GLP-1s... Reduce Schizophrenia Risk?

Wat

While I did spend some time trying to make this idea as internally consistent as possible, I am painfully aware of how little I really know about these various theories of cognition relative to people who actually work in the field. Please nitpick and challenge in the comments!

A co-resident of mine, one Quinn MacDougald, passed along a paper from Nature Medicine titled: Mapping the effectiveness and risks of GLP-1 receptor agonists by Xie et al. and it has some really interesting stuff in it. Just as a reminder, the GLP1RAs are the popular¹ weight-loss drugs like semaglutide (Ozempic) and tirzepartide (Zepbound/Mounjaro).

You might expect that I’m going to talk about them in the context of addiction. There have been rumblings for some time now that the GLP1RAs might be effective drugs for the treatment of various substance use disorders (SUDs). A few very large observational studies have shown significant associations between GLP1RA usage and a decreased risk for the various negative outcomes associated with SUDs (e.g. hospitalization with a SUD as the admitting problem).

The Xie et al. paper, however, shows something that I think is far more interesting. They found that the odds of acquiring a diagnosis of schizophrenia was something like 20% lower in individuals on GLP1RAs than in individuals not on a GLP1RA. That’s a bigger effect size than for every single SUD they looked at!

Adapted from Xie et al. 2025. Fig. 6

The first thing I thought was that this was just a secondary effect of reducing substance use across the board. If you’ve been doing psychiatry for long enough you have 100% seen people who are well past the normal age for developing psychosis, not even a whiff of a psychotic prodrome, who did a just a little too much weed/crack/meth and oops! now they have schizophrenia. The data is murky, but there seems to be a dose-response thing going on here as well, so it would make sense that if you reduce the amount of people who are using any drug really heavily you’ll also reduce the number of them who develop a permanent substance induced psychosis.

This is the obvious, most simple answer, but then why didn’t the paper find an effect of the GLP1RAs on the risk of developing any of the mood or anxiety disorders? A possible and interesting answer is that substance use does not increase the risk of developing mood disorders by nearly as much as is commonly assumed. I looked around a little and there doesn’t seem to be much in the literature that attempts to answer this question of causation for the mood disorders. Any other ideas?

Any Excuse I Can Get, Really

Here is my Galaxy Brain, too-clever-by-half-take that gives me an excuse to talk about my favorite theory of how the brain works. Predictive processing (PP).

You, reading this post, probably.

As a refresher, predictive processing contends that the brain is a hierarchical system whose goal is to try and create top-down models that accurately predict the bottom-up signals from lower levels in the processing hierarchy (for a more detailed explanation, see my essay Reconsidering Dopamine, or read Surfing Uncertainty by Andy Clark). In other words, our brains try to reduce the difference in prediction error between top-down models and bottom up information.

There are two obvious ways to reduce that error: (1) change your top down model to better match the bottom up signal or (2) act in ways that changes the bottom up signal to better match what is expected by the top down model.

One of the very cool ideas of predictive processing is that we use that second approach to produce action itself. For example, bending your index finger is accomplished by implementing a sequence of top down models that each predict the sensory inputs that would accompany the bending of your index finger, and the error is resolved by actually moving your finger. In other words, you move by imagining really, really hard that you want to move!

This is a fascinating idea, made even more compelling by experiencing it for yourself. Take a moment and extend your index finger, and then focus really hard on the thought of bending it, without actually doing so. It is not terribly hard to prevent yourself from bending your finger, but (at least for me) there is a sensation not only that your finger “wants” to move but that it already has!

If we can accept this framing, I can’t help but see the parallel here with the sensation of hunger. Hunger is typically accompanied by thoughts of eating food that satiates you and is dissipated by a series of actions that result in you actually eating that food and actually feeling satiated. Same with thirst. Hopefully, it shouldn’t be too weird for me to now draw the parallel of hunger/thirst to any experience of craving.²

Through the lens of PP, I think these experiences can be tied together quite nicely. I would argue that all of these experiences are examples of error in which there is an overwhelming preference to reduce the error through action, as opposed to modifying the top-down model. It might be that this preference is baked into the top-down model itself,³ and in the case of hunger might exist as something close to “I am not hungry.”

