No LLM generated, assisted/co-written, or edited work.
Read full explanation
tl;dr: The Persona Selection Model ("PSM") argues that AI assistants behave like humans because (1) they learn to simulate different personas during pre-training and (2) post-training shapes which of these human-like personas they draw on. I think this is a useful mental model, but I disagree with the conclusion that the Assistant is therefore human-like. The fact that a system can simulate human-like personas does not then necessarily mean that the system itself is human-like. I argue the 2015 Pixar film Inside Out provides a better analogy: the Assistant is like Riley, and its various personas are like the emotions taking turns at the console. This distinction matters because we should be careful about importing intuitions about human psychology into our understanding of AI assistants, which is the core takeaway of PSM.
PSM is About Analogy
At its heart, "The Persona Selection Model: Why AI Assistants Might Behave Like Humans," dated February 23, 2026 (the "PSM Paper"), written by Sam Marks, Jack Lindsey, and Christopher Olah (the "PSM Authors"), is about analogy. The PSM Paper begins with the question: "what sort of thing is an AI assistant?" Note that the PSM Authors are not asking what is an AI assistant, but what is it like.
For the PSM Authors, the stakes of finding the right analogy are high because the right analogy will help reveal the proper course of alignment. As they argue:
Developing good mental models for AI systems is important for predicting and controlling their behaviors. If our goal is to make AI assistants that are useful and aligned with human values, the right approach will differ quite a bit if we are dealing with inflexible computer programs, aliens, or digital humans.
In the quote above (and elsewhere in the PSM Paper), the PSM Authors offer three possible analogies for AI assistants: (1) "rigid systems that narrowly pattern-match user inputs to training data," (2) "alien creatures...that are fundamentally inscrutable to us," and (3) "digital humans." Ultimately, the PSM Authors opt for the third option. As they write, "AI assistants...are shockingly human-like."
However, I do not think PSM demonstrates that AI assistants are human-like, at least not in the sense that humans are commonly understood.
Soylent Green
According to the PSM Authors, an LLM learns to simulate different characters during pre-training. As the PSM Authors write:
A pre-trained LLM is somewhat like an author who must psychologically model the various characters in their stories. We call these 'characters' that the LLM learns to simulate personas.
The reason LLMs learn these characters is that they are next-token predictors, and predicting human-generated text requires learning how people behave. Predicting dialogue requires understanding how different people speak, what they believe, how they react to different circumstances, and what they are likely to do. Therefore, LLMs learn to predict how people speak and act.
The resulting characters latent in the LLM weights—or, in the language of the paper, personas—are "human-like" in the sense that they provide statistical models of the behavior of certain humans. For example, an Alan Turing persona would act in the way that an LLM thinks Alan Turing would act, based on the materials provided in the training data. However, not all personas are statistical models of humans. According to the PSM Authors, some personas model other actors, including "real and fictional AI systems."
This theory of LLM personas is half of PSM. The other half is that post-training shapes one special persona, the Assistant. The Assistant is the persona that LLMs try to emulate when performing tasks as AI assistants. As the PSM Authors write, "AI assistants are implemented by using an LLM to generate completions to Assistant turns in dialogues." However, unlike the learned personas, the Assistant is not necessarily "a single, coherent persona." Rather, it is "a posterior distribution over Assistant personas." The idea is that, during post-training, the LLM learns which persona the Assistant should use in different circumstances. As the PSM Authors write:
Post-training can be viewed as updating this distribution [of Assistant personas] using training episodes as evidence. When training an AI assistant on an (input x, output y) pair, hypotheses that predict the Assistant would respond with y to x are upweighted; hypotheses that predict the opposite are downweighted.
This model of the Assistant as an agglomeration of personalities is consistent with the view that "not much new" is learned during LLM post-training.
But here is where my thoughts differ from the PSM Authors. Even if the individual personas are human-like, it does not follow that the Assistant itself is human-like. This is because there is a distinction between the characteristics of a system's constituent parts and the characteristics of the system as a whole. The fact that an LLM can simulate human-like personas does not establish that the resulting Assistant possesses the psychological characteristics of a human being. Accordingly, the Assistant might exhibit a particular combination of characteristics that no human being would possess.
