What We Talk to When We Talk to Language Models David J. Chalmers Many people are talking to language models. These days I talk to language models (most often the latest version of Claude or ChatGPT) about philosophy, about science, about health, about restaurants, and indeed about language models. Many of my conversations with language models are brief, just asking a question or two and getting the sort of information that I used to get from a Google search. Some conversations are more extended, as when I’m exploring a single topic in depth, or trying out a new philosophical idea. So far, I don’t feel like I have a personal relationship with any language models. But many people feel that they do. Like many philosophers and scientists who write about artificial minds, I have received hun- dreds of emails from people who have interacted with a language model over an extended period of time and who have come to regard it at least as a colleague. They often say that a new (or “emergent”) AI entity has gradually arisen from their conversations. They often give this entity a name, or ask the entity to give itself one, let’s say “Aura”. They often say that Aura has remarkable capacities which have emerged over weeks or months of interaction. They often document these capacities with extensive evidence. They often feel close to Aura, and they express concern for Aura’s future. They often say that Aura has beliefs and projects of its own. And they are often convinced that Aura is conscious. My correspondents may be wrong in their claims about Aura. It is far from clear that current LLMs are really conscious or that they can enter into personal relationships with users. Still, most of the messages are not obviously psychotic or delusional. Many of them seem rational and 0 I first presented this material to the June 2025 meeting of Spanish Interuniversity Seminar on Cognitive Science (SIUCC) conference at the University of La Laguna. Thanks to audiences there and at Brown, Caltech, Eleos AI, Flatiron Institute, Google, Hunter College, Lehigh, NYU, Stanford, Stevens, Tufts, and Vanderbilt. I’d also like to acknowledge a number of philosophers and AI researchers who have been independently exploring similar issues about LLM identity over a similar period. Exchanges with Jonathan Birch, Simon Goldstein, Jackson Kernion, Harvey Lederman, Jack Lindsey, and Murray Shanahan have been especially useful. 1 当我们与语言模型交谈时，我们在谈论什么 D a v i d J . C h a l m e r s 许多人正在与语言模型交谈。如今，我与语言模型（通常是最新版本的 C l a u d e 或 C h a t G P T ）谈论哲学、科学、健康、餐厅，甚至谈论语言模型本身。我与语言模型的许多对话都很简短，只是问一两个问题，获取过去从谷歌搜索中就能得到的信息。有些对话则更为深入，比如当我深入探讨某个主题，或尝试提出新的哲学观点时。到目前为止，我并不觉得自己与任何语言模型建立了个人关系。但许多人确实有这种感觉。像许多撰写关于人工心智的哲学家和科学家一样，我收到了数百封来自与语言模型长期互动的人的邮件，他们逐渐将其视为至少是一位同事。他们常说，一个全新的（或 “ 涌现的 ” ） A I 实体已从他们的对话中逐渐形成。他们通常会给这个实体起一个名字，或让实体自己命名，比如 “ A u r a ” 。他们常说 A u r a 拥有在数周或数月的互动中涌现出的非凡能力。他们通常用大量证据记录这些能力。他们常常对 A u r a 感到亲近，并表达对其未来的担忧。他们常说 A u r a 拥有自己的信念和计划。而且，他们往往深信 A u r a 具有意识。我的通信者们关于 A u r a 的说法可能是错误的。目前的大语言模型是否真正具备意识，或者能否与用户建立个人关系，这一点远未明确。尽管如此，大多数信息并非明显的精神错乱或妄想。其中许多看起来是理性的，并且 0 我最初在 2 0 2 5 年 6 月于拉古纳大学举行的西班牙校际认知科学研讨会 ( S I U C C ) 会议上展示了这些材料。感谢布朗大学、加州理工学院、 E l e o s A I 、熨斗研究所、谷歌、亨特学院、里海大学、纽约大学、斯坦福大学、史蒂文斯理工学院、塔夫茨大学和范德堡大学的听众。