This is a lay guide to our paper Prediction gone awry: A computational framework for emotion dysregulation across autism and ADHD
I am autistic and have ADHD, so emotion dysregulation is not an abstract topic for me. I know what it is like for an emotion to be far more complicated on the inside than it looks from the outside. Something can feel unexpectedly intense. A feeling can linger long after the event has passed. At other times, attention and emotion can shift so quickly that it is hard to find stable footing.
These experiences are usually described by what others can see: irritability, impulsivity, escalation, difficulty calming down. Those descriptions may be accurate, but they do not explain what the brain and body are doing—or why two people who look similarly “dysregulated” may have reached that point through very different pathways.
That gap led to our paper, Prediction Gone Awry: A Computational Framework for Emotion Dysregulation Across Autism and Attention-Deficit/Hyperactivity Disorder where we ask a fairly simple question: What if emotion regulation depends partly on how the brain predicts what will happen next?
To explore that question, we draw on an established neuroscientific framework called predictive coding. The basic idea is that the brain does not passively wait for information. It is constantly forming expectations about the world and the body, comparing them with what actually happens, and adjusting when the two do not match. That mismatch is called a prediction error, the nervous system registering, This is not what I expected. Prediction errors are not bad. We need them to learn. If I expect a room to be quiet and a fire alarm suddenly goes off, that mismatch should grab my attention. Problems may arise when an unexpected signal receives too much weight, too little weight, or a different amount of weight from moment to moment. A small change can then feel much larger than it looks to someone else. Or the nervous system may continue responding as though the original situation is still happening, even after it has passed.
This gives us several possible routes to dysregulation. Incoming signals may feel too loud. Expectations may update too slowly, leaving an emotion “stuck.” Or the importance assigned to new information may fluctuate rapidly, which we suggest may be especially relevant to ADHD.
These processes could help explain some patterns of emotion dysregulation in autism, ADHD, and their co-occurrence, often called AuDHD. I deliberately say could. This is a theoretical framework that generates testable hypotheses, not a claim that every autistic or ADHD person’s brain works this way. The experiences are real. Our particular computational explanation still needs to be tested.
The brain has to decide what to trust
Imagine I am about to give a talk. My brain does not enter the room with a blank slate. It brings expectations shaped by previous talks, other people’s reactions, sensations in my body, and what I know about the setting. These expectations are called priors, the brain’s existing best guesses about what is likely to happen. Meanwhile, new information is arriving. My heart may beat faster, my muscles may tense, or the room may be noisier than I expected. Some of this fits my expectations; some creates a prediction error.
The brain then has to decide what to trust. Should it rely on the prior expectation that the talk will be manageable? Or should it give more weight to my racing heart, the unexpected noise, or an ambiguous expression in the audience?
In predictive coding, this weighting is called precision. Precision does not mean accuracy here. It means how much confidence the brain places in a signal. A signal assigned high precision is treated as reliable and important; one assigned low precision has less influence. This balance matters. If the brain gives new information too little weight, it may miss a genuine change. If it gives every unexpected sensation too much weight, the world and the body, can feel constantly surprising.
Suppose my heart rate rises before the talk. The brain might interpret that as ordinary anticipation: I am activated because I am about to speak. But if the same sensation is assigned very high precision, it may take over the interpretation: Something is wrong. I am not safe. The sensation alone is not the whole emotion. What matters is how it is interpreted alongside prior experience, the environment, and other incoming information. This is part of interoception: sensing and making sense of what is happening inside the body, including changes in heart rate, breathing, temperature, tension, hunger, or nausea.
The system must also keep learning. If the talk goes well, the brain should use that new evidence to revise its expectations. This is called updating. But updating can happen at different speeds. If expectations change too readily, every new event can push the system in another direction. If they change too slowly, new evidence may not shift the emotional state. The talk may be over, yet the nervous system continues responding as though the situation is unresolved.
This is one way an emotion may become “stuck.” The person is not choosing to hold onto it; the system responsible for revising emotional expectations may simply be updating more slowly.
One final term is gain control, the brain’s moment-to-moment adjustment of how strongly incoming signals are amplified or dampened. It is a little like a continuously moving volume control. When gain fluctuates, a cue that barely registers at one moment may feel urgent at another.
These concepts let us ask more specific questions about dysregulation. Are existing expectations too strong or too weak? How much weight is being given to new information? How quickly are expectations changing? And is the amplification of incoming signals stable?
