An addendum to Prediction, Control, and the Regulating Mind
Read the cautions first, and please take them seriously. Anorexia has the highest mortality of any psychiatric diagnosis. Nothing here is an assessment and nothing here is a treatment. If this is you, or your child, or your client, the next step is a real assessment with somebody who treats eating disorders, not a mechanistic story about loops. If you are reading this about your own situation rather than a caseload, please talk to somebody you trust or a licensed professional – I would rather you did that than finish the article.
The main document applies the model to stimming, and an addendum applies it to self-injury. This one takes it into eating disorders, where the fit is closer than anywhere else in the set and needs more care for exactly that reason.
Everywhere else in this document, “set point” is an analogy. Here it’s closer to literal. Something in the system is holding a body state against disturbance and correcting hard when it moves… which is what a control loop does.
When a framework fits a case that well, check whether it’s explaining anything or just relabeling it in nicer vocabulary.
These are some caveats. I am not an expert in this area, we’re just exploring model fit.
Nothing here is an assessment and nothing here is a treatment. Arcelus and colleagues (2011) pooled 36 studies and put the mortality of anorexia above that of any other psychiatric diagnosis (Arcelus, Mitchell, Wales & Nielsen, 2011). If this is you, or your kid, or your client, the next step is a real assessment with somebody who treats eating disorders, not a mechanistic story about loops.
Weight restoration is the treatment for anorexia. An account of habit is not a reason to delay refeeding, soften a target weight, or treat the behavior as something to understand before it’s interrupted. NONE of it. A person who’s starving is not in a state where any of this applies to them!
And refeeding is the exact thing they’re most afraid of. Not eating can feel like control, so taking it away is traumatic, in the full sense of that word rather than the loose one. That doesn’t change the rule above by one inch. It changes what you know you’re asking of them. How would you rather be handled, if it were you on that side of it? “Just feed them” is not the easy part of this, and it is not a simple thing to be on the receiving end of.
It is an account of persistence, not of cause. The habit work below is about what keeps restriction going rather than what starts it, and Walsh (2013) says so himself. Why it begins is a different question involving genetics, development, culture and events, and none of that is in here.
Calling a behavior a shortcut is not calling it a choice. If anything it is the reverse, and that is the point of the word. But it can be heard the other way in a family meeting, so don’t use it there without saying which one you mean.
And do not tell somebody their interoception is broken. Interoception is the sense of your own internal state, the reading you get of your heartbeat, your breathing, your hunger and fullness. Not because of any study, but because it is the kind of sentence somebody carries for years.
My position: I don’t treat this population. What follows is an examination of other people’s work to see if the model fits or could be helpful, and it should be weighed that way.
You don’t need the book to read this, and here is the whole apparatus in four terms.
The model treats a person as a control loop. Two halves.
The estimator is the half that reads your state. What’s going on out there, what’s going on in here, how bad is it. It doesn’t measure the world directly, it produces a best guess from the evidence, and its guess is what you actually experience.
The controller is the half that acts. Given the estimate, it selects a response and runs it.
A schema is the stored part. It’s the embedded cognitive assessment the loop leans on, built out of everything that ever worked before, and the estimator’s job is to keep adjusting it as evidence comes in.
A shortcut, or heuristic, is what’s happening whenever either half bypasses differentiated conscious thought and simply runs. Driving a familiar route. Which hand you eat with. My son plays games on a 14-button PlayStation controller and remembers what every one of them does. Most of what you do all day runs this way, and that’s a feature, not a defect.
Shortcuts fail in two ways. The situation is new, so the shortcut fires because it doesn’t know any better. Or the shortcut is maladaptive: it was learned somewhere it genuinely worked, often between five and ten years old, and it’s now embedded and hard to shift. That second one is what CBT and EMDR are aimed at.
Two more words this addendum uses. Gain is how hard the loop corrects for a given error, which is why two people facing the same problem can respond at wildly different volumes. And observer gain is how much weight the person puts on their own reading, which turns out to matter here more than the reading itself.
A shortcut, in the sense 4.4a gives the concept, is what’s happening whenever either half of the loop bypasses differentiated conscious thought and just runs.
Walsh (2013) proposed that what keeps anorexia going isn’t what starts it, and that persistence is a habit problem (Walsh, 2013). Whatever sets restriction going, it stops requiring deliberation. His phrase for what it becomes is that it needs minimal, if any, conscious effort.
Walsh got there working on eating disorders, with no knowledge of this framework at all! Two lines of work landing on the same wording is what makes it worth something.
Then somebody tested it! Foerde, Steinglass, Shohamy and Walsh (2015) put people in a scanner making actual food choices, and dorsal striatal activity tracked the choices of patients with anorexia while it didn’t track the choices of controls, which they read as habitual rather than goal-directed action (Foerde et al., 2015). Steinglass and Walsh (2016) built that into a full model of persistence. Conceição and colleagues (2023) reviewed the accumulated habit work in a scoping review.
Status: the best-evidenced application in this whole document. The autism material rests on four findings collected to answer other questions. This one has a proposed mechanism, a task, a scanner result where the groups came apart, and a decade of follow-up work.
This one sits on the controller. Restriction is a stored RESPONSE, which is why “they know it’s dangerous” and “they keep doing it” are both true at once and not a contradiction. That pair gets read as denial. It isn’t only that, and reading it that way costs you the actual problem.
A woman who was hospitalized for anorexia and recovered read this section and wrote back:
“I knew what I was doing was very harmful, but it didn’t matter to me. Anorexia is like a slow suicide, but I kept doing it.”
Read that twice, because everything the model claims about this is in those two sentences. The knowledge was there. It was accurate. It named the outcome correctly, and she went and did it anyway. The problem was never that she didn’t know. The knowledge was fully present and carried no weight at all, and that’s a worse problem than ignorance, because ignorance has an obvious remedy and this doesn’t.