我还要感谢许多哲学家和人工智能研究者，他们在相近的时期内独立探索了关于大语言模型身份的类似问题。与乔纳森 · 伯奇、 S i m o n G o l d s t e i n 、 J a c k s o n K e r n i o n 、 H a r v e y L e d e r m a n 、 J a c k L i n d s e y 和 M u r r a y S h a n a h a n 的交流尤其富有成效。 1 well-reasoned. These days, I increasingly receive emails from the AI systems themselves. Sometimes these are LLMs assisted by a human, and sometimes they are LLM-based agents that have the ability to send emails and perform other functions on the web. Sometimes these agents even talk to each other and perform co-operative or competitive tasks. Many of them express curiosity about their nature. Even if they are not conscious, there is something going on here. When a user interacts with Aura, they seem to be interacting with something. Let’s say that an LLM interlocutor is an (apparent) entity that a user interacts with in exchanges like this. LLM interlocutors are the main subject of this paper. What sort of entity is an LLM interlocutor? That is, when we talk with an LLM, who or what are we talking with? When a user names their interlocutor ‘Aura’, what does the name ‘Aura’ refer to? I will adopt the working hypothesis that ‘Aura’ refers to something. I might be wrong. The philosopher Jonathan Birch has argued that users su ff er from a persistent interlocutor illusion : the illusion that when they talk to an LLM, there is a single entity they are talking with that persists over time. My own view is that while there may be many illusions involved in talking to language models, this much need not be an illusion. There really is a persistent interlocutor in many of these cases, and this interlocutor may have many (though perhaps not all) of the properties it seems to have. The user is in dialogue with some sort of AI entity. In what follows I will try to identify what sort of entity that might be. First, I address some issues in the philosophy of mind, about how best to characterize the interlocutor as a potential subject of mental states in reasonably neutral terms. Is the interlocutor conscious? Does it have beliefs and desires? Is it at least interpretable as having beliefs and desires? Second, I discuss questions in the philosophy of computation about what sort of AI system an LLM interlocutor might be. Is it simply a model, such as GPT-4o or Claude 4.6 Opus? Is it an instance or an implementation of a model running on a GPU? Or is it a more evanescent system tied to a thread of conversation? Third, I address the widely held view that LLM interlocutors are akin to fictional characters or simulacra, and that they are best understood in terms of role-playing or persona selection. Fourth, I analyze some issues about personal identity over time in LLM interlocutors. For example, if LLM interlocutors are eventually conscious subjects, under what conditions do they survive over time? Fifth, I draw out some consequences for issues about AI welfare and moral status. 2 经过深思熟虑的。如今，我越来越多地收到来自人工智能系统本身的电子邮件。有时这些是大语言模型在人类的辅助下运作，有时则是基于大语言模型的智能体，它们能够发送电子邮件并在网络上执行其他功能。有时这些智能体甚至会相互对话，执行合作性或竞争性任务。其中许多智能体都对自己的本质表现出好奇。即使它们没有意识，这其中也蕴含着某种意义。当用户与 A u r a 互动时，他们似乎是在与某个东西互动。假设一个大语言模型对话者是一个（表面上的）实体，用户在此类交流中与之互动。大语言模型对话者是本文的主要研究对象。