These questions do not replace a person’s history, relationships, sensory environment, or social context. Predictive coding does not mean emotion happens only inside the brain. Expectations are learned through experience. A chronically unpredictable, inaccessible, or threatening environment may give the nervous system good reason to stay alert. What looks like miscalibration in a laboratory may sometimes be an understandable adaptation to the world someone has had to navigate.
What this could mean in autism
Several predictive-coding accounts of autism propose that unexpected sensory information may sometimes be assigned unusually high precision. In other words, a mismatch between expectation and experience may carry a lot of weight. A sound others dismiss, a change in routine, or an unexpected bodily sensation may continue demanding attention rather than fading into the background. This is called overprecision of prediction errors. It does not mean perceiving the world “too accurately.” It means that the brain may treat a mismatch as especially reliable and important. That could help explain why seemingly minor changes sometimes have major effects. Of course, minor is usually an outsider judgment. A change may look small to someone who can quickly absorb it into the larger situation. It may not feel small when it remains vivid and unresolved for the person experiencing it.
Our framework adds a second possibility: slow updating of priors. Even when new information arrives, existing expectations may take longer to change. I may understand someone telling me that a stressful situation is over without my body immediately responding as though it is over.
These processes could produce both strong reactions to new information and difficulty shifting away from an established emotional state. The scientific term for that persistence is emotional inertia. An emotion does not have to remain equally intense; it simply continues shaping attention, bodily responses, and how the situation is interpreted. From the outside, that can look like dwelling, refusing to move on, or overreacting. From the inside, it may feel more like the nervous system has not yet received or trusted enough evidence that the situation has changed.
This may be one reason predictability matters. Routines, advance notice, clear communication, control over sensory input, and even stimming can provide structure when the environment is difficult to anticipate. They are not automatically symptoms to eliminate; they can be sensible ways of reducing cognitive and physiological load.
And sometimes the environment really is the problem. If a light is painful, a room is chaotic, or communication is unclear, the answer may be to change the environment, not train the autistic person to tolerate it. Predictive coding should not become another way to relocate every difficulty into the disabled person.
Nor is there one autistic predictive profile. Findings differ across people, tasks, ages, and sensory systems. Our proposal is narrower: some autistic people, in some contexts, may experience dysregulation through highly weighted prediction errors, slow updating, or both.
That qualification matters to me as an autistic author. Computational language can make a hypothesis sound universal and settled. A useful framework should help us ask more precise questions about individual experience, not compress autistic people into another single story.
What this could mean in ADHD
For ADHD, we propose a somewhat different imbalance. The system may have difficulty maintaining stable expectations about what matters from one moment to the next.
One part of this account involves weaker priors, expectations held with less confidence. When the brain relies less strongly on what it already expects, incoming information has more power to redirect attention and emotion. The second part is unstable gain control. Gain control adjusts how strongly incoming signals are amplified or dampened. In ADHD, that amplification may fluctuate. Something that barely registers at one moment may suddenly feel urgent at another.
Imagine receiving a mildly critical email. At first, it seems unimportant. Ten minutes later, one sentence becomes impossible to ignore and triggers a rush of frustration or anxiety. Then something else captures my attention, and the intensity drops almost as quickly as it appeared. The system may be changing how much importance it gives the same information.
The scientific term for rapid emotional shifts is affective lability: affect means emotional state, and lability means that it changes readily. This is not simply “being emotional.” It refers to how quickly and variably emotions rise, fall, or change direction. It may also help explain why attention and emotion feel so intertwined in ADHD. Both depend partly on deciding which signals matter now. If that weighting fluctuates, a new event can abruptly take control of both.
The distinction between a “stuck” autistic emotion and a rapidly shifting ADHD emotion looks neat on paper. My own experience of being autistic and having ADHD is less tidy. Attention may move while my body remains activated. A new feeling may arrive before the earlier one has resolved. Rapid shifts and emotional inertia can coexist.
ADHD is also highly varied. Some people experience high arousal and visible emotional volatility; others have lower arousal or difficulty mobilizing attention and action. We are not claiming that every person with ADHD has weak priors or unstable gain. Evidence for altered gain comes mainly from studies of attention, perception, and learning. Applying it to emotion is still a hypothesis. A technical label is not progress by itself. “Unstable gain” should not become a fancier way of saying “emotionally impulsive.” The framework matters only if researchers can measure the proposed process and distinguish it from stress, sleep, medication effects, executive demands, and the environment.
The larger point is that emotional shifts in ADHD may not reflect a failure to care or try. They may partly involve instability in the systems that decide what receives attention and how strongly it is felt.
