whenever the new fear model detects features in either kind of sensory information which it associates with the catastrophic events, it will feed “fear”-type “distress” into the consciousness workspace.
So if the robot sees things which remind it of poking at hot stove, it will be inclined to go somewhere else; if it imagines doing something which would cause it to poke at the hot stove, then it will be inclined to imagine doing something else.
Introducing managers
But is this actually enough? We’ve now basically set up an algorithm which warns the robot when it sees things which have previously preceded a bad outcome. This might be enough for dealing with static tasks, such as not burning yourself at a stove. But it seems insufficient for dealing with things like predators or technological unemployment protesters, who might show up in a wide variety of places and actively try to hunt you down. By the time you see a sign of them, you’re already in danger. It would be better if we could learn to avoid them entirely, so that the fear model would never even be triggered.
As we ponder this dilemma, we surf the web and run across this blog post summarizing Saunders, Sastry, Stuhlmüller & Evans (2017). They are also concerned with preventing reinforcement learning agents from running into catastrophes, but have a somewhat different approach. In their approach, a reinforcement learner is allowed to do different kinds of things, which a human overseer then allows or blocks. A separate “blocker” model is trained to predict which actions the human overseer would block. In the future, if the robot would ever take an action which the “blocker” predicts the human overseer would disallow, it will block that action. In effect, the system consists of two separate subagents, one subagent trying to maximize rewards and the other subagent trying to block non-approved actions.
Since our robot has a nice modular architecture into which we can add various subagents which are listening in and taking actions, we decide to take inspiration from this idea. We create a system for spawning dedicated subprograms which try to predict and and block actions which would cause the fear model to be triggered. In theory, this is unnecessary: given enough time, even standard reinforcement learning should learn to avoid the situations which trigger the fear model. But again, trial-and-error can take a very long time to learn exactly which situations trigger fear, so we dedicate a separate subprogram to the task of pre-emptively figuring it out.
Each fear model is paired with a subagent that we’ll call a manager. While the fear model has associated a bunch of cues with the notion of an impending catastrophe, the manager learns to predict which situations would cause the fear model to trigger. Despite sounding similar, these are not the same thing: one indicates when you are already in danger, the other is trying to figure out what you can do to never end up in danger in the first place. A fear model might learn to recognize signs which technological unemployment protesters commonly wear. Whereas a manager might learn the kinds of environments where the fear model has noticed protesters before: for instance, near the protester HQ.
Then, if a manager predicts that a given action (such as going to the protester HQ) would eventually trigger the fear model, it will block that action and promote some other action. We can use the interaction of these subsystems to try to ensure that the robot only feels fear in situations which already resemble the catastrophic situation so much as to actually be dangerous. At the same time, the robot will be unafraid to take safe actions in situations from which it could end up in a danger zone, but are themselves safe to be in.
As an added benefit, we can recycle the manager component to also do the same thing as the blocker component in the Saunders et al. paper originally did. That is, if the robot has a human overseer telling it in strict terms not to do some things, it can create a manager subprogram which models that overseer and likewise blocks the robot from doing things which the model predicts that the overseer would disapprove of.
Putting together a toy model
If the robot does end up in a situation where the fear model is sounding an alarm, then we want to get it out of the situation as quickly as possible. It may be worth spawning a specialized subroutine just for this purpose. Technological unemployment activists could, among other things, use flamethrowers that set the robot on fire. So let’s call these types of subprograms dedicated to escaping from the danger zone, firefighters.
So how does the system as a whole work? First, the different subagents act by sending into the consciousness workspace various mental objects, such as an emotion of fear, or an intent to e.g. make breakfast. If several subagents are submitting identical mental objects, we say that they are voting for the same object. On each time-step, one of the submitted objects is chosen at random to become the contents of the workspace, with each object having a chance to be selected that’s proportional to its number of votes. If a mental object describing a physical action (an “intention”) ends up in the workspace and stays chosen for several time-steps, then that action gets executed by a motor subsystem.
Depending on the situation, some subagents will have more votes than others. E.g. a fear model submitting a fear object gets a number of votes proportional to how strongly it is activated. Besides the specialized subagents we’ve discussed, there’s also a default planning subagent, which is just taking whatever actions (that is, sending to the workspace whatever mental objects) it thinks will produce the greatest reward. This subagent only has a small number of votes.
Finally, there’s a self-narrative agent which is constructing a narrative of the robot’s actions as if it was a unified agent, for social purposes and for doing reasoning afterwards. After the motor system has taken an action, the self-narrative agent records this as something like “I, Robby the Robot, made breakfast by cooking eggs and bacon”, transmitting this statement to the workspace and saving it to an episodic memory store for future reference.
Consequences of the model
Is this design any good? Let’s consider a few of its implications.
First, in order for the robot to take physical actions, the intent to do so has to be in its consciousness for a long enough time for the action to be taken. If there are any subagents that wish to prevent this from happening, they must muster enough votes to bring into consciousness some other mental object replacing that intention before it’s been around for enough time-steps to be executed by the motor system. (This is analogous to the concept of the final veto in humans, where consciousness is the last place to block pre-consciously initiated actions before they are taken.)
Second, the different subagents do not see each other directly: they only see the consequences of each other’s actions, as that’s what’s reflected in the contents of the workspace. In particular, the self-narrative agent has no access to information about which subagents were responsible for generating which physical action. It only sees the intentions which preceded the various actions, and the actions themselves. Thus it might easily end up constructing a narrative which creates the internal appearance of a single agent, even though the system is actually composed of multiple subagents.
Third, even if the subagents can’t directly see each other, they might still end up forming alliances. For example, if the robot is standing near the stove, a curiosity-driven subagent might propose poking at the stove (“I want to see if this causes us to burn ourselves again!”), while the default planning system might propose cooking dinner, since that’s what it predicts will please the human owner. Now, a manager trying to prevent a fear model agent from being activated, will eventually learn that if it votes for the default planning system’s intentions to cook dinner (which it saw earlier), then the curiosity-driven agent is less likely to get its intentions into consciousness. Thus, no poking at the stove, and the manager’s and the default planning system’s goals end up aligned.
Fourth, this design can make it really difficult for the robot to even become aware of the existence of some managers. A manager may learn to support any other mental processes which block the robot from taking specific actions. It does it by voting in favor of mental objects which orient behavior towards anything else. This might manifest as something subtle, such as a mysterious lack of interest towards something that sounds like a good idea in principle, or just repeatedly forgetting to do something, as the robot always seems to get distracted by something else. The self-narrative agent, not having any idea of what’s going on, might just explain this as “Robby the Robot is forgetful sometimes” in its internal narrative.
Fifth, the default planning subagent here is doing something like rational planning, but given its weak voting power, it’s likely to be overruled if other subagents disagree with it (unless some subagents also agree with it). If some actions seem worth doing, but there are managers which are blocking it and the default planning subagent doesn’t have an explicit representation of them, this can manifest as all kinds of procrastinating behaviors and numerous failed attempts for the default planning system to “try to get itself to do something”, using various strategies. But as long as the managers keep blocking those actions, the system is likely to remain stuck.
Sixth, the purpose of both managers and firefighters is to keep the robot out of a situation that has been previously designated as dangerous. Managers do this by trying to pre-emptively block actions that would cause the fear model agent to activate; firefighters do this by trying to take actions which shut down the fear model agent after it has activated. But the fear model agent activating is not actually the same thing as being in a dangerous situation. Thus, both managers and firefighters may fall victim to Goodhart’s law, doing things which block the fear model while being irrelevant for escaping catastrophic situations.
For example, “thinking about the consequences of going to the activist HQ” is something that might activate the fear model agent, so a manager might try to block just thinking about it. This has obvious consequence that the robot can’t think clearly about that issue. Similarly, once the fear model has already activated, a firefighter might Goodhart by supporting any action which helps activate an agent with a lot of voting power that’s going to think about something entirely different. This could result in compulsive behaviors which were effective at pushing the fear aside, but useless for achieving any of the robot’s actual aims.