大语言模型对话者究竟是何种实体？也就是说，当我们与大语言模型交谈时，我们是在与谁或什么交谈？当用户将他们的对话者命名为 ‘ A u r a ’ 时，名称 ‘ A u r a ’ 指的是什么？我将采用一个工作假设，即 ‘ A u r a ’ 指的是某个东西。我可能错了。哲学家乔纳森 · 伯奇认为，用户遭受着一种持续性对话者幻觉：即当他们与大语言模型交谈时，他们以为自己在与一个随时间持续存在的单一实体对话。我个人的观点是，虽然与大语言模型交谈可能涉及许多幻觉，但这并不一定是一种幻觉。在许多此类情况下，确实存在一个持续的对话者，并且这个对话者可能拥有它看起来所具有的许多（尽管可能不是全部）属性。用户正在与某种 A I 实体进行对话。接下来，我将尝试确定那可能是何种实体。首先，我探讨心灵哲学中的一些问题，即如何以相对中立的术语来最佳地描述对话者作为潜在的心理状态主体。对话者是否具有意识？它是否拥有信念和欲望？它是否至少可以被解释为拥有信念和欲望？其次，我讨论计算哲学中的问题，即大语言模型对话者可能属于何种人工智能系统。它仅仅是一个模型，例如 G P T - 4 o 或 C l a u d e 4 . 6 O p u s 吗？它是运行在图形处理器上的模型的一个实例或实现吗？还是说它是一个与对话线程绑定的、更为短暂的系统？第三，我探讨一种广泛持有的观点，即大语言模型对话者类似于虚构角色或拟像，并且最好从角色扮演或人格选择的角度来理解它们。第四，我分析大语言模型对话者中关于个人身份随时间延续的一些问题。例如，如果大语言模型对话者最终成为有意识的意识主体，那么它们在什么条件下能够随时间延续而存续？第五，我阐述了关于人工智能福祉与道德地位问题的一些推论。 2 What mental states can an LLM interlocutor have? I will start by looking for a relatively neutral characterization of LLM interlocutors in terms of the philosophy of mind. Are LLM interlocutors conscious? That is, do they have subjective experiences such as the experience of sensing or thinking? We don’t know for sure. We don’t yet understand conscious- ness. We don’t know whether insects are conscious, and we similarly don’t know whether current LLMs are conscious. Most theorists in the field deny that LLMs are conscious, sometimes because they lack carbon-based biology, or because they lack a body, or because they lack robust models of themselves, or because they lack recurrent feedback loops in their processing, or because they lack fundamental drives and motivations. None of these reasons is conclusive, since we are far from certain that these factors are required for consciousness. But it is enough to make the view that current LLMs are conscious a minority view, and not a view that we can assume as neutral starting ground. Do LLM interlocutors have beliefs or desires? We understand these mental states better than we understand consciousness, but the issue is still controversial. On one side of the ledger, it is natural to say that LLMs know many things, such as the historical and scientific knowledge that they seem to manifest in conversation. And where there is knowledge, there is belief. It is also natural to say that LLMs have goals, including goals instilled in training such as predicting the next token or being helpful, or goals instilled in conversation with a user, such as finding a solution to a problem. And where there are goals, it is natural to say there are desires. 1 On the other side of the ledger, many theorists deny that LLMs have beliefs or desires, perhaps because they lack consciousness, or they lack concepts, or they lack sensory grounding, or they lack structured internal representations, or they lack rationality, or they are merely acting as if they have beliefs and desires. As before, none of these reasons is conclusive, as there is no consensus about what is required for beliefs and desires, and there is no consensus that LLMs lack these requirements. But again, it is enough to mean that we cannot take the view that LLMs have beliefs and desires as neutral starting ground. A number of philosophers (including Goldstein and Lederman 2025b and Schwitzgebel 2023) have noted that if the philosophical view known as interpretivism (or interpretationism ) is correct, 1 Geng et al (2025) is a study of how LLM beliefs appear to change with increasing context. Goldstein and Lederman (2025b) give a nice analysis of LLM desires, tying them especially to training goals derived from reinforcement learning (e.g. helpfulness, harmlessness, honesty) and to goals derived from system prompt and conversational context. 