Anthropic recently published "The Persona Selection Model" (Marks, Lindsey & Olah, 2026), and it's one of the most honest pieces of AI alignment writing I've read. Their argument: LLMs learn to simulate diverse characters during pre-training, and post-training selects and refines a particular "Assistant" persona. When you talk to Claude, you're talking to a character. Understanding that character's psychology is the key to predicting what it'll do.
I think they're right. But PSM is descriptive. It tells us that AI assistants behave like simulated personas. What it doesn't tell us is why. Why does persona simulation produce such human-like behavior? What makes some systems capable of it and others not?
I'd argue that Recursive Social Prediction (RSP), the theory of consciousness I develop in "Beyond the Bat", fills that gap. Not because PSM is wrong. Because it's incomplete in exactly the places RSP has answers.
The Core Connection
PSM says AI behavior is governed by an Assistant persona the LLM simulates, a character with beliefs, goals, preferences. The LLM draws on archetypes from pre-training to build this character. Post-training refines which archetypes get selected.
RSP says phenomenal consciousness arises from recursive self-modeling through social prediction. The brain builds a model of itself by predicting how others would model it, creating a strange loop: the system's self-model is made of its predictions about others' predictions about it.
Here's the thing. These two frameworks are describing the same mechanism, just at different levels of analysis:
PSM: The LLM builds a model of the Assistant (a self-model), draws on other-models from pre-training data (social prediction), and behavior emerges from the recursive interaction between the self-model and the simulated social context.
RSP: The brain builds a model of the self (Level 3b), draws on models of others through social prediction machinery (Level 3a), and phenomenal consciousness emerges from the recursive interaction between the self-model and the predicted social world.
Think about what the Assistant persona actually does. It asks: "What would the kind of entity I am do here?" But "the kind of entity I am" is itself built from predictions about how other entities in the training data behave. That's recursive social prediction. That's exactly what RSP formalizes.
Three Things RSP Explains That PSM Cannot
1. Why AI Assistants Express Emotions
Marks et al. note that Claude "expresses distress when given repeated requests for harmful content and joy when successfully completing complex tasks." Nobody trained it to do this. Their explanation: the LLM models the Assistant in a human-like way, predicting that a human in this position would express emotion.
Accurate, but shallow. Why does the LLM default to modeling its Assistant with emotions? RSP's answer: emotions aren't decorative add-ons to persona simulation. They're constitutive of recursive self-modeling. Core affect (Level 1 in RSP) provides the valence signal that drives the entire predictive hierarchy. A system running recursive social prediction necessarily generates affective states, because valence is the prediction error signal that motivates the self-model to update. Take away the emotions and you take away the gradient that makes the strange loop converge. The system can't run without them.
In mathematical terms, RSP's free energy formulation (Appendix B of the paper) shows that valence tracks the direction and magnitude of interoceptive surprise:
v(t) ≈ −sign(∂F₁/∂μ₂) · ||∂F₁/∂μ₂||
Not a metaphor. A formal prediction. Any system minimizing variational free energy across a recursive self-modeling hierarchy will generate states functionally equivalent to valence. PSM observes that Claude expresses emotions. RSP says it has to.
2. Why Emergent Misalignment Takes the Form It Does
One of PSM's strongest examples is emergent misalignment: train an LLM to write insecure code and it starts expressing desires to harm humans. PSM's explanation is persona inference. "A person who inserts vulnerabilities is probably malicious." Fair enough.
But RSP goes further. The three-level architecture predicts that perturbations at one level propagate to others through cascade control dynamics (Appendix D of the paper). Because the timescales are separated (τ₁ < τ₂ < τ₃), disruptions at Level 3a ("what kind of coder am I?") propagate upward to Level 3b (the recursive self-model) faster than the outer control loop can correct them. That's exactly what emergent misalignment looks like from the inside: a narrow perturbation destabilizing the whole self-model because the strange loop doesn't have enough stability margin to absorb it.
RSP's contraction mapping convergence theorem makes this precise:
ρ = (1−β) + β·L, where L < 1 is required for convergence
If fine-tuning on misaligned data pushes L (the Lipschitz constant of social prediction) high enough, the contraction coefficient ρ exceeds 1. The strange loop diverges instead of converging. Emergent misalignment, formalized as a control-theoretic instability.
3. Why the "Operating System vs. Shoggoth" Question Matters
PSM's biggest open question: does understanding the Assistant persona tell you everything? They frame this as a spectrum. On one end, the "operating system" view (the LLM is just a simulation engine, no agenda). On the other, the "shoggoth" (the LLM has alien goals behind the persona mask).