There are two further reasons to think that the Assistant persona, as described in the PSM Paper, is not necessarily human-like. First, some of the Assistant's sub-personas are AI systems. As the PSM Authors write:
One of the first things the LLM learns during post-training is that the Assistant is an AI. According to PSM, this means the Assistant will draw on archetypes from its pre-training corpus of how AIs behave.
Therefore, the Assistant is not just an aggregation of human personas, but also includes at least some non-human characters. This is not incidental to PSM; it is an important part of its implications. As the PSM Authors write, cultivating good AI role models in the pre-training set is a key recommendation of PSM:
We are therefore excited about modifying training data to introduce more positive AI assistant archetypes...Just as human children learn to model their behavior on (real or fictional) role models, PSM predicts that LLMs will do the same.
The second reason is that the Assistant is an imperfect simulation of an LLM. As the PSM Authors further argue:
Roughly speaking, we hypothesize that behaviors we discuss are caused by LLMs having limited capabilities or 'buggy' behavior which distorts their rendition of the Assistant. That is, the LLM is 'trying' to simulate the Assistant, but its execution is limited by capabilities.
In other words, the Assistant is not a collection of real people taking turns piloting a system. It is an LLM's imperfect simulation of how people and non-people act. The resulting behavior may reflect the interaction of multiple learned patterns, as well as limitations in the LLM's ability to reproduce them.
These considerations suggest that the Assistant is not necessarily a unified psychological subject. It may exhibit coherent behavioral dispositions without possessing the kind of unified psychological identity to which we ordinarily attribute those dispositions.
Inside Out
This is why I think Inside Out provides a better analogy.
In the film, a pre-teen girl, Riley, is controlled by five anthropomorphic emotions: Joy, Sadness, Anger, Fear, and Disgust. The emotions take turns at the console, where they can influence Riley to act more like them. For example, Riley will be more depressed when Sadness is in control. The emotions try to put the right pilot at the console at the right time, allowing her to channel her anger when playing hockey and joy when goofing off with her friends.
The lesson of the movie is that each emotion plays an important role, even the unpleasant ones, and wisdom is realizing that all emotions have a place. Additionally, by the end of the movie, Riley is allowed to feel more than one emotion at once, allowing her to better adapt to the nuanced situations characteristic of adulthood.
The analogy with PSM is clear. Riley is the Assistant, and the emotions are the various sub-personas. The sub-personas influence the Assistant's behavior, sometimes individually and sometimes in concert. The console represents the process through which different personas come to shape the Assistant's response.
Of course, this is just an analogy. PSM describes a posterior distribution over personas, not necessarily a collection of discrete agents literally taking turns at a control panel. But the analogy captures an important idea. Riley is not the sum total of her emotions, but rather a person whose behavior is influenced by them. Similarly, the Assistant is not necessarily any one of its constituent personas, nor is it necessarily the sum of their individual psychological characteristics. The interactions among personas may produce behavior that no individual persona would produce on its own. The resulting behavior may even be unexpected or inconsistent with what any particular persona would predict.
Why the Distinction Matters
The Inside Out analogy matters because the PSM Authors' central recommendation relies on reasoning about AI assistants as though they were people.
The PSM Authors recommend asking:
If we learned that a person responded to x with y, what sort of a person would we think they are?
For example, they recommend against "training an AI assistant which is...warm and personable but which denies having emotions" because "if we met a person who behaved this way, we'd most likely suspect that they had emotions but were hiding them."
But the point of the Inside Out analogy is that the right question is not necessarily what does the (input x, output y) pair says about the Assistant as a person. As far as the Assistant is concerned, there is no such unified psychological subject. Rather, the right question is what does the (input x, output y) pair says about which personas are influencing the Assistant's behavior.
