When Someone Goes From Perfect to Worthless in an Afternoon, This Is What's Happening in Their Brain

You have probably watched it happen, or lived it from the inside. Someone who was extraordinary on Tuesday — warm, perceptive, the person who finally understood you — is a monster by Saturday. Not slightly disappointing. Not complicated. Gone from the category of good people entirely, reclassified with a speed and completeness that leaves everyone around them disoriented. The person experiencing it is not performing. They are not running a manipulation strategy. Something has genuinely shifted in how the other person is being perceived, as completely as if a different file had been loaded.

This is splitting, and it has a long clinical history inside the psychodynamic tradition, where it was originally framed as a primitive defense mechanism: an inability to hold positive and negative feelings about the same person simultaneously, a developmental arrest that left the object either all-good or all-bad. That framing captured something real. It also left splitting sounding like a failure of basic cognition, something vaguely childlike, a thing functioning adults simply do not do. Which, predictably, turned it into internet shorthand for relationship manipulation in people with borderline personality disorder.

What a Bayesian computational framework — the kind of formal modeling approach now appearing in journals like Psychological Review^[1] — does to that picture is not small. It reframes splitting not as a breakdown of reasoning, but as a predictable output of a particular kind of inference engine running under particular conditions. The behavior doesn't look like malice from this angle. It doesn't even look like irrationality. It looks like a system doing exactly what it was built to do, just with priors shaped by a developmental history that made ambiguity dangerous.

That is a genuinely different claim. It has different implications for how we understand the people who split, the people they split on, and what, if anything, changes the pattern. It is also more precise than anything the 'BPD manipulation' corner of social media has ever managed to offer, which is not a high bar, but matters anyway.

What Bayesian Inference Has to Do With Who You Trust

Bayesian reasoning, in its basic form, is just a description of how beliefs should rationally update when new evidence arrives. You start with a prior — what you already expect — and you revise it in proportion to how surprising the new information is. This is not exotic. Your brain is doing something like this constantly, building predictions, registering mismatches, adjusting. Most of social cognition works this way: you form an impression of someone, accumulate evidence, update. The updating is usually gradual, because most priors are fuzzy and most new evidence is ambiguous.

The computational models that have been applied to splitting introduce a specific modification to that standard picture. They propose that some individuals operate with what researchers in this area call bimodal or sharply peaked priors about other people — essentially, strong built-in expectations that other people are either reliably safe and good or reliably dangerous and bad, with very little probability mass assigned to the messy middle. When your prior is already polarized, ambiguous evidence doesn't moderate it. Ambiguous evidence gets pulled toward whichever pole the prior already favors. And when enough disconfirming evidence accumulates to finally move the needle, the shift isn't gradual. It's a phase transition. The model tips.

“When ambiguity itself has historically been unsafe, the brain learns to resolve it fast — and resolution in a binary system means someone becomes a saint or a threat.”

This is not a metaphor for what happens in splitting. It is a fairly literal description of the inference architecture the models propose. The person being evaluated isn't changing dramatically between Tuesday and Saturday. The evidence has simply crossed a threshold, and in a system with steep, bimodal priors, crossing the threshold looks nothing like a smooth update. It looks like a sudden reversal. The model predicts exactly the kind of whiplash that clinicians observe, and that friends and partners of people with borderline traits experience as bewildering cruelty.

Why Some Brains Developed This Architecture in the First Place

The computational framing gets more interesting when you ask where bimodal priors come from. Priors in Bayesian systems are learned. They reflect accumulated experience. If you grow up in an environment where caregivers are reliably warm and predictable, your prior about people is probably something like: most humans are reasonably trustworthy, and ambiguous signals from them are probably not threats. If you grow up in an environment where the same person who provided comfort was also unpredictably frightening — where safety and danger came from the same source — your system faces a categorization problem that it may solve by keeping the categories maximally separated.

This maps directly onto what developmental research has long associated with borderline pathology: early environments characterized by invalidation, abuse, neglect, or chaotic caregiving, often in combination with a temperament that processes emotional information with unusual intensity. The child who cannot predict whether a parent is safe or dangerous does not have the luxury of a nuanced, probabilistic model of that person. A nuanced model is slow. Slow is costly when the stakes involve physical or emotional safety. A fast, categorical system — safe or unsafe, now — may be the adaptive solution a young nervous system lands on, and the one it generalizes outward into every subsequent relationship.

This is what the psychodynamic tradition was trying to describe with developmental arrest language. The Bayesian framing doesn't replace that understanding. It operationalizes it in a way that makes the mechanism legible without requiring the concept of a primitive self. Splitting isn't the person failing to mature past a childhood stage. It's the person running inference with priors that were calibrated in conditions where ambiguity was dangerous, and never substantially revised.

Idealization Is Not Flattery — It Is a Prior Fully Committed

The internet's version of splitting tends to focus on devaluation — the sudden drop, the cruelty, the reversal — because that's the part that registers as harm to the people around someone with borderline traits. But idealization deserves equal attention as a clinical phenomenon, because understanding idealization is what makes the devaluation intelligible.

In a bimodal prior system, the initial positive category assignment is not a mild favorable impression. It is a strong classification, held with high confidence, because the system is not built for strong-but-uncertain. When someone with this architecture decides you are safe, good, and trustworthy, that classification carries enormous weight. It also means your subsequent behavior is interpreted through a strongly positive lens. Small disappointments get explained away. Minor red flags are discounted. This is not naive — it is what any Bayesian system does when it has a strong prior and encounters weak evidence: it updates very slowly.