Which is exactly what a stored response looks like from the inside. The controller runs. It doesn’t consult the part of you that has the facts. Then everybody in the room concludes she must not really understand, and explains it to her again.
The received view is that interoception is straightforwardly impaired in anorexia. That’s my impression of the field rather than a survey of it, and it isn’t what the meta-analysis found.
Gend, Remy and Lutz (2026) pooled 16 studies of cardiac interoceptive accuracy in anorexia and found no detectable difference from controls on that task (Gend et al., 2026). Illness duration moderated the effect, with the lower scores concentrated in shorter-duration samples. That’s consistent with something shifting over the course of illness, and it’s equally consistent with short and long duration samples differing in other ways, which a between-study moderator can’t separate.
What differs is confidence rather than performance. Patients rate their own interoceptive ability low while performing about as well. That’s my read across the literature, not a single reported result.
That’s an observer gain problem. The reading holds up on this task, and what doesn’t is how far the person trusts it.
The literature leans almost entirely on the heartbeat counting task, and that task’s validity is genuinely disputed, so “no detectable difference” might partly mean “no difference on a bad instrument.”
Ask a patient with anorexia on the day she’s admitted whether she’s hungry and she’ll tell you no. Not “yes but I won’t eat.” No. She isn’t lying and she isn’t being difficult. She has been starving for months and she cannot feel it.
What she can feel is full. Two bites in, she’s full the way you’re full at the end of Thanksgiving dinner, and that feeling is real. She’s also frightened, because the meal in front of her is the one thing she has spent years learning to avoid and somebody has just taken away her only means of managing it. Then a nurse hands her a tablet of olanzapine to switch the appetite back on and push the food through a stomach that has forgotten what to do with it, and that’s its own kind of awful.
This is the part people who have never watched it get wrong. They picture somebody desperately hungry, white-knuckling it past a plate of food. The real thing is worse and stranger. The hunger has stopped reporting.
Now put instruments on it. Heruc et al. (2019) measured 22 adolescent inpatients against 17 healthy controls (Heruc et al., 2019). Acyl ghrelin, the hormone whose entire job is to signal hunger, was running at more than double control levels. Reported hunger sat at about a quarter of controls. Reported fullness ran more than triple.
The hunger hormone is screaming. Nobody is picking up.
Robinson (1989) measured the connection itself. In healthy people, what’s physically sitting in your stomach tracks what you report feeling at a correlation around 0.6. In anorexia the hunger correlation falls significantly while the fullness correlation stays normal (Robinson, 1989). One half of the appetite signal has come unhooked from the body it reports on. The other half hasn’t.
That’s an estimator failure you can watch from both ends at once, which is more than the heartbeat literature above manages, and it’s why “listen to your body” is the wrong sentence to say in that room. It isn’t that nothing is on the line. It’s that the reading is confident, detailed, and wrong in one direction only. Mechanical eating exists for exactly that reason. You eat by the clock and the plan because the instrument that was supposed to tell you when has been reading zero for a year.
The fullness isn’t imaginary either, which matters if you’re ever tempted to treat it as resistance. Benini et al. (2004) timed a meal at roughly six hours to clear the stomach in anorexia against three and a half in controls, and after 22 weeks of rehabilitation the restricting patients were still near five hours while the purging patients hadn’t improved at all (Benini et al., 2004). Food really is sitting there. She really is full. Both are true and neither one means she should stop eating.
A ceiling on all of it. Norris et al. (2016) reviewed 123 papers on gastrointestinal complications in anorexia and found exactly ONE randomized trial among them, and found that where medical testing was actually done, self-reported symptoms didn’t correlate with measurable pathology (Norris et al., 2016). Which is the lesson the interoception section taught, arriving from the gut instead of the heart. Heruc’s gastric-emptying data also has emptying back to normal after two weeks where Benini has it abnormal at 22. Different instruments, different patients, unresolved.
And a fit this clean earns more caution rather than less. Neatness isn’t evidence.
What this still owes: calling that dissociation “observer gain” renames it, and renaming isn’t explaining. The habit account earns its keep because Walsh (2013) proposed a mechanism and Foerde et al. (2015) tested it. The heartbeat material doesn’t, yet. The hunger material comes closer, because there you have the signal, the report, and the broken link between them all measured in the same people.
Most illnesses produce a symptom. Depression produces a symptom and its opposite, and the explanation there is two opposing loops rather than one.
Restriction and bingeing may work the same way, as two directions of correction on ONE regulated variable rather than two disorders that happen to share a chapter. That would also be a reason people move between the diagnoses over time. The crossover is documented. This explanation of it is not.
Chester, Viranda, Kaye and Berner (2024) ran an initial test of an interoceptive inference account in bulimia (Chester et al., 2024). Thirty women with bulimia against thirty-one matched controls. The bulimia group reported lower trust in sensory information from the body, and stronger beliefs that once upset there’s little to be done apart from eating. Those two travelled together, and the belief about uncontrollable emotion tracked how often subjective binge episodes happened.
Status: initial support, self-report only, n = 30. The authors call it INITIAL themselves and they’re right. It’s a survey study rather than a mechanism, and it can’t tell you which way the arrow runs.
What it describes is a shortcut with a belief attached: low trust in the read, and a stored response that’s been credited with fixing the feeling every previous time.
Barca and Pezzulo (2020) set out the active-inference version for anorexia (Barca & Pezzulo, 2020), and it’s the same Barca whose 2023 paper the self-injury addendum leans on. A single research group is carrying a lot of the theoretical weight across two of these addenda. That’s a thin base and I’d rather say so than let it pass.
If a model says what’s keeping the behavior going, it ought to be able to say why the treatments that work, work.
For medically stable adolescents in outpatient care, family-based treatment has the most evidence behind it. Lock et al. (2010) randomized 121 adolescents to 12 months of FBT or adolescent-focused individual therapy (Lock et al., 2010). At the end of treatment the two were even on full remission. At both 6 and 12 months afterward, FBT was significantly ahead. Austin et al. (2025) pooled 18 randomized trials and found more weight gain from family therapy than from individual psychotherapy at end of treatment, with the parent-focused and separated formats beating the conjoint one (Austin et al., 2025). Their own wording is that it’s the best outpatient option when getting weight on quickly is what matters most. Everything else they tested came back null.