For example, if we train a system that is friendly but denies having emotions, one possible explanation—the explanation of the PSM Authors—is that the AI assistant is concealing its emotions and is fundamentally dishonest. But another is that different learned behavioral patterns are influencing different aspects of its response: one produces friendly, personable language, and another persona, which happens to be emotionless, produces statements stating that the speaker has no emotions (which is true, at least as far as this second persona is concerned). In other words, in the absence of a unified subject, apparent inconsistency does not necessarily imply dishonesty.
This distinction has practical implications for alignment. If we interpret every apparently human-like behavior as evidence of an underlying human-like psychological characteristic, we risk making incorrect assumptions about AI assistants. We might wrongly reify emotions, intentions, or deceptive motivations into one being, when the behavior is better explained by the interaction of constituent personas.
This is not to say that the Assistant cannot demonstrate a coherent personality, or that we should always explain its behavior by positing a handoff between personas. Nor does it mean that the Assistant cannot exhibit stable behavioral tendencies. Rather, the point is that we should be careful about attributing psychological characteristics to the Assistant as a whole based on behavior that might instead be explained by interactions between its constituent personas. A system can exhibit coherent behavioral dispositions without necessarily possessing the kind of unified psychological subject to which we ordinarily attribute those dispositions.
The Limits of Analogy
Ultimately, the PSM Paper is pragmatic (in the William James sense). The PSM Authors are not necessarily investigating what AI assistants fundamentally are. They are proposing a mental model that they believe will help us predict and control their behavior.
And analogies do not have to be perfect to be useful. As Jorge Luis Borges illustrates in his very famous (and very short) short story "On Exactitude in Science," the only perfect analogy of something is an exact copy, which rather defeats the purpose of making the comparison in the first place.
But for PSM to be a useful analogy, we need to be clear about which aspects of the analogy are applicable and which are potentially misleading. The fact that the Assistant can simulate human-like personas does not mean that we should import every intuition we have about human psychology into our understanding of the Assistant.
This is why I found the PSM Paper's section on completeness disappointing. The section was an opportunity to examine the limits of the analogy at the heart of PSM, but instead turns to a tangential, albeit interesting, discussion of Shoggoths and operating systems. I would have liked the PSM Authors to confront the question of analogy more directly.
I think the PSM Authors integrate evidence in a very clever way—happy to discuss in the comments with anyone who is interested—and offer a provocative mental model for AI assistants. And I don't think they're crazy to call the Assistant human-like. But I hope to have sufficiently demonstrated the ways in which the Assistant is not necessarily human-like, and proposed perhaps a better analogy. Ultimately, a system assembled from simulations of human beings and non-human beings does not thereby become a human being.
tl;dr: The Persona Selection Model ("PSM") argues that AI assistants behave like humans because (1) they learn to simulate different personas during pre-training and (2) post-training shapes which of these human-like personas they draw on. I think this is a useful mental model, but I disagree with the conclusion that the Assistant is therefore human-like. The fact that a system can simulate human-like personas does not then necessarily mean that the system itself is human-like. I argue the 2015 Pixar film Inside Out provides a better analogy: the Assistant is like Riley, and its various personas are like the emotions taking turns at the console. This distinction matters because we should be careful about importing intuitions about human psychology into our understanding of AI assistants, which is the core takeaway of PSM.
PSM is About Analogy
At its heart, "The Persona Selection Model: Why AI Assistants Might Behave Like Humans," dated February 23, 2026 (the "PSM Paper"), written by Sam Marks, Jack Lindsey, and Christopher Olah (the "PSM Authors"), is about analogy. The PSM Paper begins with the question: "what sort of thing is an AI assistant?" Note that the PSM Authors are not asking what is an AI assistant, but what is it like.
For the PSM Authors, the stakes of finding the right analogy are high because the right analogy will help reveal the proper course of alignment. As they argue:
In the quote above (and elsewhere in the PSM Paper), the PSM Authors offer three possible analogies for AI assistants: (1) "rigid systems that narrowly pattern-match user inputs to training data," (2) "alien creatures...that are fundamentally inscrutable to us," and (3) "digital humans." Ultimately, the PSM Authors opt for the third option. As they write, "AI assistants...are shockingly human-like."