At worst, this could cause loops of mutually activating subagents pushing in opposite directions. First, a stove-phobic robot runs away from the stove as it was about to make breakfast. Then a firefighter trying to suppress that fear, causes the robot to get stuck looking at pictures of beautiful naked robots, which is engrossing and thus great for removing the fear of the stove. Then another fear model starts to activate, this one afraid of failure and of spending so much time looking at pictures of beautiful naked robots that the robot won’t accomplish its goal of making breakfast. A separate firefighter associated with this second fear model has learned that focusing the robot’s attention on the pictures of beautiful naked robots even more is the most effective action for keeping this new fear temporarily subdued. So the two firefighters are allied and temporarily successful at their goal, but then the first one—seeing that the original stove fear has disappeared—turns off. Without the first firefighter’s votes supporting the second firefighter, the fear manages to overwhelm the second firefighter, causing the robot to rush into making breakfast. This again activates its fear of the stove, but if the fear of failure remains strong enough, it might overpower its fear of the stove so that the robot manages to make breakfast in time...
Hmm. Maybe this design isn’t so great after all. Good thing we noticed these failure modes, so that there aren’t any mind architectures like this going around being vulnerable to them!
The Internal Family Systems model
But enough hypothetical robot design; let’s get to the topic of IFS. The IFS model hypothesizes the existence of three kinds of “extreme parts” in the human mind:
Exiles are said to be parts of the mind which hold the memory of past traumatic events, which the person did not have the resources to handle. They are parts of the psyche which have been split off from the rest and are frozen in time of the traumatic event. When something causes them to surface, they tend to flood the mind with pain. For example, someone may have an exile associated with times when they were romantically rejected in the past.
Managers are parts that have been tasked with keeping the exiles permanently exiled from consciousness. They try to arrange a person’s life and psyche so that exiles never surface. For example, managers might keep someone from reaching out to potential dates due to a fear of rejection.
Firefighters react when exiles have been triggered, and try to either suppress the exile’s pain or distract the mind from it. For example, after someone has been rejected by a date, they might find themselves drinking in an attempt to numb the pain.
Some presentations of the IFS model simplify things by combining Managers and Firefighters into the broader category of Protectors, so only talk about Exiles and Protectors.
Exiles are not limited to being created from the kinds of situations that we would commonly consider seriously traumatic. They can also be created from things like relatively minor childhood upsets, as long as the child didn’t feel like they could handle the situation.
IFS further claims that you can treat these parts as something like independent subpersonalities. You can communicate with them, consider their worries, and gradually persuade managers and firefighters to give you access to the exiles that have been kept away from consciousness. When you do this, you can show them that you are no longer in the situation which was catastrophic before, and now have the resources to handle it if something similar was to happen again. This heals the exile, and also lets the managers and firefighters assume better, healthier roles.
As I mentioned in the beginning, when I first heard about IFS, I was turned off by it for several different reasons. For instance, here were some of my thoughts at the time:
The whole model about some parts of the mind being in pain, and other parts trying to suppress their suffering. The thing about exiles was framed in terms of a part of the mind splitting off in order to protect the rest of the mind against damage. What? That doesn’t make any evolutionary sense! A traumatic situation is just sensory information for the brain, it’s not literal brain damage: it wouldn’t have made any sense for minds to evolve in a way that caused parts of it to split off, forcing other parts of the mind to try to keep them suppressed. Why not just… never be damaged in the first place?
That whole thing about parts being personalized characters that you could talk to. That… doesn’t describe anything in my experience.
Also, how does just talking to yourself fix any trauma or deeply ingrained behaviors?
IFS talks about everyone having a “True Self”. Quote from Wikipedia: IFS also sees people as being whole, underneath this collection of parts. Everyone has a true self or spiritual center, known as the Self to distinguish it from the parts. Even people whose experience is dominated by parts have access to this Self and its healing qualities of curiosity, connectedness, compassion, and calmness. IFS sees the therapist’s job as helping the client to disentangle themselves from their parts and access the Self, which can then connect with each part and heal it, so that the parts can let go of their destructive roles and enter into a harmonious collaboration, led by the Self. That… again did not sound particularly derived from any sensible psychology.
Hopefully, I’ve already answered my past self’s concerns about the first point. The model itself talks in terms of managers protecting the mind from pain, exiles being exiled from consciousness in order for their pain to remain suppressed, etc. Which is a reasonable description of the subjective experience of what happens. But the evolutionary logic—as far as I can guess—is slightly different: to keep us out of dangerous situations.
The story of the robot describes the actual “design rationale”. Exiles are in fact subagents which are “frozen in the time of a traumatic event”, but they didn’t split off to protect the rest of the mind from damage. Rather, they were created as an isolated memory block to ensure that the memory of the event wouldn’t be forgotten. Managers then exist to keep the person away from such catastrophic situations, and firefighters exist to help escape them. Unfortunately, this setup is vulnerable to various failure modes, similar to those that the robot is vulnerable to.
With that said, let’s tackle the remaining problems that I had with IFS.
Personalized characters
IFS suggests that you can experience the exiles, managers and firefighters in your mind as something akin to subpersonalities—entities with their own names, visual appearances, preferences, beliefs, and so on. Furthermore, this isn’t inherently dysfunctional, nor indicative of something like Dissociative Identity Disorder. Rather, even people who are entirely healthy and normal may experience this kind of “multiplicity”.
Now, it’s important to note right off that not everyone has this to a major extent: you don’t need to experience multiplicity in order for the IFS process to work. For instance, my parts feel more like bodily sensations and shards of desire than subpersonalities, but IFS still works super-well for me.
In the book Internal Family Systems Therapy, Richard Schwartz, the developer of IFS, notes that if a person’s subagents play well together, then that person is likely to feel mostly internally unified. On the other hand, if a person has lots of internal conflict, then they are more likely to experience themselves as having multiple parts with conflicting desires.
I think that this makes a lot of sense, assuming the existence of something like a self-narrative subagent. If you remember, this is the part of the mind which looks at the actions that the mind-system has taken, and then constructs an explanation for why those actions were taken. (See e.g. the posts on the limits of introspection and on the Apologist and the Revolutionary for previous evidence for the existence of such a confabulating subagent with limited access to our true motivations.) As long as all the exiles, managers and firefighters are functioning in a unified fashion, the most parsimonious model that the self-narrative subagent might construct is simply that of a unified self. But if the system keeps being driven into strongly conflicting behaviors, then it can’t necessarily make sense of them from a single-agent perspective. Then it might naturally settle on something like a multiagent approach and experience itself as being split into parts.
Kevin Simler, in Neurons Gone Wild, notes how people with strong addictions seem particularly prone to developing multi-agent narratives:
This American Life did a nice segment on addiction a few years back, in which the producers — seemingly on a lark — asked people to personify their addictions. “It was like people had been waiting all their lives for somebody to ask them this question,” said the producers, and they gushed forth with descriptions of the ‘voice’ of their inner addict:
“The voice is irresistible, always. I’m in the thrall of that voice.”
“Totally out of control. It’s got this life of its own, and I can’t tame it anymore.”
“I actually have a name for the voice. I call it Stan. Stan is the guy who tells me to have the extra glass of wine. Stan is the guy who tells me to smoke.”
This doesn’t seem like it explains all of it, though. I’ve frequently been very dysfunctional, and have always found very intuitive the notion of the mind being split into very parts. Yet I mostly still don’t seem to experience my subagents anywhere near as person-like as some others clearly do. I know at least one person who ended up finding IFS because of having all of these talking characters in their head, and who was looking for something that would help them make sense of it. Nothing like that has ever been the case for me: I did experience strongly conflicting desires, but they were just that, strongly conflicting desires.
I can only surmise that it has something to do with the same kinds of differences which cause some people to think mainly verbally, others mainly visually, and others yet in some other hard-to-describe modality. Some fiction writers spontaneously experience their characters as real people who speak to them and will even bother the writer when at the supermarket, and some others don’t.
It’s been noted that the mechanisms which use to model ourselves and other people overlap—not very surprisingly, since both we and other people are (presumably) humans. So it seems reasonable that some of the mechanisms for representing other people, would sometimes also end up spontaneously recruited for representing internal subagents or coalitions of them.
Why should this technique be useful for psychological healing?