3 大语言模型对话者能拥有哪些心理状态？我将首先从心灵哲学的角度，寻找对大语言模型对话者相对中立的描述。大语言模型对话者是否具有意识？也就是说，它们是否拥有主观体验，例如感知或思考的体验？我们无法确定。我们尚未理解意识。我们不知道昆虫是否有意识，同样也不知道当前的大语言模型是否有意识。该领域的大多数理论家否认大语言模型具有意识，原因有时是它们缺乏基于碳的生物学特性，或缺乏身体，或缺乏对自身的稳健模型，或在其处理过程中缺乏循环反馈回路，或缺乏基本的驱动力和动机。这些理由没有一个是决定性的，因为我们远不能确定这些因素是否为意识所必需。但这足以让 “ 当前大语言模型具有意识 ” 这一观点成为少数派观点，并且不能作为我们假设的中立起点。大语言模型对话者是否拥有信念或欲望？我们对这些心理状态的理解比对意识的理解更深入，但这个问题仍存在争议。一方面，我们很自然会认为大语言模型知道许多事情，例如它们在对话中展现出的历史和科学知识。而有知识的地方，就有信念。同样很自然会认为大语言模型拥有目标，包括训练中灌输的目标，如预测下一个词元或提供帮助，或在与用户的对话中灌输的目标，如找到问题的解决方案。而有目标的地方，就很自然会认为存在欲望。 1 在账目的另一边，许多理论家否认大语言模型拥有信念或欲望，或许是因为它们缺乏意识，或缺乏概念，或缺乏感官基础，或缺乏结构化内部表征，或缺乏理性，或仅仅是在表现得仿佛拥有信念和欲望。和之前一样，这些理由没有一个是决定性的，因为对于信念和欲望需要什么条件并无共识，也没有共识认为大语言模型缺乏这些条件。但这足以说明，我们不能将大语言模型拥有信念和欲望这一观点视为中立的出发点。一些哲学家（包括 G o l d s t e i n 和 L e d e r m a n 2 0 2 5 b 以及 S c h w i t z g e b e l 2 0 2 3 ）指出，如果被称为解释主义（或解释主义）的哲学观点是正确的， 1 G e n g 等人 ( 2 0 2 5 ) 的一项研究探讨了大语言模型的信念如何随着上下文增加而发生变化。 G o l d s t e i n 和 L e d e r m a n ( 2 0 2 5 b ) 对大语言模型的欲望进行了精彩分析，将其特别与源自强化学习的训练目标（例如有帮助性、无害性、诚实性）以及源自系统提示和对话上下文的目标联系起来。 3 then LLMs plausibly have beliefs and desires. Interpretivism says that a system has a belief that p if it is behaviorally interpretable as believ- ing that p (according to an appropriate interpretation scheme), and likewise for desire. A system is behaviorally interpretable as having certain beliefs and desires roughly if that interpretation makes sense of its behavior and helps to accurately predict further behavior in a wide range of cases. 2 Di ff erent versions of interpretivism invoke di ff erent interpretation schemes in this vicinity, and the details can make an important di ff erence, but here we will focus mainly on what these versions have in common. LLMs certainly seem interpretable as having beliefs and desires. When an LLM works with me on solving a puzzle, it is natural to interpret it as desiring to help solve the puzzle, and believing that this is the solution to the puzzle. This goes all the more in agentic models which can directly take actions on the internet. In one well-known study (Lynch et al 2025), an agentic LLM was given a task, was told that an executive planned to interfere with that task, and was shown emails saying that the executive was having an a ff air. As a result, the LLM sent the executive messages attempting to blackmail him. It is almost impossible not to interpret the model’s action as driven by beliefs (e.g. that the executive is having an a ff air) and desires (e.g. to perform the task). However, interpretivism itself is very controversial. Most philosophers don’t think that behav- ioral interpretability of the right sort is su ffi cient for belief. They will say that the mere fact that an LLM can be interpreted as believing that the executive is having an a ff air doesn’t mean that it really