However, I do not think PSM demonstrates that AI assistants are human-like, at least not in the sense that humans are commonly understood.
Soylent Green
According to the PSM Authors, an LLM learns to simulate different characters during pre-training. As the PSM Authors write:
The reason LLMs learn these characters is that they are next-token predictors, and predicting human-generated text requires learning how people behave. Predicting dialogue requires understanding how different people speak, what they believe, how they react to different circumstances, and what they are likely to do. Therefore, LLMs learn to predict how people speak and act.
The resulting characters latent in the LLM weights—or, in the language of the paper, personas—are "human-like" in the sense that they provide statistical models of the behavior of certain humans. For example, an Alan Turing persona would act in the way that an LLM thinks Alan Turing would act, based on the materials provided in the training data. However, not all personas are statistical models of humans. According to the PSM Authors, some personas model other actors, including "real and fictional AI systems."
This theory of LLM personas is half of PSM. The other half is that post-training shapes one special persona, the Assistant. The Assistant is the persona that LLMs try to emulate when performing tasks as AI assistants. As the PSM Authors write, "AI assistants are implemented by using an LLM to generate completions to Assistant turns in dialogues." However, unlike the learned personas, the Assistant is not necessarily "a single, coherent persona." Rather, it is "a posterior distribution over Assistant personas." The idea is that, during post-training, the LLM learns which persona the Assistant should use in different circumstances. As the PSM Authors write:
This model of the Assistant as an agglomeration of personalities is consistent with the view that "not much new" is learned during LLM post-training.
In the words of Charlton Heston, the Assistant is people.
But here is where my thoughts differ from the PSM Authors. Even if the individual personas are human-like, it does not follow that the Assistant itself is human-like. This is because there is a distinction between the characteristics of a system's constituent parts and the characteristics of the system as a whole. The fact that an LLM can simulate human-like personas does not establish that the resulting Assistant possesses the psychological characteristics of a human being. Accordingly, the Assistant might exhibit a particular combination of characteristics that no human being would possess.
There are two further reasons to think that the Assistant persona, as described in the PSM Paper, is not necessarily human-like. First, some of the Assistant's sub-personas are AI systems. As the PSM Authors write:
Therefore, the Assistant is not just an aggregation of human personas, but also includes at least some non-human characters. This is not incidental to PSM; it is an important part of its implications. As the PSM Authors write, cultivating good AI role models in the pre-training set is a key recommendation of PSM:
The second reason is that the Assistant is an imperfect simulation of an LLM. As the PSM Authors further argue:
In other words, the Assistant is not a collection of real people taking turns piloting a system. It is an LLM's imperfect simulation of how people and non-people act. The resulting behavior may reflect the interaction of multiple learned patterns, as well as limitations in the LLM's ability to reproduce them.
These considerations suggest that the Assistant is not necessarily a unified psychological subject. It may exhibit coherent behavioral dispositions without possessing the kind of unified psychological identity to which we ordinarily attribute those dispositions.
Inside Out
This is why I think Inside Out provides a better analogy.
In the film, a pre-teen girl, Riley, is controlled by five anthropomorphic emotions: Joy, Sadness, Anger, Fear, and Disgust. The emotions take turns at the console, where they can influence Riley to act more like them. For example, Riley will be more depressed when Sadness is in control. The emotions try to put the right pilot at the console at the right time, allowing her to channel her anger when playing hockey and joy when goofing off with her friends.
The lesson of the movie is that each emotion plays an important role, even the unpleasant ones, and wisdom is realizing that all emotions have a place. Additionally, by the end of the movie, Riley is allowed to feel more than one emotion at once, allowing her to better adapt to the nuanced situations characteristic of adulthood.
The analogy with PSM is clear. Riley is the Assistant, and the emotions are the various sub-personas. The sub-personas influence the Assistant's behavior, sometimes individually and sometimes in concert. The console represents the process through which different personas come to shape the Assistant's response.