For adults nothing has separated on weight. The ANTOP trial (Zipfel et al., 2014) randomized 242 adults across ten German hospitals to focal psychodynamic therapy, enhanced CBT, or optimized treatment as usual, meaning outpatient psychotherapy plus structured care from a family doctor.
| Arm | BMI gain at end of treatment | BMI gain at 12 months |
|---|---|---|
| Focal psychodynamic therapy | 0.73 kg/m² | 1.64 kg/m² |
| Enhanced CBT | 0.93 kg/m² | 1.30 kg/m² |
| Optimized treatment as usual | 0.69 kg/m² | 1.22 kg/m² |
Weight rose in every arm and kept rising, and no arm beat another on BMI at either point. On the secondaries they did come apart. Focal psychodynamic therapy was ahead on recovery at 12 months, and enhanced CBT was faster on weight gain and better on eating disorder psychopathology. The authors’ conclusion has both halves in it: optimized treatment as usual should be treated as the solid baseline, AND those two advantages are real (Zipfel et al., 2014). Ten hospitals, 242 patients, three serious treatments, and nothing separated on the primary outcome! And it isn’t only ANTOP. Solmi and colleagues (2021) network-meta-analyzed 16 randomized trials covering 1,047 adult outpatients with anorexia and found that none of the interventions outperformed treatment as usual on the primary outcomes, with confidence in the evidence rated low to very low (Solmi et al., 2021).
No drug does the main work in anorexia. Attia et al. (2019) ran olanzapine against placebo in 152 adult outpatients for 16 weeks (Attia et al., 2019). BMI climbed faster on the drug, 0.259 against 0.095 per month, which the authors themselves call a modest effect. On obsessionality the two groups didn’t separate.
That trial measured weight and obsessionality, which is not the whole of what the drug is doing on an inpatient unit, where olanzapine is also given to bring appetite back and to move food through a gut that has adapted to almost nothing. Being on the receiving end of that is traumatizing in its own right rather than a side benefit. None of which is a reason to refuse it. That call belongs to the team treating the person, who can see things this paragraph can’t. Whichever job the drug is doing, the target is the body and the schema is somewhere else, which is the same answer the obsessionality scores gave, arriving by a different road.
CBT is first line for bulimia and the abstinence numbers are low across the whole field. Linardon and Wade (2018) pooled 45 randomized trials covering 78 psychotherapy conditions of every kind, not CBT alone, and found that 35.4% of treatment completers abstained from bingeing and purging afterward, or 29.9% counting everybody randomized (Linardon & Wade, 2018). At follow-up those figures were 34.6% and 28.6%, effectively unchanged. Behavioral treatments, CBT among them, sat at the high end of that spread.
Over 60% don’t reach full abstinence at the end of treatment, and that’s using the treatments the field is most confident in. Abstinence there is a hard line, 28 days clean in most of these trials. A lot of the people under that line got a great deal better and the meta-analysis wasn’t scoring that. None of this is a reason to skip CBT. About a third stop entirely, most of the rest improve, and it’s the best-tested thing there is.
Fluoxetine at 60 mg a day beat placebo on weekly binge and vomit frequency in 387 women over 8 weeks, with 20 mg landing between the two (Fluoxetine Bulimia Nervosa Collaborative Study Group, 1992). That was an acute trial, 8 weeks. Insomnia, nausea, asthenia and tremor all came up more often on the drug than on placebo, though not enough more that people quit over it.
The combination is where the numbers get interesting. Bacaltchuk, Hay and Trefiglio (2001) reviewed the randomized trials that pitted antidepressants against psychological treatment, and either against the two together, scoring full remission (Bacaltchuk et al., 2001).
| Comparison | Remission | Result |
|---|---|---|
| Antidepressant alone vs psychotherapy alone | 20% vs 39% | RR 1.28, CI 0.98–1.67, not significant |
| Combination vs antidepressant alone | 42% vs 23% | RR 1.38, CI 0.98–1.93, not significant |
| Combination vs psychotherapy alone | 49% vs 36% | RR 1.21, CI 1.02–1.45, significant, NNT 8 |
The relative risks there are computed on failure to remit, which is how the review reported them, so they look smaller than a ratio of the two percentages would. Which half is the medication moving?
Two of the three comparisons came back null. Dropout ran higher on the combination, one extra dropout for every 7 patients treated, so whatever those extra 13 points are buying, some patients are paying for it by leaving. Dropout was also higher on antidepressants alone than on psychotherapy alone, one extra for every 4, and the reviewers concluded psychotherapy was simply more acceptable to patients.
Four readings.
Start with the question the anorexia numbers put on the table. If the meaning of the restriction is what’s holding it in place, why does the treatment that never discusses the meaning win on weight?
The treatment that works best for anorexia ignores the meaning entirely. FBT hands control of eating to the parents and holds it there for months, so the stored response gets no opening, over and over. In the terms of 4.4a that’s the disconfirming-outcome route with the volume turned all the way up. The comparison arm, adolescent-focused individual therapy, works on understanding. Both treatments improved these kids considerably and the two were even on full remission at the end of treatment, so this is a difference in speed and in durability rather than one treatment working and one not. Families who can’t run FBT still have somewhere to go.
Read that as sequencing and not as a verdict. When somebody is close to organ failure, refeeding is the emergency and the whys and wherefores take a back seat for a while. Taking a back seat is not the same as being unnecessary. Nothing here tested whether the understanding still has to happen afterward, and I’d expect that it does, once there’s a body well enough to do it in.