When autism and ADHD occur together
Autism and ADHD frequently co-occur. The informal term AuDHD describes people who are both autistic and have ADHD, although it is not a separate clinical diagnosis. Rather than treating AuDHD as a simple sum of traits, our framework asks whether the underlying processes might interact. We call this a computational collision. This does not mean that two defective systems crash together. It means that slow updating and unstable gain may operate within the same predictive system and alter each other’s effects.
Slow updating may keep the nervous system in an established emotional state. At the same time, unstable gain may shift attention and emotional intensity rapidly in response to new information. A person might remain physically activated after a stressful interaction while moving through anger, worry, distraction, and back to the original event. The emotions appear to change, but the underlying state never fully settles.
AuDHD can also involve needs that look contradictory from the outside. I may need predictability but struggle to create or maintain routines. I may seek novelty while finding unexpected change destabilizing. I may become intensely absorbed in something but be unable to direct that attention when I need it.
Separating updating speed from gain control offers one way to understand these tensions. Different parts of the predictive system may be changing at different speeds. What looks inconsistent may reflect genuinely competing needs, not a lack of effort or self-knowledge.
This is also the most speculative part of our model. We do not yet know whether AuDHD produces an interaction between these mechanisms, whether they simply contribute independently, or whether there are several different AuDHD profiles.
Testing this will require more than comparing average questionnaire scores. Researchers would need to measure how quickly expectations update and how much signal weighting varies from trial to trial, then ask whether their interaction explains experiences that neither measure captures alone. Lived experience can sharpen that scientific question without proving the answer. Accounts of needing but struggling to create predictability, or of rapidly changing emotions within prolonged distress, point toward dynamics that broad diagnostic labels may miss. The computational model must then earn its value by explaining and predicting those dynamics.
How could we test any of this?
A theoretical framework is useful only if it makes predictions that could turn out to be wrong. Otherwise, computational terms risk becoming impressive-sounding ways to redescribe what we already know. Precision, gain, and updating speed cannot be observed directly. Researchers must estimate them from behavior and physiology. Our paper identifies several possible measures, but none maps neatly onto a single computational process.
One is heart-rate variability, or HRV: small variations in the time between heartbeats. A healthy heart does not beat like a metronome; the intervals change as the autonomic nervous system responds to breathing, activity, and stress. Rather than looking only at resting HRV, we may learn more from how it changes during stress and how quickly it recovers. Slow recovery could be consistent with slow emotional updating, but it would not prove it.
Pupils provide another window. They respond not only to light but also to attention, effort, arousal, and surprise. A large pupil response to something unexpected might be consistent with a strongly weighted prediction error, while highly variable responses could point toward unstable gain control. But many processes affect pupil size, so it cannot provide an answer by itself.
We also discuss mismatch negativity, or MMN, a pattern of brain activity measured using electroencephalography, or EEG. If someone hears a series of identical tones followed by an occasional different one, the brain registers that deviation even without conscious effort. The size, timing, and variability of MMN may tell us something about how the nervous system processes unexpected information. Still, a larger or smaller MMN does not translate directly into “more” or “less” precision.
Then there is interoception: sensing and interpreting signals from inside the body. Traditional tasks often ask people to count or detect their heartbeats, but performance can also reflect attention, guessing, and beliefs about heart rate. Heartbeat-evoked potentials, or HEPs, offer another approach by measuring brain activity linked to each heartbeat. Yet an HEP is not a direct readout of “interoceptive precision” either.
This caution matters because HRV, pupil responses, MMN, and HEPs are all multidetermined; many biological and psychological processes shape them. None is a thermometer for predictive dysregulation. A strong study would therefore combine several kinds of evidence. A participant might complete a task in which the emotional meaning of cues changes, allowing researchers to estimate how quickly expectations update. At the same time, we could record HRV, pupil responses, or EEG. Brief check-ins during daily life, called ecological momentary assessment, could show whether the laboratory results relate to actual episodes of dysregulation. If slower updating on the task predicts slower physiological recovery and longer emotional episodes in daily life, that would support the model. If these measures do not relate as predicted, the model would need to change.
Eventually, this approach might help explain differences within diagnoses. Two people may show similar outward dysregulation while one responds strongly to unexpected signals, another recovers slowly, and a third fluctuates from moment to moment. For now, though, this is a research agenda, not a clinical test. No single HRV, pupil, or EEG measurement can reveal someone’s computational “type.” The bridge between the theory and these measures still has to be built.
What might this change about support?