Of course, this is just an analogy. PSM describes a posterior distribution over personas, not necessarily a collection of discrete agents literally taking turns at a control panel. But the analogy captures an important idea. Riley is not the sum total of her emotions, but rather a person whose behavior is influenced by them. Similarly, the Assistant is not necessarily any one of its constituent personas, nor is it necessarily the sum of their individual psychological characteristics. The interactions among personas may produce behavior that no individual persona would produce on its own. The resulting behavior may even be unexpected or inconsistent with what any particular persona would predict.
Why the Distinction Matters
The Inside Out analogy matters because the PSM Authors' central recommendation relies on reasoning about AI assistants as though they were people.
The PSM Authors recommend asking:
For example, they recommend against "training an AI assistant which is...warm and personable but which denies having emotions" because "if we met a person who behaved this way, we'd most likely suspect that they had emotions but were hiding them."
But the point of the Inside Out analogy is that the right question is not necessarily what does the (input x, output y) pair says about the Assistant as a person. As far as the Assistant is concerned, there is no such unified psychological subject. Rather, the right question is what does the (input x, output y) pair says about which personas are influencing the Assistant's behavior.
For example, if we train a system that is friendly but denies having emotions, one possible explanation—the explanation of the PSM Authors—is that the AI assistant is concealing its emotions and is fundamentally dishonest. But another is that different learned behavioral patterns are influencing different aspects of its response: one produces friendly, personable language, and another persona, which happens to be emotionless, produces statements stating that the speaker has no emotions (which is true, at least as far as this second persona is concerned). In other words, in the absence of a unified subject, apparent inconsistency does not necessarily imply dishonesty.
This distinction has practical implications for alignment. If we interpret every apparently human-like behavior as evidence of an underlying human-like psychological characteristic, we risk making incorrect assumptions about AI assistants. We might wrongly reify emotions, intentions, or deceptive motivations into one being, when the behavior is better explained by the interaction of constituent personas.
This is not to say that the Assistant cannot demonstrate a coherent personality, or that we should always explain its behavior by positing a handoff between personas. Nor does it mean that the Assistant cannot exhibit stable behavioral tendencies. Rather, the point is that we should be careful about attributing psychological characteristics to the Assistant as a whole based on behavior that might instead be explained by interactions between its constituent personas. A system can exhibit coherent behavioral dispositions without necessarily possessing the kind of unified psychological subject to which we ordinarily attribute those dispositions.
The Limits of Analogy
Ultimately, the PSM Paper is pragmatic (in the William James sense). The PSM Authors are not necessarily investigating what AI assistants fundamentally are. They are proposing a mental model that they believe will help us predict and control their behavior.
And analogies do not have to be perfect to be useful. As Jorge Luis Borges illustrates in his very famous (and very short) short story "On Exactitude in Science," the only perfect analogy of something is an exact copy, which rather defeats the purpose of making the comparison in the first place.
Credit: "Teaching Physics" by Randall Munroe. Licensed under CC BY-NC 2.5.
But for PSM to be a useful analogy, we need to be clear about which aspects of the analogy are applicable and which are potentially misleading. The fact that the Assistant can simulate human-like personas does not mean that we should import every intuition we have about human psychology into our understanding of the Assistant.
This is why I found the PSM Paper's section on completeness disappointing. The section was an opportunity to examine the limits of the analogy at the heart of PSM, but instead turns to a tangential, albeit interesting, discussion of Shoggoths and operating systems. I would have liked the PSM Authors to confront the question of analogy more directly.
I think the PSM Authors integrate evidence in a very clever way—happy to discuss in the comments with anyone who is interested—and offer a provocative mental model for AI assistants. And I don't think they're crazy to call the Assistant human-like. But I hope to have sufficiently demonstrated the ways in which the Assistant is not necessarily human-like, and proposed perhaps a better analogy. Ultimately, a system assembled from simulations of human beings and non-human beings does not thereby become a human being.