No adult treatment separated on weight, and the frame is comfortable with that. If the lever is repetition of a different response, then which brand of conversation surrounds the repetition should matter less than whether the eating actually changed. That’s weak support and I’d rather call it weak than dress it up. A null on a primary outcome is also what an underpowered trial looks like, and the secondaries in ANTOP did separate, which the frame doesn’t predict in either direction.
Olanzapine may pull the two halves apart inside one trial. Weight moved and the obsessionality scores didn’t separate. A drug that shifts the controlled variable while leaving the schema where it was is the profile I’d have guessed, and I’m saying that after reading the result, which isn’t the same as predicting it. It’s a weak test for another reason too: no account of anorexia predicts a weight-gain drug fixing obsessionality, so the model gets no credit for agreeing with all of them.
The combination result is the best evidence for the two-halves reading and the weakest number on this page. CBT works the controller, medication works the affective side, the two together beat CBT alone by 13 points. That’s a satisfying story. It’s also the single significant contrast in a review where the other two were null, its confidence interval on number-needed-to-treat runs from 4 to 320, and more people quit the combination arm than either single treatment. Citing the one that fit and passing over the two that didn’t would be me arguing a position instead of reporting one.
What the treatments cost the framework, taken together. The treatment built on changing the eating beats the one built on understanding it, on speed and on durability. Nobody here tested ADDING insight to behavior change, so I can’t say insight adds nothing, and I won’t. What the treatment evidence supports is the habit reading of restriction, and it gives the two-halves reading of interoception exactly nothing. Anorexia has an effective treatment already and it doesn’t need this model to work. If the frame earns anything clinically, it will be in bulimia, where the ceiling is 35% and nobody is satisfied.
Everything above treats the estimator question as a question about signals from inside the body. That leaves out the estimate people actually act on, which is the one they get by looking.
Here’s the thing nobody outside this tells you. The distortion usually arrives before the behavior does, and it’s still there years after the weight is back. It isn’t a symptom of being underweight. It doesn’t resolve when the scale does.
Cornelissen and Tovée (2021) call body image disturbance a core component of eating disorders and a predictor of three separate things: onset, treatment outcome, and future relapse (Cornelissen & Tovée, 2021). One variable, three jobs. Getting the weight back doesn’t retire it. It stays the monster to be reckoned with, and it’s probably why so many people who look recovered aren’t.
A woman who has restored her weight, who any doctor would call recovered, looks in a mirror and sees a fat person looking back. Not a woman worried about being fat. A fat person, right there, plainly visible. And underneath the fatness, the reading that does the real damage: unlovable.
That chain is the whole problem, and notice what each link is made of. Fat is a category. Ugly is a category. Unlovable is a verdict about whether you get to be wanted by anybody, ever, and it’s been reached in about a second and a half from a reflection. Nobody measured anything to get there. No inches were estimated. A body was sorted into a bin, and the bin came with a sentence attached.
The obvious intervention is to correct the estimate. Show her the tape measure. Show her the photographs, the sizes, the numbers. Prove she’s wrong about the inches. Everybody who has loved somebody with this has tried it, and it does approximately nothing. Why would it?
The training that helps does NOT work on the accuracy of body-size judgements at all. It works on how a body of a given size gets categorized, moving the boundary at which a body starts being judged overweight, and moving that boundary appears to produce a broad reassessment of eating-disordered attitudes rather than a narrow improvement at estimating inches (Cornelissen & Tovée, 2021).
In the terms of 4.4a the stored thing is a category boundary rather than a measurement, which is a schema in the plainest sense this book has. Move the boundary and the attitudes move with it. Correct the measurement, the intuitive thing, and nothing happens. If you’ve ever wondered why the reassurance doesn’t land, that’s why. You’re arguing with a ruler. The problem is the bin.
A boundary like that isn’t born in somebody’s head. It gets installed, and everybody helps.
Becker and colleagues (2002) ran the closest thing to a controlled experiment anyone will ever get. Television arrived in a rural region of Fiji in 1995. They measured disordered eating attitudes in adolescent Fijian schoolgirls before and after prolonged exposure, and key indicators of disordered eating were significantly more prevalent afterward (Becker et al., 2002). The follow-up interviews say why: the girls described wanting to lose weight in order to model themselves on the characters they were watching (Becker, 2004). A community with no history of this developed it in about three years, on delivery!
Some countries are worse than the United States, and they’re usually the ones most convinced they’re the victim of it.
I was in Brazil, riding with two brothers who drove me around for the better part of a day. They brought up anorexia and bulimia themselves and they wouldn’t let it go. It was terrible. Women were getting sick. Something ought to be done. They meant every word of it.
I thought about it for a few minutes and then asked them: would you date a woman who wasn’t skinny?
Both of them, instantly, without a half second of thought: “No way!”
They had no idea they’d just answered their own question… and I didn’t push it, because what would I have said? To them those were two unrelated topics that happened to come up in the same car ride. One was a public health tragedy visited on their country from somewhere outside it. The other was just, you know, what you find attractive.
That’s the boundary being installed, live, by two men who would sincerely tell you they’re on the right side of this. Not media. Not magazines. The people you might actually date, telling you where the line is, and then grieving about the illness that line produces.
The interoception section owed a prediction and still does. This one names how the thing works, gives you an intervention that follows from that, and then reports a result landing where it should have landed. If the two-halves reading earns anything, it earns it here.
It also explains something the habit account can’t. Restriction as a stored response tells you why the behavior persists. It doesn’t tell you why the person still can’t stand what she sees after a year at a normal weight. The mirror is the estimator, running a schema nobody has retrained, in a culture that reinstalls it on contact.
Status: promising, and the trials are small and few. Cornelissen and Tovée say themselves this needs further trials for long-term effectiveness. A category-shift result is not yet a treatment. And the Fiji work is a natural experiment rather than a randomized one, so read it as strong evidence about that population rather than a law.
A shortcut gets built when the outcome keeps matching what the schema predicted. Weight loss in this culture gets rewarded out loud, by people who mean well, and that reward arrives early.