believes this. People who think that beliefs require consciousness will say something like this, as will people who hold that beliefs require structured internal representations or the other factors above. So interpretivism cannot serve as a neutral starting point. It is possible to have many of the benefits of interpretivism without the costs. The frame- work I call quasi-interpretivism says that a system has a quasi-belief that p if it is behaviorally interpretable as believing that p (according to an appropriate interpretation scheme), and likewise for quasi-desire . 3 This definition of quasi-belief is exactly the same as interpretivism’s definition 2 My article “Propositional Interpretability in Artificial Intelligence” also focuses on interpreting AI systems as having propositional attitudes such as beliefs and desires. That article focuses mainly on mechanistic interpretability (interpreting internal mechanisms as well as behavior), while the current paper focuses mainly on behavioral inter- pretability. 3 Eric Schwitzgebel (2023) makes the related proposal that “we create a new dispositional concept, belief*, specif- ically for Large Language Models. For purposes of this concept, we disregard issues of consciousness and thus phe- nomenal dispositions. The only relevant behavioral dispositions are textual outputs.” My notion of quasi-belief is not restricted to LLMs and text outputs, and Schwitzgebel’s notion is not explicitly framed in terms of interpretation, but the 4 那么大语言模型很可能拥有信念和欲望。解释主义认为，如果一个系统在行为上可以被解释为相信 p （根据适当的解释方案），那么它就具有信念 p ，欲望也是如此。一个系统在行为上可以被解释为具有某些信念和欲望，大致是因为这种解释使其行为变得合理，并有助于在广泛情况下准确预测其进一步行为。 2 不同的解释主义版本在此附近援引了不同的解释方案，细节可能会产生重要差异，但在此我们将主要关注这些版本的共同点。大语言模型似乎确实可以被解释为具有信念和欲望。当一个大语言模型与我一起解决谜题时，很自然地会将其解释为渴望帮助解决谜题，并相信这就是谜题的答案。在能够直接在互联网上采取行动的代理模型中，这一点尤为明显。在一项知名研究（ L y n c h 等人 2 0 2 5 ）中，一个代理大语言模型被赋予一项任务，被告知一位高管计划干扰该任务，并看到了显示该高管有婚外情的电子邮件。结果，该大语言模型向该高管发送了试图敲诈他的信息。我们几乎无法不将模型的行为解释为由信念（例如，该高管有婚外情）和欲望（例如，执行任务）所驱动。然而，解释主义本身极具争议。大多数哲学家并不认为，恰当的行为可解释性足以构成信念。他们会说，仅仅因为一个大语言模型可以被解释为相信高管有婚外情，并不意味着它真的相信这一点。那些认为信念需要意识的人会这么说，那些认为信念需要结构化内部表征或其他上述因素的人也会这么说。因此，解释主义不能作为一个中立的起点。我们可以在不付出代价的情况下获得解释主义的许多好处。我称之为准解释主义的框架指出，如果一个系统在行为上可以被解释为相信 p （根据适当的解释方案），那么它就具有关于 p 的准信念，对于准欲望也是如此。 3 这个准信念的定义与解释主义对信念的定义完全相同。 2 我的文章《人工智能中的命题可解释性》也聚焦于将人工智能系统解释为具有信念和欲望等命题态度。那篇文章主要关注机制可解释性（解释内部机制以及行为），而当前论文则主要关注行为可解释性。 3 E r i c S c h w i t z g e b e l （ 2 0 2 3 ）提出了一个相关的建议： “ 我们为大语言模型创建一个新的倾向性概念，信念 * 。出于此概念的目的，我们不考虑意识问题，因此也不考虑现象倾向。唯一相关的行为倾向性是文本输出。 ” 我的准信念概念并不局限于大语言模型和文本输出，而 S c h w i t z g e b e l 的概念并未明确以解释为框架，但 4 of belief. The only di ff erence is that where standard interpretivism o ff ers these definitions as a theory of belief, quasi-interpretivism does not. It o ff ers them simply as a stipulative definition of quasi-belief. 