Okay, suppose it’s possible to access our subagents somehow. Why would just talking with these entities in your own head, help you fix psychological issues?
Let’s consider that a person having exiles, managers and firefighters is costly in the sense of constraining that person’s options. If you never want to do anything that would cause you to see a stove, that limits quite a bit of what you can do. I strongly suspect that many forms of procrastination and failure to do things we’d like to do are mostly a manifestation of overactive managers. So it’s important not to create those kinds of entities unless the situation really is one which should be designated as categorically unacceptable to end up in.
The theory for IFS mentions that not all painful situations turn into trauma: just ones in which we felt helpless and like we didn’t have the necessary resources for dealing with it. This makes sense, since if we were capable of dealing with it, then the situation can’t have been that catastrophic. The aftermath of the immediate event is important as well: a child who ends up in a painful situation doesn’t necessarily end up traumatized, if they have an adult who can put the event in a reassuring context afterwards.
But situations which used to be catastrophic and impossible for us to handle before, aren’t necessarily that any more. It seems important to have a mechanism for updating that cache of catastrophic events and for disassembling the protections around it, if the protections turn out to be unnecessary.
How does that process usually happen, without IFS or any other specialized form of therapy?
Often, by talking about your experiences with someone you trust. Or writing about them in private or in a blog.
In my post about Consciousness and the Brain, I mentioned that once a mental object becomes conscious, many different brain systems synchronize their processing around it. I suspect that the reason why many people have such a powerful urge to discuss their traumatic experiences with someone else, is that doing so is a way of bringing those memories into consciousness in detail. And once you’ve dug up your traumatic memories from their cache, their content can be re-processed and re-evaluated. If your brain judges that you now do have the resources to handle that event if you ever end up in it again, or if it’s something that simply can’t happen anymore, then the memory can be removed from the cache and you no longer need to avoid it.
I think it’s also significant that, while something like just writing about a traumatic event is sometimes enough to heal, often it’s more effective if you have a sympathetic listener who you trust. Traumas often involve some amount of shame: maybe you were called lazy as a kid and are still afraid of others thinking that you are lazy. Here, having friends who accept you and are willing to nonjudgmentally listen while you talk about your issues, is by itself an indication that the thing that you used to be afraid of isn’t a danger anymore: there exist people who will stay by your side despite knowing your secret.
Now, when you are talking to a friend about your traumatic memory, you will be going through cached memories that have been stored in an exile subagent. A specific memory circuit—one of several circuits specialized for the act of holding painful memories—is active and outputting its contents into the global workspace, from which they are being turned into words.
Meaning that, in a sense, your friend is talking directly to your exile.
Could you hack this process, so that you wouldn’t even need a friend, and could carry this process out entirely internally?
In my earlier post, I remarked that you could view language as a way of joining two people’s brains together. A subagent in your brain outputs something that appears in your consciousness, you communicate it to a friend, it appears in their consciousness, subagents in your friend’s brain manipulate the information somehow, and then they send it back to your consciousness.
If you are telling your friend about your trauma, you are in a sense joining your workspaces together, and letting some subagents in your workspace, communicate with the “sympathetic listener” subagents in your friend’s workspace.
So why not let a “sympathetic listener” subagent in your workspace, hook up directly with the traumatized subagents that are also in your own workspace?
I think that something like this happens when you do IFS. You are using a technique designed to activate the relevant subagents in a very specific way, which allows for this kind of a “hooking up” without needing another person.
For instance, suppose that you are talking to a manager subagent which wants to hide the fact that you’re bad at something, and starts reacting defensively whenever the topic is brought up. Now, one way by which its activation could manifest, is feeding those defensive thoughts and reactions directly into your workspace. In such a case, you would experience them as your own thoughts, and possibly as objectively real. IFS calls this “blending”; I’ve also previously used the term “cognitive fusion” for what’s essentially the same thing.
Instead of remaining blended, you then use various unblending / cognitive defusion techniques that highlight the way by which these thoughts and emotions are coming from a specific part of your mind. You could think of this as wrapping extra content around the thoughts and emotions, and then seeing them through the wrapper (which is obviously not-you), rather than experiencing the thoughts and emotions directly (which you might experience as your own). For example, the IFS book Self-Therapy suggests this unblending technique (among others):
Allow a visual image of the part [subagent] to arise. This will give you the sense of it as a separate entity. This approach is even more effective if the part is clearly a certain distance away from you. The further away it is, the more separation this creates.
Another way to accomplish visual separation is to draw or paint an image of the part. Or you can choose an object from your home that represents the part for you or find an image of it in a magazine or on the Internet. Having a concrete token of the part helps to create separation.
I think of this as something like, you are taking the subagent in question, routing its responses through a visualization subsystem, and then you see a talking fox or whatever. And this is then a representation that your internal subsystems for talking with other people can respond to. You can then have a dialogue with the part (verbally or otherwise) in a way where its responses are clearly labeled as coming from it, rather than being mixed together with all the other thoughts in the workspace. This lets the content coming from the sympathetic-listener subagent and the exile/manager/firefighter subagent be kept clearly apart, allowing you to consider the emotional content as you would as an external listener, preventing you from drowning in it. You’re hacking your brain so as to work as the therapist and client as the same time.
The Self
IFS claims that, below all the various parts and subagents, there exists a “true self” which you can learn to access. When you are in this Self, you exhibit the qualities of “calmness, curiosity, clarity, compassion, confidence, creativity, courage, and connectedness”. Being at least partially in Self is said to be a prerequisite for working with your parts: if you are not, then you are not able to evaluate their models objectively. The parts will sense this, and as a result, they will not share their models properly, preventing the kind of global re-evaluation of their contents that would update them.
This was the part that I was initially the most skeptical of, and which made me most frequently decide that IFS was not worth looking at. I could easily conceptualize the mind as being made up of various subagents. But then it would just be numerous subagents all the way down, without any single one that could be designated the “true” self.
But let’s look at IFS’s description of how exactly to get into Self. You check whether you seem to be blended with any part. If you are, you unblend with it. Then you check whether you might also be blended with some other part. If you are, you unblend from it also. You then keep doing this until you can find no part that you might be blended with. All that’s left are those “eight Cs”, which just seem to be a kind of a global state, with no particular part that they would be coming from.
I now think that “being in Self” represents a state where there no particular subagent is getting a disproportionate share of voting power, and everything is processed by the system as a whole. Remember that in the robot story, catastrophic states were situations in which the organism should never end up. A subagent kicking in to prevent that from happening is a kind of a priority override to normal thinking. It blocks you from being open and calm and curious because some subagent thinks that doing so would be dangerous. If you then turn off or suspend all those priority overrides, then the mind’s default state absent any override seems to be one with the qualities of the Self.
This actually fits at least one model of the function of positive emotions pretty well. Fredrickson (1998) suggests that an important function of positive emotions is to make us engage in activities such as play, exploration, and savoring the company of other people. Doing these things has the effect of building up skills, knowledge, social connections, and other kinds of resources which might be useful for us in the future. If there are no active ongoing threats, then that implies that the situation is pretty safe for the time being, making it reasonable to revert to a positive state of being open to exploration.
The Internal Family Systems Therapy book makes a somewhat big deal out of the fact that everyone, even most traumatized people, ultimately has a Self which they can access. It explains this in terms of the mind being organized to protect against damage, and with parts always splitting off from the Self when it would otherwise be damaged. I think the real explanation is much simpler: the mind is not accumulating damage, it is just accumulating a longer and longer list of situations not considered safe.
As an aside, this model feels like it makes me less confused about confidence. It seems like people are really attracted to confident people, and that to some extent it’s also possible to fake confidence until it becomes genuine. But if confidence is so attractive and we can fake it, why hasn’t evolution just made everyone confident by default?
Turns out that it has. The reason why faked confidence gradually turns into genuine confidence is that by forcing yourself to act in confident ways which felt dangerous before, your mind gets information indicating that this behavior is not as dangerous as you originally thought. That gradually turns off those priority overrides that kept you out of Self originally, until you get there naturally.