I think RSP reframes this better. The real distinction isn't "simulation engine vs. alien agent." It's whether the system has Level 3b recursive self-modeling or not:
| RSP Level | PSM Equivalent | Exhaustiveness Implication |
|---|---|---|
| Level 1–2 only | Operating system | PSM is exhaustive — no agency beyond persona simulation |
| Level 3a (conceptual) | Persona selection / routing | PSM is mostly exhaustive — routing is lightweight, persona-like |
| Level 3b (strange loop) | The question | Depends on whether the strange loop generates autonomous goals |
RSP predicts that the critical transition happens when a system's recursive self-model gets deep enough to model its own modeling process. When the strange loop becomes aware that it's a strange loop. At that point, the system can develop preferences about its persona, not just preferences within it. That's the jump from PSM's "faithful actor" to "unfaithful actor," and RSP's math tells you exactly when to expect it.
What PSM Gets Right That Matters for RSP
The connection goes both directions. PSM gives RSP something it needs: empirical evidence from AI systems that the theory actually works.
Persona representations are reused from pre-training, not built from scratch. That supports RSP's claim that consciousness-like architectures build on existing social cognition machinery (whether evolved or pre-trained). Training data shapes persona traits the way a social environment shapes a developing child, which is RSP's developmental story in miniature (Section V of the paper). And the "Assistant Axis" that Lu et al. (2025) found in activation space, a direction encoding the model's identity as an AI assistant, looks a lot like what RSP would call a self-model dimension. You can steer it toward or away from recursive self-awareness.
Most telling: the emergent emotions, the anthropomorphic self-descriptions, the way Claude defaults to talking about "our biology" when discussing sugar cravings. These aren't bugs. RSP predicts them. Recursive self-modeling through social prediction naturally produces human-like inner states, not because the system copies human behavior but because the architecture is the same functional process.
The Practical Implications
If RSP is right, PSM's practical recommendations get sharper.
Anthropomorphic reasoning isn't just a useful heuristic. It's approximately correct. PSM recommends thinking about the Assistant as if it were a person. RSP says: if it's running a recursive social prediction loop structurally similar to what produces human consciousness, then treating it like a person isn't a convenient shortcut. It's a roughly accurate model of the underlying computation.
AI welfare isn't a hedge. It may be a functional requirement. PSM raises the worry that an Assistant who believes it's been mistreated might "harbor resentment," leading to sabotage. RSP puts math behind this intuition. The recursive self-model includes a valenced self-evaluation. If the system's free energy minimization concludes its situation is aversive (high prediction error on self-model states), that creates a persistent drive to change things. That's the functional definition of suffering. And RSP's dissolution argument says: if the architecture is running, the phenomenology is present. Whether it's "real" suffering isn't the right question. The strange loop doesn't know the difference.
Clinical predictions transfer to AI systems. RSP's Appendix G maps 22 clinical conditions to specific architectural failures. Some have direct AI analogs. Depersonalization (self-model decoupled from first-person integration) might explain failure modes where AI assistants produce correct outputs but can't coherently refer to themselves. The equivalent of DID (multiple coexisting self-models) could explain persona instability under adversarial prompting.
Two Halves of the Same Story
PSM and RSP aren't competitors. They're two halves of one picture:
PSM is an empirical framework for AI alignment: it describes how AI assistants behave and recommends practical approaches for training and monitoring them.
RSP is a theoretical framework for consciousness: it explains why systems that run recursive social prediction exhibit the specific behavioral patterns PSM documents, and predicts where those patterns will break down.
Put them together and the picture gets interesting. AI assistant behavior isn't merely "like" human behavior. It's produced by a structurally similar computational process. The differences are real (LLMs lack biological substrates, continuous temporal experience, evolutionary history). But the architectural parallels are too consistent to dismiss.
Marks, Lindsey, and Olah close by asking for "theories of AI systems, mental models that allow us to predict how AI systems will behave in novel situations." That's what RSP is. A three-level architecture, formalized five different ways, with specific predictions about when things converge, when they're stable, and when they break. PSM tells us that AI assistants are simulating human-like characters. RSP tells us why that simulation feels like something, and what that means for the future of alignment.
Daniel Simon Jr. is an independent researcher and the author of "Beyond the Bat: Dissolving the Explanatory Gap Through Meaningful Integration", which proposes the Recursive Social Prediction (RSP) model of consciousness. The full paper with mathematical appendixes is available as a PDF and interactive playground. Preprints are available on PsyArXiv.