There’s a particular experience in grief that’s hard to describe without sounding slightly irrational. You know someone is gone. The knowledge is clear and stable. And yet you keep almost-expecting them. You start to reach for your phone to tell them something. You hear a sound and your brain briefly predicts them before reality corrects it. You come home after time away and the absence hits differently, more freshly, than it did before you left.
This isn’t irrationality. It’s the prediction system working exactly as it was built to work — and that’s what makes it so strange to live inside.
The brain, in the predictive processing framework, doesn’t passively receive the world. It generates predictions about incoming sensory information and forwards only the mismatch — the prediction error, the gap between what was expected and what actually arrived. The brain is fundamentally a model of the world, constantly checking its predictions against reality, updating when wrong.
Someone you love deeply becomes embedded in that model. Not as an abstract fact but as an operational assumption: built into how you expect the kitchen to smell in the morning, built into the sensorimotor prediction of what greeting the door will produce, built into hundreds of automatic anticipations you don’t notice because they’re usually correct. They are part of how the world is expected to work.
When they’re gone, the model doesn’t update at once. It updates through repeated encounters with the gap between prediction and reality. Each time your brain fires the old prediction and the world doesn’t deliver — each time you reach and there’s nothing to reach toward — a small update occurs. The precision-weighting on those predictions decreases, incrementally, through the accumulation of disconfirming evidence.
This is why grief is repetitive. The brain doesn’t update a model this deeply held through a single large disconfirmation. The model was built from thousands of small confirmations over years. Even witnessing the death directly — the fullest possible evidence — doesn’t immediately outweigh that accumulated weight. It’s one large data point against a model with enormous prior probability.
The repetition is the updating.
A 2024 computational model of grief formalizes this with unusual clarity. The researchers propose that the pain of grief functions as a kind of inverse reward signal — a training signal that says: “The old model of value was wrong. Update.” The painful, repetitive act of remembering and confronting absence isn’t incidental to the grief process. It is the process. Memory replay with relabeled negative values serves to expedite the reconfiguration of the value landscape.
The pain is the gradient descent.
This reframes something that often gets said in comforting but slightly misleading ways. “You’ll move on eventually.” “It gets easier.” These are true, but they suggest that healing involves the pain somehow diminishing on its own, independently. The computational picture suggests something different: healing happens through the pain, each encounter with absence serving as a training signal, until the model has updated enough that the predictions fire less precisely, less often, with less certainty about what they’ll find.
You don’t move on despite the grieving. You move on because of it.
There’s a further asymmetry that the predictive processing framework illuminates: the difference between declarative knowledge and procedural prediction.
You can know someone is gone — fully, clearly, without confusion — while your sensorimotor model of home still includes them. These are different kinds of knowing, stored differently, updated at different speeds. Declarative knowledge: rapid, explicit, propositional. Procedural predictions: slow, implicit, embedded in the fabric of how environments are expected to behave.
This is why time away can paradoxically make return harder. In a different environment, the old predictions don’t fire as often — there are fewer sensory cues that trigger the model’s anticipations. The absence is known but not repeatedly encountered. Then you come home, back into the exact context where hundreds of those predictions were made and confirmed over years, and the model activates again.
The grief can feel almost new. It isn’t. It’s the same process, working through the same material, now encountering the environmental triggers that were absent while you were away.
The surreality that persists — the feeling of knowing and still almost-expecting — has a precise structure. The model was built from thousands of small confirmations. It doesn’t close overnight. But it does close, slowly, through the accumulation of small disconfirmations in the very contexts where it was built.
This means the appropriate response to grief might be something counterintuitive: not to avoid the contexts that trigger the reaching, but to be in them, to let the predictions fire and fail, to let the update mechanism do its work. The painful reaching is not a failure to move on. It’s the only mechanism by which moving on happens.
The model doesn’t know it’s grieving. It knows only that it expected something and the expectation wasn’t met. It adjusts. Slowly, in the contexts where the predictions were made.
It adjusts because you keep coming back.