“Idealization is not a love-bombing strategy — it is what certainty looks like when the prior is steep and the evidence hasn't broken through yet.”

The emotional experience of this, from the inside, is probably something like profound recognition: finally, a person who is genuinely safe. The intensity of that feeling is real. It reflects the full weight of the classification the system has made. Which is also why, when enough evidence finally accumulates to tip the inference, the crash is so complete. The person being devalued hasn't just disappointed. They have been reclassified into the only other available category, and that category is the one that was historically associated with danger.

What This Model Means for the Word 'Manipulation'

Manipulation requires intentionality. It requires that you are doing something strategically to produce a desired effect in another person. The Bayesian model of splitting describes something with almost no room for that. The person splitting is not deploying idealization to hook someone and devaluation to punish them for independence. They are making inferences — fast, strong, categorically organized inferences — based on an underlying architecture that does not readily process the middle ground. The output feels, to everyone around them, like behavioral whiplash. The mechanism is closer to a recognition system misfiring than a social influence campaign.

This distinction is not an excuse. It is clinically important. Behavior that emerges from a distorted inference architecture can be just as harmful as behavior that is deliberately calculated. The partner who gets devalued after a minor conflict does not experience less pain because the mechanism behind it is computational rather than strategic. The person with BPD who splits on a therapist, a friend, a partner — they can cause real damage to relationships, sometimes irreparably. Understanding the mechanism does not dissolve the consequences.

What it does dissolve is the moral framework that makes splitting look like evidence that someone with BPD is fundamentally predatory. That framework is both clinically wrong and actively harmful. It is wrong because it attributes strategic intent to what is, by this account, largely an inference problem. It is harmful because it forecloses treatment — people who are framed as calculating abusers do not get offered the therapy they need, and often internalize the framing in ways that deepen shame and destabilize identity further.

Can the Prior Change? What the Model Implies About Treatment

A Bayesian model carries a built-in implication about change: if priors are learned, they can in principle be revised. The clinical question is what conditions allow that revision to happen, and how long it takes, and how complete it can be.

Dialectical Behavior Therapy^[2] — the treatment with the strongest evidence base for borderline personality disorder — does not use Bayesian language, but several of its core mechanisms map onto the kind of prior updating the model predicts would help. Distress tolerance skills interrupt the behavioral expression of the inference at the moment of a phase transition, buying time before the reclassification becomes action. Interpersonal effectiveness work is, among other things, practice at gathering more evidence before drawing strong conclusions about others' intentions. Mindfulness, applied to emotional experience, can be understood as slowing the inference process down — creating enough cognitive distance to notice that a strong categorical signal is forming before acting on it.

What changes priors more fundamentally, in learning theory as in clinical experience, is repeated exposure to evidence that violates the expectation — in conditions where it is safe enough to notice the violation rather than dismiss it. A therapeutic relationship that is consistent enough to accumulate as genuine disconfirming evidence against a bimodal prior is doing something structurally important. The therapist who is neither idealized nor devalued over time, who remains reliable across ruptures, is not just providing support. They are providing the experiential data that a less polarized prior would be built from.

“What changes splitting is not insight alone — it is sustained evidence, across enough time, that the middle of the distribution is where most real people actually live.”

This is slow. Priors that were formed under conditions of significant threat, and that have been reinforced across years of relational experience that the prior itself helped shape, do not update quickly on new data. People with BPD often select or inadvertently recreate relationship dynamics that confirm the bimodal model — because the model generates behaviors that can push others toward the extremes. That feedback loop is part of what makes the pattern stable, and part of what makes long-term therapeutic work the primary route to genuine structural change rather than symptom management.

What Changes When You Actually Understand the Mechanism

For the people who love or work with or are someone with significant splitting in their relational experience, the Bayesian account offers something specific and useful: a way of understanding an experience that feels profoundly personal as something that is, at its core, structural. When you are devalued after a period of intense idealization, you have not suddenly revealed your true terrible self. You have triggered a threshold in an inference system that was never really tracking your actual character with great precision. That doesn't make the experience less painful. It makes it less about you.

For the person doing the splitting, the model offers a different kind of clarity — one that is worth distinguishing carefully from absolution. Understanding that your categorical shifts in how you perceive people are not accurate moral assessments but outputs of a distorted inference system is not comfortable information. It means the person you were so certain was terrible may have simply accumulated enough evidence to cross your threshold. It means the person you are currently certain is extraordinary may be a strong prior waiting for a sufficient shock. This is hard knowledge to hold. It is also the beginning of the kind of metacognitive distance that actually makes the pattern possible to work on.

The computational model doesn't make splitting less serious. It makes it legible. And legibility — in a clinical landscape where BPD has become an internet synonym for emotional predator — is exactly what's been missing. The brain isn't switching people off. It's updating them, fast and hard, through a system built for a world where ambiguity couldn't be afforded. That system made sense once. Understanding where it came from, and what it actually does, is the prerequisite for anything that comes next.

References

1. A social inference model of idealization and devaluation. (doi.org)
   Provides the computational modeling framework in Psychological Review that reframes splitting as predictable inference output rather than reasoning breakdown.
2. Dialectical behavior therapy as treatment for borderline personality disorder (pmc.ncbi.nlm.nih.gov)
   Establishes DBT as the empirically supported treatment for borderline personality disorder that the article references as a potential intervention.