The reason why being in Self is a requirement for doing IFS, is the existence of conflicts between parts. For instance, recall the stove-phobic robot having a firefighter subagent that caused it to retreat from the stove into watching pictures of beautiful naked robots. This triggered a subagent which was afraid of the naked-robot-watching preventing the robot from achieving its goals. If the robot now tried to do IFS and talk with the firefighter subagent that caused it to run away from stoves, this might bring to mind content which activated the exile that was afraid of not achieving things. Then that exile would keep flooding the mind with negative memories, trying to achieve its priority override of “we need to get out of this situation”, and preventing the process from proceeding. Thus, all of the subagents that have strong opinions about the situation need to be unblended from, before integration can proceed.
IFS also has a separate concept of “Self-Leadership”. This is a process where various subagents eventually come to trust the Self, so that they allow the person to increasingly remain in Self even in various emergencies. IFS views this as a positive development, not only because it feels nice, but because doing so means that the person will have more cognitive resources available for actually dealing with the emergency in question.
I think that this ties back to the original notion of subagents being generated to invoke priority overrides for situations which the person originally didn’t have the resources to handle. Many of the subagents IFS talks about seem to emerge from childhood experiences. A child has many fewer cognitive, social, and emotional resources for dealing with bad situations, in which case it makes sense to just categorically avoid them, and invoke special overrides to ensure that this happens. A child’s cognitive capacities, models of the world, and abilities to self-regulate are also less developed, so she may have a harder time staying out of dangerous situations without having some priority overrides built in. An adult, however, typically has many more resources than a child does. Even when faced with an emergency situation, it can be much better to be able to remain calm and analyze the situation using all of one’s subagents, rather than having a few of them take over all the decision-making. Thus, it seems to me—both theoretically and practically—that developing Self-Leadership is really valuable.
That said, I do not wish to imply that it would be a good goal to never have negative emotions. Sometimes blending with a subagent, and experiencing resulting negative emotions, is the right thing to do in that situation. Rather than suppressing negative emotions entirely, Self-Leadership aims to get to a state where any emotional reaction tends to be endorsed by the mind-system as a whole. Thus, if feeling angry or sad or bitter or whatever feels appropriate to the situation, you can let yourself feel so, and then give yourself to that emotion without resisting it. As a result, negative emotions become less unpleasant to experience, since there are fewer subagents trying to fight against them. Also, if it turns out that being in a negative emotional state is no longer useful, the system as a whole can just choose to move back into Self.
Final words
I’ve now given a brief summary of the IFS model, and explained why I think it makes sense. This is of course not enough to establish the model as true. But it might help in making the model plausible enough to at least try out.
I think that most people could benefit from learning and doing IFS on themselves, either alone or together with a friend. I’ve been saying that exiles/managers/firefighters tend to be generated from trauma, but it’s important to realize that these events don’t need to be anything immensely traumatic. The kinds of ordinary, normal childhood upsets that everyone has had can generate these kinds of subagents. Remember, just because you think of a childhood event as trivial now, doesn’t mean that it felt trivial to you as a child. Doing IFS work, I’ve found exiles related to memories and events which I thought left no negative traces, but actually did.
Remember also that it can be really hard to notice the presence of some managers: if they are doing their job effectively, then you might never become aware of them directly. “I don’t have any trauma so I wouldn’t benefit from doing IFS” isn’t necessarily correct. Rather, the cues that I use for detecting a need to do internal work are:
Do I have the qualities associated with Self, or is something blocking them?
Do I feel like I’m capable of dealing with this situation rationally, and doing the things which feel like good ideas on an intellectual level?
Do my emotional reactions feel like they are endorsed by my mind-system as a whole, or is there a resistance to them?
If not, there is often some internal conflict which needs to be addressed—and IFS, combined with some other practices such as Focusing and meditation—has been very useful in learning to solve those internal conflicts.
Even if you don’t feel convinced that doing IFS personally would be a good idea, I think adopting its framework of exiles, managers and firefighters is useful for better understanding the behavior of other people. Their dynamics will be easier to recognize in other people if you’ve had some experience recognizing them in yourself, however.
If you want to learn more about IFS, I would recommend starting with Self-Therapy by Jay Earley. In terms of What/How/Why books, my current suggestions would be:
Why: The Power of Focusing, by Ann Weiser Cornell (technically not about IFS, but AWC’s variant of Focusing gets very close to IFS, and is excellent for conveying the right mindset for it)
This post was written as part of research supported by the Foundational Research Institute. Thank you to everyone who provided feedback on earlier drafts of this article: Eli Tyre, Elizabeth Van Nostrand, Jan Kulveit, Juha Törmänen, Lumi Pakkanen, Maija Haavisto, Marcello Herreshoff, Qiaochu Yuan, and Steve Omohundro.
Building up to an Internal Family Systems model
Introduction
Internal Family Systems (IFS) is a psychotherapy school/technique/model which lends itself particularly well for being used alone or with a peer. For years, I had noticed that many of the kinds of people who put in a lot of work into developing their emotional and communication skills, some within the rationalist community and some outside it, kept mentioning IFS.
So I looked at the Wikipedia page about the IFS model, and bounced off, since it sounded like nonsense to me. Then someone brought it up again, and I thought that maybe I should reconsider. So I looked at the WP page again, thought “nah, still nonsense”, and continued to ignore it.
This continued until I participated in CFAR mentorship training last September, and we had a class on CFAR’s Internal Double Crux (IDC) technique. IDC clicked really well for me, so I started using it a lot and also facilitating it to some friends. However, once we started using it on more emotional issues (as opposed to just things with empirical facts pointing in different directions), we started running into some weird things, which it felt like IDC couldn’t quite handle… things which reminded me of how people had been describing IFS. So I finally read up on it, and have been successfully applying it ever since.
In this post, I’ll try to describe and motivate IFS in terms which are less likely to give people in this audience the same kind of a “no, that’s nonsense” reaction as I initially had.
Epistemic status
This post is intended to give an argument for why something like the IFS model could be true and a thing that works. It’s not really an argument that IFS is correct. My reason for thinking in terms of IFS is simply that I was initially super-skeptical of it (more on the reasons of my skepticism later), but then started encountering things which it turned out IFS predicted—and I only found out about IFS predicting those things after I familiarized myself with it.
Additionally, I now feel that IFS gives me significantly more gears for understanding the behavior of both other people and myself, and it has been significantly transformative in addressing my own emotional issues. Several other people who I know report it having been similarly powerful for them. On the other hand, aside for a few isolated papers with titles like “proof-of-concept” or “pilot study”, there seems to be conspicuously little peer-reviewed evidence in favor of IFS, meaning that we should probably exercise some caution.
I think that, even if not completely correct, IFS is currently the best model that I have for explaining the observations that it’s pointing at. I encourage you to read this post in the style of learning soft skills—trying on this perspective, and seeing if there’s anything in the description which feels like it resonates with your experiences.
But before we talk about IFS, let’s first talk about building robots. It turns out that if we put together some existing ideas from machine learning and neuroscience, we can end up with a robot design that pretty closely resembles IFS’s model of the human mind.
What follows is an intentionally simplified story, which is simpler than either the full IFS model or a full account that would incorporate everything that I know about human brains. Its intent is to demonstrate that an agent architecture with IFS-style subagents might easily emerge from basic machine learning principles, without claiming that all the details of that toy model would exactly match human brains. A discussion of what exactly IFS does claim in the context of human brains follows after the robot story.
Wanted: a robot which avoids catastrophes
Suppose that we’re building a robot that we want to be generally intelligent. The hot thing these days seems to be deep reinforcement learning, so we decide to use that. The robot will explore its environment, try out various things, and gradually develop habits and preferences as it accumulates experience. (Just like those human babies.)
Now, there are some problems we need to address. For one, deep reinforcement learning works fine in simulated environments where you’re safe to explore for an indefinite duration. However, it runs into problems if the robot is supposed to learn in a real life environment. Some actions which the robot might take will result in catastrophic consequences, such as it being damaged. If the robot is just doing things at random, it might end up damaging itself. Even worse, if the robot does something which could have been catastrophic but narrowly avoids harm, it might then forget about it and end up doing the same thing again!