There’s one more hoop we have to jump through, though. PP says that our perceptual experiences occur when the error between our top-down and bottom-up models have been sufficiently resolved. The thing is, we’ve all had the experience of feeling a bit peckish, but then “forgetting” how hungry we are when we get caught up in doing something else. How is this possible when we haven’t “acted” to resolve the aforementioned error?

This is where the PP concept of “gain” comes in, which is pretty analogous to what we mean when we talk about “attention” and to some degree “salience.” The basic idea is that some error signals are more important than others at any given moment, and our brains need a way to say “Yeah man, I know that the mismatch between how satiated we’re supposed to feel right now and how satiated we actually are is really big, but we need to figure out if this guy yelling in our face right now is threatening send us to hell literally or metaphorically, so put that problem on the backburner for a moment.” If you tone down the magnitude of the error signal pushing you to eat, the mismatch disappears, and you don’t experience the sensation of hunger.

So, what if part of the GLP1RAs’ mechanism of action involves acting on this shared, underlying system of error modulation? This would explain nicely why a drug that is used for weight loss and appetite suppression also seems to reduce addictive behaviors.

Yeah but are you finally going to explain how schizophrenia fits in here now that you’re done explaining something you said you weren’t going to address in the second paragraph of this essay? I hear you say.

Uh, right, sorry about that.

In chapter 4 of Surfing Uncertainty, author Andy Clark discusses how it is well established that we perceive self-generated sensations as less intense than externally generated ones. We all know this anecdotally through the fact that you can’t tickle yourself… or at least not nearly as well as everyone else seems to be able to. It turns out that we’ve also consistently demonstrated this phenomena via experiment. For example, Shergill et al. did so in 2003 when they published their amazingly titled paper, Two Eyes for an Eye. They found that if you apply an external force to the tip of someone’s left finger and then ask them to reproduce that force with their right finger (i.e. you ask them to reproduce the sensation just produced in their left finger), they use about twice as much force as the original external force.

This seems a little strange and suggests to me that there is some sort of generalized/universally applied system (or systems) that discount self-generated sensation. The value of this general discounting starts to make a little more sense when Clark contends that normal individuals who display higher levels of sensory attenuation overall are less likely to form delusional beliefs.

Schizophrenics are rather famous for being bad at differentiating between internally and externally generated experiences, so in a 2005 paper Shergill et al. decided to see how schizophrenics matched up against healthy controls in a similar force-matching task. Though they still applied more force than necessary, schizophrenics used 15% less force than controls. The authors frame this as a “dysfunction in the ability to predict the sensory consequences of their actions,” but I think it may be indicative of a more global issue where schizophrenics just don’t modulate the gain on error signals related to interoceptive processes in the same way that normal people do.

So, if GLP-1RAs help with turning down the gain on these interoceptive error signals, I think we have a plausible mechanism by which they reduce schizophrenia risk! More concretely, I think they could reduce the salience attributed to various internal experiences and thus the attenuate the subsequent creation of various (delusional) top-down models that produces psychosis. The clearest signal that something real akin to this is going on here would be if GLP1RAs improve psychosis in already psychotic individuals, but I don’t know if anyone has looked at this already. There’s a free study idea for you.

---

1. But not new! Exenatide was approved in 2005

2. We talk about food cravings pretty frequently!

3. This is best demonstrated in a PP framework with the top-down models we can easily observe in our visual system when we view optical illusions. The checker shadow illusion is my favorite example of this. You can tell yourself all you want that the checkers are the same color, but even with the conscious knowledge that your top-down model is dead wrong, it will not change!