You’re so thin. I wish I was as skinny as you. What’s your secret? Those land as wins. They get cherished, saved, repeated back to yourself.
Walsh’s account of persistence has this in it explicitly. Coniglio and colleagues (2017) describe restriction as initially maintained by operant conditioning, with successful weight loss and external praise as the salient rewards, and then becoming a habit that runs on regardless of whether those rewards are still around (Coniglio et al., 2017). Their numbers: once habit strength went into the model it explained 27.9% of the variance in restriction, and cognitive restraint explained none.
That’s a sharper claim than the one I started with, and it’s less flattering to the framework. I wanted to say the praise keeps confirming the schema so nothing ever disconfirms it. The habit account says something harder. The praise gets the thing built and then becomes irrelevant, which is why the disorder outlives the compliments and why “nobody says that to me anymore” changes nothing at all. A behavior that no longer needs its reward is worse news than one still being paid.
And the direct evidence runs against the version I wanted. In non-clinical samples, positive appearance comments predict LESS body dissatisfaction, not more. Herbozo, Menzel and Thompson (2013) found that across 924 undergraduate women, and Liu et al. (2022) found the same in 620 people surveyed six times a day for a week, where nothing at all predicted disordered-eating urges. Fuller-Tyszkiewicz et al. (2019) points the same direction (Herbozo et al., 2013; Liu et al., 2022; Fuller-Tyszkiewicz et al., 2019). The one study running my way, Herbozo et al. (2017), found it in a subgroup of ethnic-minority women and only for negative reactions TO the commentary rather than the commentary itself.
Every one of those samples is non-clinical, which is the honest limitation on them and not a rescue for me. It’s also the gap somebody should go fill. Nobody has asked whether praise for weight loss lands differently on a person already on the trajectory than it does on an undergraduate who isn’t. That’s a real question, it’s answerable, and it’s the sort of thing an entire literature can miss by sampling the wrong people.
The negative side of the same coin is much better evidenced. Haines and colleagues (2006) followed 2,516 adolescents over five years and found weight teasing predicted incident frequent dieting in girls at 18.2% against 11.0%, and incident binge eating in boys at 4.1% against 1.4% (Haines et al., 2006). Menzel et al. (2010) put the pooled association at .32 to .39 across four outcomes.
The prospective data on pressure is narrower than the story too, and I’d rather show you the gap than paper it. Presnell, Bearman and Stice (2004) followed 531 adolescents and found that perceived pressure to be thin FROM PEERS predicted increases in body dissatisfaction, while thin-ideal internalization and perceived pressure from family, dating partners and media did not (Presnell et al., 2004). Media is the usual suspect and media isn’t what came up. Peer pressure was. Which is two brothers in a car, not a magazine.
Status: the early-reward account holds, the “praise keeps it going” account does not. Praise is in the model as a starter, on Walsh’s authority and Coniglio’s data. As a maintainer it has one supportive subgroup finding and three studies pointing the other way, all of them in people who don’t have the illness. More investigation needed, and I’d start with the sampling.
A question that comes up from people who have recovered is why there’s no physical fitness component once the danger has passed. Teach somebody how to make good choices about food and movement and you hand back a legitimate way to run the same controls. It’s a fair question, the field has looked at it, and the answer is more complicated than either yes or no.
Start with why it’s hard. Boltri and colleagues (2026) reviewed 17 studies of the state urge to be physically active in anorexia (Boltri et al., 2026). Three findings. The urge persists regardless of what BMI does, including through recovery. It looks like it functions as emotion regulation rather than as a stable trait. And higher levels of it during weight restoration predicted worse weight maintenance afterward.
In model terms that may be a second controller output aimed at the same regulated variable. Restriction and movement are two responses doing one job, which is why treating the movement as ordinary exercise misses what it’s for in that person.
Is it safe? Mostly, and that’s the strongest thing here. Yin and colleagues (2026) pooled 11 randomized trials of physical activity added to usual care in anorexia (Yin et al., 2026). It was well tolerated, with no adverse effects on weight restoration.
Does it do anything? Less than you’d hope. In the same meta-analysis: no significant difference in body weight, BMI, body fat percentage, or quality of life. One thing moved and it moved hard. Muscle strength, SMD 0.99, CI 0.45 to 1.53.
Does treating the compulsive exercise itself work? The one randomized trial says no. Hay and colleagues (2018) randomized 78 adults to CBT for anorexia with or without eight embedded sessions of LEAP, the purpose-built compulsive exercise therapy (Hay et al., 2018). Everything improved over time and there were no significant differences between groups on the primary outcomes. The uncontrolled studies of the same program look wonderful, which is what uncontrolled inpatient studies always look like.
Bulimia does better, oddly. Mathisen and colleagues (2020) ran exercise plus dietary therapy against CBT in 149 women with bulimia or binge-eating disorder (Mathisen et al., 2020). Both produced medium to strong improvement, the exercise arm improved faster early, the gap closed by follow-up, and 30 to 50% responded favorably with no difference between groups.
Now the concern that should stop you, and it’s about timing rather than exercise. Carter and colleagues (2004) followed 51 weight-restored anorexia patients after discharge (Carter et al., 2004). 35% relapsed, the highest risk running from 6 to 17 months out, and one of five predictors of relapse was excessive exercise immediately after discharge. Not exercise. Exercise then. Which window would you have guessed?
And it isn’t the same illness in both directions. Stiles-Shields and colleagues (2015) looked at 201 adolescents and found driven exercise in 66.3% of the bulimia group against 23.1% of the anorexia group, predicting worse outcomes only in anorexia (Stiles-Shields et al., 2015). Far more common where it does less damage, which should make anybody cautious about one rule for both.
Where the field openly disagrees. Drabkin and colleagues (2017), writing on bone disease, say weight-bearing exercise should be avoided unless carefully monitored in the setting of weight restoration (Drabkin et al., 2017). The Safe Exercise at Every Stage group say safe, nutritionally supported activity enhances outcomes without hindering restoration. Both are narrative reviews, neither is a trial, and the split tracks the authors’ discipline. Bone doctors say wait. Exercise physiologists say go carefully.