How could we deal with this? Well, let’s look at the existing literature. Lipton et al. (2016) proposed what seems like a promising idea for addressing the part about forgetting. Their approach is to explicitly maintain a memory of danger states—situations which are not the catastrophic outcome itself, but from which the learner has previously ended up in a catastrophe. For instance, if “being burned by a hot stove” is a catastrophe, then “being about to poke your finger in the stove” is a danger state. Depending on how cautious we want to be and how many preceding states we want to include in our list of danger states, “going near the stove” and “seeing the stove” can also be danger states, though then we might end up with a seriously stove-phobic robot.
In any case, we maintain a separate storage of danger states, in such a way that the learner never forgets about them. We use this storage of danger states to train a fear model: a model which is trying to predict the probability of ending up in a catastrophe from some given novel situation. For example, maybe our robot poked its robot finger at the stove in our kitchen, but poking its robot finger at stoves in other kitchens might be dangerous too. So we want the fear model to generalize from our stove to other stoves. On the other hand, we don’t want it to be stove-phobic and run away at the mere sight of a stove. The task of our fear model is to predict exactly how likely it is for the robot to end up in a catastrophe, given some situation it is in, and then make it increasingly disinclined to end up in the kinds of situations which might lead to a catastrophe.
This sounds nice in theory. On the other hand, Lipton et al. are still assuming that they can train their learner in a simulated environment, and that they can label catastrophic states ahead of time. We don’t know in advance every possible catastrophe our robot might end up in—it might walk off a cliff, shoot itself in the foot with a laser gun, be beaten up by activists protesting technological unemployment, or any number of other possibilities.
So let’s take inspiration from humans. We can’t know beforehand every bad thing that might happen to our robot, but we can identify some classes of things which are correlated with catastrophe. For instance, being beaten or shooting itself in the foot will cause physical damage, so we can install sensors which indicate when the robot has taken physical damage. If these sensors—let’s call them “pain” sensors—register a high amount of damage, we consider the situation to have been catastrophic. When they do, we save that situation and the situations preceding it to our list of dangerous situations. Assuming that our robot has managed to make it out of that situation intact and can do anything in the first place, we use that list of dangerous situations to train up a fear model.
At this point, we notice that this is starting to remind us about our experience with humans. For example, the infamous Little Albert experiment. A human baby was allowed to play with a laboratory rat, but each time that he saw the rat, a researcher made a loud scary sound behind his back. Soon Albert started getting scared whenever he saw the rat—and then he got scared of furry things in general.
Something like Albert’s behavior could be implemented very simply using something like Hebbian conditioning to get a learning algorithm which picks up on some features of the situation, and then triggers a panic reaction whenever it re-encounters those same features. For instance, it registers that the sight of fur and loud sounds tend to coincide, and then it triggers a fear reaction whenever it sees fur. This would be a basic fear model, and a “danger state” would be “seeing fur”.
Wanting to keep things simple, we decide to use this kind of an approach as the fear model of our robot. Also, having read Consciousness and the Brain, we remember a few basic principles about how those human brains work, which we decide to copy because we’re lazy and don’t want to come up with entirely new principles:
There’s a special network of neurons in the brain, called the global neuronal workspace. The contents of this workspace are roughly the same as the contents of consciousness.
We can thus consider consciousness a workspace which many different brain systems have access to. It can hold a single “chunk” of information at a time.
The brain has multiple different systems doing different things. When a mental object becomes conscious (that is, is projected into the workspace by a subsystem), many systems will synchronize their processing around analyzing and manipulating that mental object.
So here is our design:
The robot has a hardwired system scanning for signs of catastrophe. This system has several subcomponents. One of them scans the “pain” sensors for signs of physical damage. Another system watches the “hunger” sensors for signs of low battery.
Any of these “distress” systems can, alone or in combination, feed a negative reward signal into the global workspace. This tells the rest of the system that this is a bad state, from which the robot should escape.
If a certain threshold level of “distress” is reached, the current situation is designated as catastrophic. All other priorities are suspended and the robot will prioritize getting out of the situation. A memory of the situation and the situations preceding it are saved to a dedicated storage.
After the experience, the memory of the catastrophic situation is replayed in consciousness for analysis. This replay is used to train up a separate fear model which effectively acts as a new “distress” system.
As the robot walks around its environment, sensory information about the surroundings will enter its consciousness workspace. When it plans future actions, simulated sensory information about how those actions would unfold enters the workspace. Whenever the new fear model detects features in either kind of sensory information which it associates with the catastrophic events, it will feed “fear”-type “distress” into the consciousness workspace.
So if the robot sees things which remind it of poking at hot stove, it will be inclined to go somewhere else; if it imagines doing something which would cause it to poke at the hot stove, then it will be inclined to imagine doing something else.
Introducing managers
But is this actually enough? We’ve now basically set up an algorithm which warns the robot when it sees things which have previously preceded a bad outcome. This might be enough for dealing with static tasks, such as not burning yourself at a stove. But it seems insufficient for dealing with things like predators or technological unemployment protesters, who might show up in a wide variety of places and actively try to hunt you down. By the time you see a sign of them, you’re already in danger. It would be better if we could learn to avoid them entirely, so that the fear model would never even be triggered.
As we ponder this dilemma, we surf the web and run across this blog post summarizing Saunders, Sastry, Stuhlmüller & Evans (2017). They are also concerned with preventing reinforcement learning agents from running into catastrophes, but have a somewhat different approach. In their approach, a reinforcement learner is allowed to do different kinds of things, which a human overseer then allows or blocks. A separate “blocker” model is trained to predict which actions the human overseer would block. In the future, if the robot would ever take an action which the “blocker” predicts the human overseer would disallow, it will block that action. In effect, the system consists of two separate subagents, one subagent trying to maximize rewards and the other subagent trying to block non-approved actions.
Since our robot has a nice modular architecture into which we can add various subagents which are listening in and taking actions, we decide to take inspiration from this idea. We create a system for spawning dedicated subprograms which try to predict and and block actions which would cause the fear model to be triggered. In theory, this is unnecessary: given enough time, even standard reinforcement learning should learn to avoid the situations which trigger the fear model. But again, trial-and-error can take a very long time to learn exactly which situations trigger fear, so we dedicate a separate subprogram to the task of pre-emptively figuring it out.
Each fear model is paired with a subagent that we’ll call a manager. While the fear model has associated a bunch of cues with the notion of an impending catastrophe, the manager learns to predict which situations would cause the fear model to trigger. Despite sounding similar, these are not the same thing: one indicates when you are already in danger, the other is trying to figure out what you can do to never end up in danger in the first place. A fear model might learn to recognize signs which technological unemployment protesters commonly wear. Whereas a manager might learn the kinds of environments where the fear model has noticed protesters before: for instance, near the protester HQ.
Then, if a manager predicts that a given action (such as going to the protester HQ) would eventually trigger the fear model, it will block that action and promote some other action. We can use the interaction of these subsystems to try to ensure that the robot only feels fear in situations which already resemble the catastrophic situation so much as to actually be dangerous. At the same time, the robot will be unafraid to take safe actions in situations from which it could end up in a danger zone, but are themselves safe to be in.
As an added benefit, we can recycle the manager component to also do the same thing as the blocker component in the Saunders et al. paper originally did. That is, if the robot has a human overseer telling it in strict terms not to do some things, it can create a manager subprogram which models that overseer and likewise blocks the robot from doing things which the model predicts that the overseer would disapprove of.
Putting together a toy model
If the robot does end up in a situation where the fear model is sounding an alarm, then we want to get it out of the situation as quickly as possible. It may be worth spawning a specialized subroutine just for this purpose. Technological unemployment activists could, among other things, use flamethrowers that set the robot on fire. So let’s call these types of subprograms dedicated to escaping from the danger zone, firefighters.
So how does the system as a whole work? First, the different subagents act by sending into the consciousness workspace various mental objects, such as an emotion of fear, or an intent to e.g. make breakfast. If several subagents are submitting identical mental objects, we say that they are voting for the same object. On each time-step, one of the submitted objects is chosen at random to become the contents of the workspace, with each object having a chance to be selected that’s proportional to its number of votes. If a mental object describing a physical action (an “intention”) ends up in the workspace and stays chosen for several time-steps, then that action gets executed by a motor subsystem.