The practical value of this framework is that it shifts attention from the visible size of an emotion to what may be sustaining it. Instead of asking only, “How do we stop this response?” we can ask, “What is the nervous system responding to, and what might help?” That is not the same as making neurodivergent people appear calmer or more typical. Emotion regulation should not mean hiding distress so that others find it less disruptive. Nor is every intense emotion a clinical problem. Anger, fear, excitement, and frustration may be entirely reasonable responses to the circumstances.
Sometimes the environment is the right place to intervene. Predictable routines, advance notice, clear communication, sensory modifications, and control over the timing or intensity of input can reduce the number of unexpected signals a person must manage. These are not simply forms of avoidance. They can provide the stability needed for learning, communication, and participation.
The framework also points to several approaches that researchers could test, although none is a reliable or universal solution. HRV biofeedback combines paced breathing with feedback about heart rhythms. Neurofeedback provides feedback about patterns of brain activity. Interoceptive training focuses on noticing and interpreting signals from inside the body. Some people may find these approaches useful, while others may experience little benefit—or find increased attention to bodily sensations uncomfortable or distressing. Gradual exposure is another possibility, particularly when it produces expectancy violation, new evidence that an anticipated outcome does not occur. But its effects vary, and it should not be used to make someone endure pain, sensory overload, or an inaccessible environment. At most, our framework suggests hypotheses about who might benefit, under what conditions, and through which mechanism. It does not establish any of these approaches as the answer.
Our framework is therefore not a menu that matches one physiological result to one treatment. It offers questions for future research. Do changes in updating speed accompany improvements after exposure? Does HRV recovery relate to how long an emotional state persists? Does more stable pupil-linked arousal accompany more stable attention and emotion? If these relationships hold, support could become more individualized. Someone overwhelmed by unpredictable sensory information may need greater environmental control. Someone whose distress continues after a situation changes may need help noticing and trusting evidence of safety. Someone with highly variable arousal may need greater physiological stability.
For me, the most important shift is away from treating dysregulation as noncompliance or a failure of willpower. A person may know that a situation is over while their nervous system has not caught up. They may want to regulate while lacking the conditions that make regulation possible. A computational explanation does not make emotion less personal or remove its social context. Ideally, it gives us one more way to understand what is happening—and leads to more precise support with less blame.
Where the framework goes from here
The title Prediction Gone Awry is intentionally provocative. It does not mean that autistic or ADHD brains are simply broken prediction machines. Prediction develops through the relationship among the brain, body, and environment. What looks poorly calibrated in one setting may be an understandable adaptation to uncertainty, sensory overload, inconsistent expectations, or repeated threat.
Predictive coding also cannot explain every form of emotion dysregulation. Sleep, pain, trauma, communication barriers, medication, sensory conditions, executive demands, relationships, and the wider environment all matter. This framework may help connect some of them, but it should not become an explanation for everything.
What it does offer is a sharper set of questions. Is an unexpected signal receiving unusually high weight? Is an emotional state persisting because expectations are updating slowly? Is the importance assigned to new information fluctuating? And do these processes interact?
Testing this will require studies that combine computational tasks with physiology, behavior, and emotional experiences in daily life. Researchers also need to look beyond average differences between diagnostic groups. Autism, ADHD, and AuDHD each include enormous variation.
That research must include people with different communication styles, intellectual abilities, sensory profiles, and support needs. A model cannot claim to explain neurodevelopmental diversity if it is tested on only a narrow slice of the people it is supposed to describe.
Lived experience can help reveal distinctions that behavior alone may miss: an emotion may change while the body remains activated; novelty can be appealing while unexpected change is destabilizing; a person can understand that a situation is over without yet feeling that it is over.
But lived experience and computational theory have different jobs. First-person accounts can show researchers which questions they are missing. The model must then make predictions that can be measured, challenged, and potentially disproved.
As an autistic person with ADHD, I do not need a model to tell me that emotions can feel too loud, shift rapidly, or become stuck. Its value lies in helping explain how these experiences might coexist, why similar outward behavior may arise through different pathways, and why the same support will not work for everyone.
Prediction Gone Awry does not claim to have solved emotion dysregulation in autism and ADHD. It offers a way to ask better questions about what the nervous system is predicting, which signals it trusts, and how readily it can change course. Better questions will not fix inaccessible environments. But they may help us offer support that is more individualized, less judgmental, and better aligned with what a person is actually experiencing.
The full open-access paper is:
Srinivasan, H., Cascio, C. J., & Wallace, M. T. (2026). Prediction gone awry: A computational framework for emotion dysregulation across autism and attention-deficit/hyperactivity disorder. Journal of Child and Adolescent Psychopharmacology, 36, 461–472. https://doi.org/10.1177/10445463261448490
No comments:
Post a Comment