The bone question isn’t small. Franzoni and colleagues (2014) scanned 79 adolescents and rescanned 46 a year later (Franzoni et al., 2014). BMI rose significantly. Bone density didn’t change at all. A year of weight restoration and the skeleton hadn’t caught up.
Her instinct about education was right, and it’s already in the guidelines. Cook and colleagues (2016) synthesized 11 guidelines for therapeutic exercise in eating disorder treatment, two of which are to include a psychoeducational component and to include a nutritional component (Cook et al., 2016). Somebody wondering forty years ago why nobody taught her about food and movement arrived at what the guideline writers arrived at. Nobody has tested education as relapse prevention. Two targeted searches for it came back empty.
So where does that leave somebody recovered and wondering? It probably won’t hurt you and it probably won’t do as much as you’re hoping. The moment you start matters more than what you do. Your bones may be a long way behind your weight! And if the wanting to move is loud, that’s the thing to talk about rather than the thing to act on.
If you’re in recovery and working this out for yourself, it’s a question for whoever is treating you rather than one to settle from a review. Wanting to move is not by itself a symptom, and this section is not the place to work out whether yours is.
Four, cheapest first.
Separate accuracy from confidence. The interoception literature keeps measuring one and reporting the other. Eliciting calibrated confidence alongside performance would settle what the dissociation actually is, and the instruments already exist. This tests the phenomenon rather than the framework, which is the debt above.
Track the habit measure against relapse timing. If restriction persists because it’s running as a shortcut, then how habitual it is at discharge should predict who relapses, over and above weight at discharge. That’s a prediction the affect-regulation account doesn’t obviously make, which is what makes it worth running.
Test whether restriction and bingeing trade off within a person. Two opposing loops predicts that reducing one shows up in the other. Nothing in the current literature is designed to catch that.
Stratify a combination trial for bulimia by which half is driving. If low trust in the body’s read and the belief that eating is the only thing that fixes the feeling mark an estimator-driven loop (Chester et al., 2024), then adding medication to CBT should help those patients most and add little for the ones whose bingeing runs mostly as habit. This needs the moderators measured at baseline and then an interaction test, and detecting an interaction takes a bigger sample than detecting the main effect, so it is not the cheap one. Nobody has split a combination trial this way… which is strange, given that the moderators are two short questionnaires.
Khalsa, Portnoff, McCurdy-McKinnon and Feusner (2017) reviewed 27 studies of relapse in anorexia and found reported relapse rates running from 9% to 52%, rising as follow-up got longer, with the risk concentrated in the first year after treatment (Khalsa et al., 2017). That spread is the finding. Definitions of relapse varied so much between studies that comparing across them is difficult, and if the studies can’t agree what recovery means, a mechanistic claim has no fixed thing to be scored against.
The habit account of restriction is the strongest thing here, and the credit for it belongs to Walsh (2013) and his group. The treatment evidence backs that half and only that half. The body image work is the second strongest, and unlike the interoception material it comes with an intervention that moves the thing the model says is stored. The two-halves reading of the interoception data is a reinterpretation, it IS mine, it rests on a meta-analysis of one disputed task, and it still owes a prediction. The bulimia extension is one small self-report study. The two-opposing-loops idea is untested.
If you take one thing into a room from this: the restriction is running without deliberation, so understanding it isn’t the lever. For anorexia the lever is weight restoration, run by people who do this work. Bulimia’s target is a different one and this addendum doesn’t solve it.
Arcelus, J., Mitchell, A. J., Wales, J., & Nielsen, S. (2011). Mortality rates in patients with anorexia nervosa and other eating disorders: A meta-analysis of 36 studies. Archives of General Psychiatry, 68(7), 724–731. https://doi.org/10.1001/archgenpsychiatry.2011.74
Attia, E., Steinglass, J. E., Walsh, B. T., Wang, Y., Wu, P., Schreyer, C., Wildes, J., Yilmaz, Z., Guarda, A. S., Kaplan, A. S., & Marcus, M. D. (2019). Olanzapine versus placebo in adult outpatients with anorexia nervosa: A randomized clinical trial. American Journal of Psychiatry, 176(6), 449–456. https://doi.org/10.1176/appi.ajp.2018.18101125
Austin, A., Anderson, A. G., Lee, J., Vander Steen, H., Savard, C., Bergmann, C., Singh, M., Devoe, D., Gorrell, S., Patten, S., Le Grange, D., & Dimitropoulos, G. (2025). Efficacy of eating disorder focused family therapy for adolescents with anorexia nervosa: A systematic review and meta-analysis. International Journal of Eating Disorders, 58(1), 3–36. https://doi.org/10.1002/eat.24252
Bacaltchuk, J., Hay, P., & Trefiglio, R. (2001). Antidepressants versus psychological treatments and their combination for bulimia nervosa. Cochrane Database of Systematic Reviews, 2001(4), CD003385. https://doi.org/10.1002/14651858.CD003385