Depending on the situation, some subagents will have more votes than others. E.g. a fear model submitting a fear object gets a number of votes proportional to how strongly it is activated. Besides the specialized subagents we’ve discussed, there’s also a default planning subagent, which is just taking whatever actions (that is, sending to the workspace whatever mental objects) it thinks will produce the greatest reward. This subagent only has a small number of votes.
Finally, there’s a self-narrative agent which is constructing a narrative of the robot’s actions as if it was a unified agent, for social purposes and for doing reasoning afterwards. After the motor system has taken an action, the self-narrative agent records this as something like “I, Robby the Robot, made breakfast by cooking eggs and bacon”, transmitting this statement to the workspace and saving it to an episodic memory store for future reference.
Consequences of the model
Is this design any good? Let’s consider a few of its implications.
First, in order for the robot to take physical actions, the intent to do so has to be in its consciousness for a long enough time for the action to be taken. If there are any subagents that wish to prevent this from happening, they must muster enough votes to bring into consciousness some other mental object replacing that intention before it’s been around for enough time-steps to be executed by the motor system. (This is analogous to the concept of the final veto in humans, where consciousness is the last place to block pre-consciously initiated actions before they are taken.)
Second, the different subagents do not see each other directly: they only see the consequences of each other’s actions, as that’s what’s reflected in the contents of the workspace. In particular, the self-narrative agent has no access to information about which subagents were responsible for generating which physical action. It only sees the intentions which preceded the various actions, and the actions themselves. Thus it might easily end up constructing a narrative which creates the internal appearance of a single agent, even though the system is actually composed of multiple subagents.
Third, even if the subagents can’t directly see each other, they might still end up forming alliances. For example, if the robot is standing near the stove, a curiosity-driven subagent might propose poking at the stove (“I want to see if this causes us to burn ourselves again!”), while the default planning system might propose cooking dinner, since that’s what it predicts will please the human owner. Now, a manager trying to prevent a fear model agent from being activated, will eventually learn that if it votes for the default planning system’s intentions to cook dinner (which it saw earlier), then the curiosity-driven agent is less likely to get its intentions into consciousness. Thus, no poking at the stove, and the manager’s and the default planning system’s goals end up aligned.
Fourth, this design can make it really difficult for the robot to even become aware of the existence of some managers. A manager may learn to support any other mental processes which block the robot from taking specific actions. It does it by voting in favor of mental objects which orient behavior towards anything else. This might manifest as something subtle, such as a mysterious lack of interest towards something that sounds like a good idea in principle, or just repeatedly forgetting to do something, as the robot always seems to get distracted by something else. The self-narrative agent, not having any idea of what’s going on, might just explain this as “Robby the Robot is forgetful sometimes” in its internal narrative.
Fifth, the default planning subagent here is doing something like rational planning, but given its weak voting power, it’s likely to be overruled if other subagents disagree with it (unless some subagents also agree with it). If some actions seem worth doing, but there are managers which are blocking it and the default planning subagent doesn’t have an explicit representation of them, this can manifest as all kinds of procrastinating behaviors and numerous failed attempts for the default planning system to “try to get itself to do something”, using various strategies. But as long as the managers keep blocking those actions, the system is likely to remain stuck.
Sixth, the purpose of both managers and firefighters is to keep the robot out of a situation that has been previously designated as dangerous. Managers do this by trying to pre-emptively block actions that would cause the fear model agent to activate; firefighters do this by trying to take actions which shut down the fear model agent after it has activated. But the fear model agent activating is not actually the same thing as being in a dangerous situation. Thus, both managers and firefighters may fall victim to Goodhart’s law, doing things which block the fear model while being irrelevant for escaping catastrophic situations.
For example, “thinking about the consequences of going to the activist HQ” is something that might activate the fear model agent, so a manager might try to block just thinking about it. This has obvious consequence that the robot can’t think clearly about that issue. Similarly, once the fear model has already activated, a firefighter might Goodhart by supporting any action which helps activate an agent with a lot of voting power that’s going to think about something entirely different. This could result in compulsive behaviors which were effective at pushing the fear aside, but useless for achieving any of the robot’s actual aims.
At worst, this could cause loops of mutually activating subagents pushing in opposite directions. First, a stove-phobic robot runs away from the stove as it was about to make breakfast. Then a firefighter trying to suppress that fear, causes the robot to get stuck looking at pictures of beautiful naked robots, which is engrossing and thus great for removing the fear of the stove. Then another fear model starts to activate, this one afraid of failure and of spending so much time looking at pictures of beautiful naked robots that the robot won’t accomplish its goal of making breakfast. A separate firefighter associated with this second fear model has learned that focusing the robot’s attention on the pictures of beautiful naked robots even more is the most effective action for keeping this new fear temporarily subdued. So the two firefighters are allied and temporarily successful at their goal, but then the first one—seeing that the original stove fear has disappeared—turns off. Without the first firefighter’s votes supporting the second firefighter, the fear manages to overwhelm the second firefighter, causing the robot to rush into making breakfast. This again activates its fear of the stove, but if the fear of failure remains strong enough, it might overpower its fear of the stove so that the robot manages to make breakfast in time...
Hmm. Maybe this design isn’t so great after all. Good thing we noticed these failure modes, so that there aren’t any mind architectures like this going around being vulnerable to them!
The Internal Family Systems model
But enough hypothetical robot design; let’s get to the topic of IFS. The IFS model hypothesizes the existence of three kinds of “extreme parts” in the human mind:
Exiles are said to be parts of the mind which hold the memory of past traumatic events, which the person did not have the resources to handle. They are parts of the psyche which have been split off from the rest and are frozen in time of the traumatic event. When something causes them to surface, they tend to flood the mind with pain. For example, someone may have an exile associated with times when they were romantically rejected in the past.
Managers are parts that have been tasked with keeping the exiles permanently exiled from consciousness. They try to arrange a person’s life and psyche so that exiles never surface. For example, managers might keep someone from reaching out to potential dates due to a fear of rejection.
Firefighters react when exiles have been triggered, and try to either suppress the exile’s pain or distract the mind from it. For example, after someone has been rejected by a date, they might find themselves drinking in an attempt to numb the pain.
Some presentations of the IFS model simplify things by combining Managers and Firefighters into the broader category of Protectors, so only talk about Exiles and Protectors.
Exiles are not limited to being created from the kinds of situations that we would commonly consider seriously traumatic. They can also be created from things like relatively minor childhood upsets, as long as the child didn’t feel like they could handle the situation.
IFS further claims that you can treat these parts as something like independent subpersonalities. You can communicate with them, consider their worries, and gradually persuade managers and firefighters to give you access to the exiles that have been kept away from consciousness. When you do this, you can show them that you are no longer in the situation which was catastrophic before, and now have the resources to handle it if something similar was to happen again. This heals the exile, and also lets the managers and firefighters assume better, healthier roles.
As I mentioned in the beginning, when I first heard about IFS, I was turned off by it for several different reasons. For instance, here were some of my thoughts at the time:
The whole model about some parts of the mind being in pain, and other parts trying to suppress their suffering. The thing about exiles was framed in terms of a part of the mind splitting off in order to protect the rest of the mind against damage. What? That doesn’t make any evolutionary sense! A traumatic situation is just sensory information for the brain, it’s not literal brain damage: it wouldn’t have made any sense for minds to evolve in a way that caused parts of it to split off, forcing other parts of the mind to try to keep them suppressed. Why not just… never be damaged in the first place?
That whole thing about parts being personalized characters that you could talk to. That… doesn’t describe anything in my experience.
Also, how does just talking to yourself fix any trauma or deeply ingrained behaviors?
IFS talks about everyone having a “True Self”. Quote from Wikipedia: IFS also sees people as being whole, underneath this collection of parts. Everyone has a true self or spiritual center, known as the Self to distinguish it from the parts. Even people whose experience is dominated by parts have access to this Self and its healing qualities of curiosity, connectedness, compassion, and calmness. IFS sees the therapist’s job as helping the client to disentangle themselves from their parts and access the Self, which can then connect with each part and heal it, so that the parts can let go of their destructive roles and enter into a harmonious collaboration, led by the Self. That… again did not sound particularly derived from any sensible psychology.