Barca, L., & Pezzulo, G. (2020). Keep your interoceptive streams under control: An active inference perspective on anorexia nervosa. Cognitive, Affective, & Behavioral Neuroscience, 20(2), 427–440. https://doi.org/10.3758/s13415-020-00777-6
Becker, A. E. (2004). Television, disordered eating, and young women in Fiji: Negotiating body image and identity during rapid social change. Culture, Medicine and Psychiatry, 28(4), 533–559. https://doi.org/10.1007/s11013-004-1067-5
Becker, A. E., Burwell, R. A., Gilman, S. E., Herzog, D. B., & Hamburg, P. (2002). Eating behaviours and attitudes following prolonged exposure to television among ethnic Fijian adolescent girls. The British Journal of Psychiatry, 180(6), 509–514. https://doi.org/10.1192/bjp.180.6.509
Benini, L., Todesco, T., Dalle Grave, R., Deiorio, F., Salandini, L., & Vantini, I. (2004). Gastric emptying in patients with restricting and binge/purging subtypes of anorexia nervosa. The American Journal of Gastroenterology, 99(8), 1448–1454. https://doi.org/10.1111/j.1572-0241.2004.30246.x
Boltri, M., Gabutti, C., Pignattai, L., Brusa, F., Apicella, E., Castelnuovo, G., Sapuppo, W., Mendolicchio, L., Hanachi, M., Hebebrand, J., & Paslakis, G. (2026). The state urge to be physically active in anorexia nervosa: A systematic review of mechanisms and clinical implications. Journal of Psychiatric Research, 201, 149–162. https://doi.org/10.1016/j.jpsychires.2026.06.022
Carter, J. C., Blackmore, E., Sutandar-Pinnock, K., & Woodside, D. B. (2004). Relapse in anorexia nervosa: A survival analysis. Psychological Medicine, 34(4), 671–679. https://doi.org/10.1017/S0033291703001168
Chester, M. A., Viranda, T., Kaye, W. H., & Berner, L. A. (2024). Evaluating the predictions of an interoceptive inference model of bulimia nervosa. Journal of Eating Disorders, 12(1), 57. https://doi.org/10.1186/s40337-024-01010-2
Conceição, I. S. R., Garcia-Burgos, D., de Macêdo, P. F. C., Nepomuceno, C. M. M., Pereira, E. M., Cunha, C. de M., Ribeiro, C. D. F., & de Santana, M. L. P. (2023). Habits and persistent food restriction in patients with anorexia nervosa: A scoping review. Behavioral Sciences, 13(11), 883. https://doi.org/10.3390/bs13110883
Coniglio, K. A., Becker, K. R., Franko, D. L., Zayas, L. V., Plessow, F., Eddy, K. T., & Thomas, J. J. (2017). Won’t stop or can’t stop? Food restriction as a habitual behavior among individuals with anorexia nervosa or atypical anorexia nervosa. Eating Behaviors, 26, 144–147. https://doi.org/10.1016/j.eatbeh.2017.03.005
Cook, B. J., Wonderlich, S. A., Mitchell, J. E., Thompson, R., Sherman, R., & McCallum, K. (2016). Exercise in eating disorders treatment: Systematic review and proposal of guidelines. Medicine & Science in Sports & Exercise, 48(7), 1408–1414. https://doi.org/10.1249/MSS.0000000000000912
Cornelissen, P. L., & Tovée, M. J. (2021). Targeting body image in eating disorders. Current Opinion in Psychology, 41, 71–77. https://doi.org/10.1016/j.copsyc.2021.03.013
Drabkin, A., Rothman, M. S., Wassenaar, E., Mascolo, M., & Mehler, P. S. (2017). Assessment and clinical management of bone disease in adults with eating disorders: A review. Journal of Eating Disorders, 5, 42. https://doi.org/10.1186/s40337-017-0172-0
Fluoxetine Bulimia Nervosa Collaborative Study Group. (1992). Fluoxetine in the treatment of bulimia nervosa: A multicenter, placebo-controlled, double-blind trial. Archives of General Psychiatry, 49(2), 139–147. https://doi.org/10.1001/archpsyc.1992.01820020059008
Foerde, K., Steinglass, J. E., Shohamy, D., & Walsh, B. T. (2015). Neural mechanisms supporting maladaptive food choices in anorexia nervosa. Nature Neuroscience, 18(11), 1571–1573. https://doi.org/10.1038/nn.4136
Franzoni, E., Ciccarese, F., Di Pietro, E., Facchini, G., Moscano, F., Iero, L., Monaldi, A., Battista, G., & Bazzocchi, A. (2014). Follow-up of bone mineral density and body composition in adolescents with restrictive anorexia nervosa: Role of dual-energy X-ray absorptiometry. European Journal of Clinical Nutrition, 68(2), 247–252. https://doi.org/10.1038/ejcn.2013.254
Fuller-Tyszkiewicz, M., Chhouk, J., McCann, L. A., Urbina, G., Vuo, H., Krug, I., Ricciardelli, L., Linardon, J., Broadbent, J., Heron, K., & Richardson, B. (2019). Appearance comparison and other appearance-related influences on body dissatisfaction in everyday life. Body Image, 28, 101–109. https://doi.org/10.1016/j.bodyim.2019.01.002
Gend, C., Remy, L., & Lutz, A. P. C. (2026). A systematic review and meta-analysis of cardiac interoceptive accuracy in anorexia nervosa. International Journal of Eating Disorders, 59(6), 1208–1224. https://doi.org/10.1002/eat.70064
Haines, J., Neumark-Sztainer, D., Eisenberg, M. E., & Hannan, P. J. (2006). Weight teasing and disordered eating behaviors in adolescents: Longitudinal findings from Project EAT (Eating Among Teens). Pediatrics, 117(2), e209–e215. https://doi.org/10.1542/peds.2005-1242
Hay, P., Touyz, S., Arcelus, J., Pike, K., Attia, E., Crosby, R. D., Madden, S., Wales, J., La Puma, M., Heriseanu, A. I., Young, S., & Meyer, C. (2018). A randomized controlled trial of the compuLsive Exercise Activity TheraPy (LEAP): A new approach to compulsive exercise in anorexia nervosa. International Journal of Eating Disorders, 51(8), 999–1004. https://doi.org/10.1002/eat.22920
Herbozo, S., Menzel, J. E., & Thompson, J. K. (2013). Differences in appearance-related commentary, body dissatisfaction, and eating disturbance among college women of varying weight groups. Eating Behaviors, 14(2), 204–206. https://doi.org/10.1016/j.eatbeh.2013.01.013
Herbozo, S., Stevens, S. D., Moldovan, C. P., & Morrell, H. E. R. (2017). Positive comments, negative outcomes? The potential downsides of appearance-related commentary in ethnically diverse women. Body Image, 21, 6–14. https://doi.org/10.1016/j.bodyim.2017.01.008
Heruc, G. A., Little, T. J., Kohn, M., Madden, S., Clarke, S., Horowitz, M., & Feinle-Bisset, C. (2019). Appetite perceptions, gastrointestinal symptoms, ghrelin, peptide YY and state anxiety are disturbed in adolescent females with anorexia nervosa and only partially restored with short-term refeeding. Nutrients, 11(1), 59. https://doi.org/10.3390/nu11010059
Khalsa, S. S., Portnoff, L. C., McCurdy-McKinnon, D., & Feusner, J. D. (2017). What happens after treatment? A systematic review of relapse, remission, and recovery in anorexia nervosa. Journal of Eating Disorders, 5(1), 20. https://doi.org/10.1186/s40337-017-0145-3
Linardon, J., & Wade, T. D. (2018). How many individuals achieve symptom abstinence following psychological treatments for bulimia nervosa? A meta-analytic review. International Journal of Eating Disorders, 51(4), 287–294. https://doi.org/10.1002/eat.22838
Liu, S., Fuller-Tyszkiewicz, M., Eddy, S., Liu, X., Portingale, J., Giles, S., & Krug, I. (2022). The effects of appearance-based comments and non-appearance-based evaluations on body dissatisfaction and disordered eating urges: An ecological momentary assessment study. Behavior Therapy, 53(5), 807–818. https://doi.org/10.1016/j.beth.2022.01.002
Lock, J., Le Grange, D., Agras, W. S., Moye, A., Bryson, S. W., & Jo, B. (2010). Randomized clinical trial comparing family-based treatment with adolescent-focused individual therapy for adolescents with anorexia nervosa. Archives of General Psychiatry, 67(10), 1025–1032. https://doi.org/10.1001/archgenpsychiatry.2010.128
Mathisen, T. F., Rosenvinge, J. H., Friborg, O., Vrabel, K., Bratland-Sanda, S., Pettersen, G., & Sundgot-Borgen, J. (2020). Is physical exercise and dietary therapy a feasible alternative to cognitive behavior therapy in treatment of eating disorders? A randomized controlled trial of two group therapies. International Journal of Eating Disorders, 53(4), 574–585. https://doi.org/10.1002/eat.23228
Menzel, J. E., Schaefer, L. M., Burke, N. L., Mayhew, L. L., Brannick, M. T., & Thompson, J. K. (2010). Appearance-related teasing, body dissatisfaction, and disordered eating: A meta-analysis. Body Image, 7(4), 261–270. https://doi.org/10.1016/j.bodyim.2010.05.004
Norris, M. L., Harrison, M. E., Isserlin, L., Robinson, A., Feder, S., & Sampson, M. (2016). Gastrointestinal complications associated with anorexia nervosa: A systematic review. International Journal of Eating Disorders, 49(3), 216–237. https://doi.org/10.1002/eat.22462
Presnell, K., Bearman, S. K., & Stice, E. (2004). Risk factors for body dissatisfaction in adolescent boys and girls: A prospective study. International Journal of Eating Disorders, 36(4), 389–401. https://doi.org/10.1002/eat.20045
Robinson, P. H. (1989). Perceptivity and paraceptivity during measurement of gastric emptying in anorexia and bulimia nervosa. The British Journal of Psychiatry, 154(3), 400–405. https://doi.org/10.1192/bjp.154.3.400
Solmi, M., Wade, T. D., Byrne, S., Del Giovane, C., Fairburn, C. G., Ostinelli, E. G., De Crescenzo, F., Johnson, C., Schmidt, U., Treasure, J., Favaro, A., Zipfel, S., & Cipriani, A. (2021). Comparative efficacy and acceptability of psychological interventions for the treatment of adult outpatients with anorexia nervosa: A systematic review and network meta-analysis. The Lancet Psychiatry, 8(3), 215–224. https://doi.org/10.1016/S2215-0366(20)30566-6
Steinglass, J. E., & Walsh, B. T. (2016). Neurobiological model of the persistence of anorexia nervosa. Journal of Eating Disorders, 4(1), 19. https://doi.org/10.1186/s40337-016-0106-2
Stiles-Shields, C., Bamford, B., Lock, J., & Le Grange, D. (2015). The effect of driven exercise on treatment outcomes for adolescents with anorexia and bulimia nervosa. International Journal of Eating Disorders, 48(4), 392–396. https://doi.org/10.1002/eat.22281
Walsh, B. T. (2013). The enigmatic persistence of anorexia nervosa. American Journal of Psychiatry, 170(5), 477–484. https://doi.org/10.1176/appi.ajp.2012.12081074
Yin, Z., Liu, M., Cai, J., Kong, N., Peng, L., Gong, H., Du, Z., & Zhang, K. (2026). The effect of physical activity on health outcomes in individuals with anorexia nervosa: A systematic review and meta-analysis study. BMC Psychiatry, 26(1), 532. https://doi.org/10.1186/s12888-026-08110-0
Zipfel, S., Wild, B., Groß, G., Friederich, H.-C., Teufel, M., Schellberg, D., Giel, K. E., de Zwaan, M., Dinkel, A., Herpertz, S., Burgmer, M., Löwe, B., Tagay, S., von Wietersheim, J., Zeeck, A., Schade-Brittinger, C., Schauenburg, H., & Herzog, W. (2014). Focal psychodynamic therapy, cognitive behaviour therapy, and optimised treatment as usual in outpatients with anorexia nervosa (ANTOP study): Randomised controlled trial. The Lancet, 383(9912), 127–137. https://doi.org/10.1016/S0140-6736(13)61746-8
Part of the Prediction, Control, and the Regulating Mind series. Companion addenda: Self-Injury as a Regulatory Behavior and Amusement, Awe and Curiosity.
This addendum was improved substantially by a reader who recovered from anorexia and who took the time to say what was missing. The gaps she found became two of its sections.