Hopefully, I’ve already answered my past self’s concerns about the first point. The model itself talks in terms of managers protecting the mind from pain, exiles being exiled from consciousness in order for their pain to remain suppressed, etc. Which is a reasonable description of the subjective experience of what happens. But the evolutionary logic—as far as I can guess—is slightly different: to keep us out of dangerous situations.
The story of the robot describes the actual “design rationale”. Exiles are in fact subagents which are “frozen in the time of a traumatic event”, but they didn’t split off to protect the rest of the mind from damage. Rather, they were created as an isolated memory block to ensure that the memory of the event wouldn’t be forgotten. Managers then exist to keep the person away from such catastrophic situations, and firefighters exist to help escape them. Unfortunately, this setup is vulnerable to various failure modes, similar to those that the robot is vulnerable to.
With that said, let’s tackle the remaining problems that I had with IFS.
Personalized characters
IFS suggests that you can experience the exiles, managers and firefighters in your mind as something akin to subpersonalities—entities with their own names, visual appearances, preferences, beliefs, and so on. Furthermore, this isn’t inherently dysfunctional, nor indicative of something like Dissociative Identity Disorder. Rather, even people who are entirely healthy and normal may experience this kind of “multiplicity”.
Now, it’s important to note right off that not everyone has this to a major extent: you don’t need to experience multiplicity in order for the IFS process to work. For instance, my parts feel more like bodily sensations and shards of desire than subpersonalities, but IFS still works super-well for me.
In the book Internal Family Systems Therapy, Richard Schwartz, the developer of IFS, notes that if a person’s subagents play well together, then that person is likely to feel mostly internally unified. On the other hand, if a person has lots of internal conflict, then they are more likely to experience themselves as having multiple parts with conflicting desires.
I think that this makes a lot of sense, assuming the existence of something like a self-narrative subagent. If you remember, this is the part of the mind which looks at the actions that the mind-system has taken, and then constructs an explanation for why those actions were taken. (See e.g. the posts on the limits of introspection and on the Apologist and the Revolutionary for previous evidence for the existence of such a confabulating subagent with limited access to our true motivations.) As long as all the exiles, managers and firefighters are functioning in a unified fashion, the most parsimonious model that the self-narrative subagent might construct is simply that of a unified self. But if the system keeps being driven into strongly conflicting behaviors, then it can’t necessarily make sense of them from a single-agent perspective. Then it might naturally settle on something like a multiagent approach and experience itself as being split into parts.
Kevin Simler, in Neurons Gone Wild, notes how people with strong addictions seem particularly prone to developing multi-agent narratives:
This doesn’t seem like it explains all of it, though. I’ve frequently been very dysfunctional, and have always found very intuitive the notion of the mind being split into very parts. Yet I mostly still don’t seem to experience my subagents anywhere near as person-like as some others clearly do. I know at least one person who ended up finding IFS because of having all of these talking characters in their head, and who was looking for something that would help them make sense of it. Nothing like that has ever been the case for me: I did experience strongly conflicting desires, but they were just that, strongly conflicting desires.
I can only surmise that it has something to do with the same kinds of differences which cause some people to think mainly verbally, others mainly visually, and others yet in some other hard-to-describe modality. Some fiction writers spontaneously experience their characters as real people who speak to them and will even bother the writer when at the supermarket, and some others don’t.
It’s been noted that the mechanisms which use to model ourselves and other people overlap—not very surprisingly, since both we and other people are (presumably) humans. So it seems reasonable that some of the mechanisms for representing other people, would sometimes also end up spontaneously recruited for representing internal subagents or coalitions of them.
Why should this technique be useful for psychological healing?
Okay, suppose it’s possible to access our subagents somehow. Why would just talking with these entities in your own head, help you fix psychological issues?
Let’s consider that a person having exiles, managers and firefighters is costly in the sense of constraining that person’s options. If you never want to do anything that would cause you to see a stove, that limits quite a bit of what you can do. I strongly suspect that many forms of procrastination and failure to do things we’d like to do are mostly a manifestation of overactive managers. So it’s important not to create those kinds of entities unless the situation really is one which should be designated as categorically unacceptable to end up in.
The theory for IFS mentions that not all painful situations turn into trauma: just ones in which we felt helpless and like we didn’t have the necessary resources for dealing with it. This makes sense, since if we were capable of dealing with it, then the situation can’t have been that catastrophic. The aftermath of the immediate event is important as well: a child who ends up in a painful situation doesn’t necessarily end up traumatized, if they have an adult who can put the event in a reassuring context afterwards.
But situations which used to be catastrophic and impossible for us to handle before, aren’t necessarily that any more. It seems important to have a mechanism for updating that cache of catastrophic events and for disassembling the protections around it, if the protections turn out to be unnecessary.
How does that process usually happen, without IFS or any other specialized form of therapy?
Often, by talking about your experiences with someone you trust. Or writing about them in private or in a blog.
In my post about Consciousness and the Brain, I mentioned that once a mental object becomes conscious, many different brain systems synchronize their processing around it. I suspect that the reason why many people have such a powerful urge to discuss their traumatic experiences with someone else, is that doing so is a way of bringing those memories into consciousness in detail. And once you’ve dug up your traumatic memories from their cache, their content can be re-processed and re-evaluated. If your brain judges that you now do have the resources to handle that event if you ever end up in it again, or if it’s something that simply can’t happen anymore, then the memory can be removed from the cache and you no longer need to avoid it.
I think it’s also significant that, while something like just writing about a traumatic event is sometimes enough to heal, often it’s more effective if you have a sympathetic listener who you trust. Traumas often involve some amount of shame: maybe you were called lazy as a kid and are still afraid of others thinking that you are lazy. Here, having friends who accept you and are willing to nonjudgmentally listen while you talk about your issues, is by itself an indication that the thing that you used to be afraid of isn’t a danger anymore: there exist people who will stay by your side despite knowing your secret.
Now, when you are talking to a friend about your traumatic memory, you will be going through cached memories that have been stored in an exile subagent. A specific memory circuit—one of several circuits specialized for the act of holding painful memories—is active and outputting its contents into the global workspace, from which they are being turned into words.
Meaning that, in a sense, your friend is talking directly to your exile.
Could you hack this process, so that you wouldn’t even need a friend, and could carry this process out entirely internally?
In my earlier post, I remarked that you could view language as a way of joining two people’s brains together. A subagent in your brain outputs something that appears in your consciousness, you communicate it to a friend, it appears in their consciousness, subagents in your friend’s brain manipulate the information somehow, and then they send it back to your consciousness.
If you are telling your friend about your trauma, you are in a sense joining your workspaces together, and letting some subagents in your workspace, communicate with the “sympathetic listener” subagents in your friend’s workspace.
So why not let a “sympathetic listener” subagent in your workspace, hook up directly with the traumatized subagents that are also in your own workspace?
I think that something like this happens when you do IFS. You are using a technique designed to activate the relevant subagents in a very specific way, which allows for this kind of a “hooking up” without needing another person.
For instance, suppose that you are talking to a manager subagent which wants to hide the fact that you’re bad at something, and starts reacting defensively whenever the topic is brought up. Now, one way by which its activation could manifest, is feeding those defensive thoughts and reactions directly into your workspace. In such a case, you would experience them as your own thoughts, and possibly as objectively real. IFS calls this “blending”; I’ve also previously used the term “cognitive fusion” for what’s essentially the same thing.
Instead of remaining blended, you then use various unblending / cognitive defusion techniques that highlight the way by which these thoughts and emotions are coming from a specific part of your mind. You could think of this as wrapping extra content around the thoughts and emotions, and then seeing them through the wrapper (which is obviously not-you), rather than experiencing the thoughts and emotions directly (which you might experience as your own). For example, the IFS book Self-Therapy suggests this unblending technique (among others):
I think of this as something like, you are taking the subagent in question, routing its responses through a visualization subsystem, and then you see a talking fox or whatever. And this is then a representation that your internal subsystems for talking with other people can respond to. You can then have a dialogue with the part (verbally or otherwise) in a way where its responses are clearly labeled as coming from it, rather than being mixed together with all the other thoughts in the workspace. This lets the content coming from the sympathetic-listener subagent and the exile/manager/firefighter subagent be kept clearly apart, allowing you to consider the emotional content as you would as an external listener, preventing you from drowning in it. You’re hacking your brain so as to work as the therapist and client as the same time.
The Self
IFS claims that, below all the various parts and subagents, there exists a “true self” which you can learn to access. When you are in this Self, you exhibit the qualities of “calmness, curiosity, clarity, compassion, confidence, creativity, courage, and connectedness”. Being at least partially in Self is said to be a prerequisite for working with your parts: if you are not, then you are not able to evaluate their models objectively. The parts will sense this, and as a result, they will not share their models properly, preventing the kind of global re-evaluation of their contents that would update them.
This was the part that I was initially the most skeptical of, and which made me most frequently decide that IFS was not worth looking at. I could easily conceptualize the mind as being made up of various subagents. But then it would just be numerous subagents all the way down, without any single one that could be designated the “true” self.
But let’s look at IFS’s description of how exactly to get into Self. You check whether you seem to be blended with any part. If you are, you unblend with it. Then you check whether you might also be blended with some other part. If you are, you unblend from it also. You then keep doing this until you can find no part that you might be blended with. All that’s left are those “eight Cs”, which just seem to be a kind of a global state, with no particular part that they would be coming from.
I now think that “being in Self” represents a state where there no particular subagent is getting a disproportionate share of voting power, and everything is processed by the system as a whole. Remember that in the robot story, catastrophic states were situations in which the organism should never end up. A subagent kicking in to prevent that from happening is a kind of a priority override to normal thinking. It blocks you from being open and calm and curious because some subagent thinks that doing so would be dangerous. If you then turn off or suspend all those priority overrides, then the mind’s default state absent any override seems to be one with the qualities of the Self.
This actually fits at least one model of the function of positive emotions pretty well. Fredrickson (1998) suggests that an important function of positive emotions is to make us engage in activities such as play, exploration, and savoring the company of other people. Doing these things has the effect of building up skills, knowledge, social connections, and other kinds of resources which might be useful for us in the future. If there are no active ongoing threats, then that implies that the situation is pretty safe for the time being, making it reasonable to revert to a positive state of being open to exploration.
The Internal Family Systems Therapy book makes a somewhat big deal out of the fact that everyone, even most traumatized people, ultimately has a Self which they can access. It explains this in terms of the mind being organized to protect against damage, and with parts always splitting off from the Self when it would otherwise be damaged. I think the real explanation is much simpler: the mind is not accumulating damage, it is just accumulating a longer and longer list of situations not considered safe.
As an aside, this model feels like it makes me less confused about confidence. It seems like people are really attracted to confident people, and that to some extent it’s also possible to fake confidence until it becomes genuine. But if confidence is so attractive and we can fake it, why hasn’t evolution just made everyone confident by default?
Turns out that it has. The reason why faked confidence gradually turns into genuine confidence is that by forcing yourself to act in confident ways which felt dangerous before, your mind gets information indicating that this behavior is not as dangerous as you originally thought. That gradually turns off those priority overrides that kept you out of Self originally, until you get there naturally.
The reason why being in Self is a requirement for doing IFS, is the existence of conflicts between parts. For instance, recall the stove-phobic robot having a firefighter subagent that caused it to retreat from the stove into watching pictures of beautiful naked robots. This triggered a subagent which was afraid of the naked-robot-watching preventing the robot from achieving its goals. If the robot now tried to do IFS and talk with the firefighter subagent that caused it to run away from stoves, this might bring to mind content which activated the exile that was afraid of not achieving things. Then that exile would keep flooding the mind with negative memories, trying to achieve its priority override of “we need to get out of this situation”, and preventing the process from proceeding. Thus, all of the subagents that have strong opinions about the situation need to be unblended from, before integration can proceed.
IFS also has a separate concept of “Self-Leadership”. This is a process where various subagents eventually come to trust the Self, so that they allow the person to increasingly remain in Self even in various emergencies. IFS views this as a positive development, not only because it feels nice, but because doing so means that the person will have more cognitive resources available for actually dealing with the emergency in question.
I think that this ties back to the original notion of subagents being generated to invoke priority overrides for situations which the person originally didn’t have the resources to handle. Many of the subagents IFS talks about seem to emerge from childhood experiences. A child has many fewer cognitive, social, and emotional resources for dealing with bad situations, in which case it makes sense to just categorically avoid them, and invoke special overrides to ensure that this happens. A child’s cognitive capacities, models of the world, and abilities to self-regulate are also less developed, so she may have a harder time staying out of dangerous situations without having some priority overrides built in. An adult, however, typically has many more resources than a child does. Even when faced with an emergency situation, it can be much better to be able to remain calm and analyze the situation using all of one’s subagents, rather than having a few of them take over all the decision-making. Thus, it seems to me—both theoretically and practically—that developing Self-Leadership is really valuable.
That said, I do not wish to imply that it would be a good goal to never have negative emotions. Sometimes blending with a subagent, and experiencing resulting negative emotions, is the right thing to do in that situation. Rather than suppressing negative emotions entirely, Self-Leadership aims to get to a state where any emotional reaction tends to be endorsed by the mind-system as a whole. Thus, if feeling angry or sad or bitter or whatever feels appropriate to the situation, you can let yourself feel so, and then give yourself to that emotion without resisting it. As a result, negative emotions become less unpleasant to experience, since there are fewer subagents trying to fight against them. Also, if it turns out that being in a negative emotional state is no longer useful, the system as a whole can just choose to move back into Self.
Final words
I’ve now given a brief summary of the IFS model, and explained why I think it makes sense. This is of course not enough to establish the model as true. But it might help in making the model plausible enough to at least try out.
I think that most people could benefit from learning and doing IFS on themselves, either alone or together with a friend. I’ve been saying that exiles/managers/firefighters tend to be generated from trauma, but it’s important to realize that these events don’t need to be anything immensely traumatic. The kinds of ordinary, normal childhood upsets that everyone has had can generate these kinds of subagents. Remember, just because you think of a childhood event as trivial now, doesn’t mean that it felt trivial to you as a child. Doing IFS work, I’ve found exiles related to memories and events which I thought left no negative traces, but actually did.
Remember also that it can be really hard to notice the presence of some managers: if they are doing their job effectively, then you might never become aware of them directly. “I don’t have any trauma so I wouldn’t benefit from doing IFS” isn’t necessarily correct. Rather, the cues that I use for detecting a need to do internal work are:
Do I have the qualities associated with Self, or is something blocking them?
Do I feel like I’m capable of dealing with this situation rationally, and doing the things which feel like good ideas on an intellectual level?
Do my emotional reactions feel like they are endorsed by my mind-system as a whole, or is there a resistance to them?
If not, there is often some internal conflict which needs to be addressed—and IFS, combined with some other practices such as Focusing and meditation—has been very useful in learning to solve those internal conflicts.
Even if you don’t feel convinced that doing IFS personally would be a good idea, I think adopting its framework of exiles, managers and firefighters is useful for better understanding the behavior of other people. Their dynamics will be easier to recognize in other people if you’ve had some experience recognizing them in yourself, however.
If you want to learn more about IFS, I would recommend starting with Self-Therapy by Jay Earley. In terms of What/How/Why books, my current suggestions would be:
How: Self-Therapy by Jay Earley.
What: Internal Family Systems Therapy, by Richard Schwartz
Why: The Power of Focusing, by Ann Weiser Cornell (technically not about IFS, but AWC’s variant of Focusing gets very close to IFS, and is excellent for conveying the right mindset for it)
This post was written as part of research supported by the Foundational Research Institute. Thank you to everyone who provided feedback on earlier drafts of this article: Eli Tyre, Elizabeth Van Nostrand, Jan Kulveit, Juha Törmänen, Lumi Pakkanen, Maija Haavisto, Marcello Herreshoff, Qiaochu Yuan, and Steve Omohundro.