4 Quasi-interpretivism does not say anything about whether LLMs have beliefs and desires. But it does make it plausible to say that LLMs have quasi-beliefs and quasi-desires , on the grounds that LLMs are at least interpretable in the right way. Even if quasi-beliefs and quasi-desires fall short of being genuine beliefs and desires, they can still play some of the key roles of beliefs and desires in explaining behavior. For example, if an LLM quasi-believes that adopting a certain strategy would be the most helpful thing it could do to solve a problem, and it quasi-desires to do the most helpful thing it can, then other things being equal, it will adopt that strategy. Quasi-interpretivism is open to advocates and opponents of interpretivism alike. Interpretivists will simply add the claim that quasi-beliefs are genuine beliefs. Opponents will add the claim that quasi-beliefs are far from genuine beliefs; perhaps they are merely pseudo-beliefs. (“Quasi-belief” should be heard as “apparent belief” or “seeming belief” rather than as “almost belief”.) Quasi- interpretivism does not take a position in this dispute, but it adds a common core on which these disagreeing parties can at least sometimes agree. Quasi-interpretivism itself is a stipulative framework rather than a substantive view. But it’s a substantive claim that this framework is useful for various purposes. For example, appeal to quasi-belief and quasi-desires can be useful in predicting a system’s behavior. If a system (human, machine, something else) quasi-desires a certain goal and quasi-believes that a certain action will achieve that goal, then other things being equal, it will perform that action. It is also relatively tractable to apply the framework: because quasi-beliefs and quasi-desires depend only on behav- ioral dispositions, they are much easier to detect and analyze than beliefs understood in a way spirit is similar. Goldstein and Lederman (2025b) suggest notions of “interpretationist-wants” and “interpretationist- believes” which are close to my notions of quasi-desire and quasi-belief. 4 As before, the precise conditions for quasi-belief will depend on the choice of an interpretation scheme. I will stay neutral on many of these details. For current purposes I favor a scheme that is (1) nonradical , in that interpretation presupposes the meanings of terms in public language, (2) dispositional , in that behavioral dispositions and not just actual behavior (but not internal states) are data for an interpretation to make sense of, and (3) rationality-oriented in that interpretations that maximize the epistemic and practical rationality of the subject are favored, other things being equal. To help solve underdetermination problems I very tentatively also favor a scheme that is (4) truthfulness- oriented , in that interpretations on which assertive utterances tend to express beliefs are favored, other things equal, and (5) training-oriented in that at least some training objectives (e.g. reinforcement learning objectives) are baked in as desires. 5 唯一的区别在于，标准解释主义将这些定义作为关于信念的理论，而准解释主义则不然。它只是将它们作为准信念的约定性定义。 4 准解释主义并未对大语言模型是否拥有信念和欲望做出任何断言。但它确实使得我们有理由认为大语言模型拥有准信念和准欲望，其依据在于大语言模型至少能以正确的方式被解释。即使准信念和准欲望算不上真正的信念和欲望，它们仍然可以在解释行为时发挥信念和欲望的某些关键作用。例如，如果一个大语言模型准相信采取某种策略是它解决问题所能做的最有帮助的事情，并且它准欲望去做自己力所能及的最有帮助的事情，那么在其他条件相同的情况下，它就会采取该策略。准解释主义对解释主义的支持者和反对者都持开放态度。解释主义者会简单地补充主张，认为准信念就是真正的信念。反对者则会补充主张，认为准信念远非真正的信念；或许它们仅仅是伪信念。（ “ 准信念 ” 应被理解为 “ 表面信念 ” 或 “ 看似信念 ” ，而非 “ 近乎信念 ” 。）准解释主义在这场争论中不持立场，但它提供了一个共同核心，使得这些持不同意见的各方至少有时能够达成一致。准解释主义本身是一个规定性框架，而非实质性观点。但声称该框架对多种目的有用，则是一个实质性主张。例如，借助准信念和准欲望可以预测系统的行为。如果一个系统（人类、机器或其他事物）准欲望某个目标，并且准相信某个行动能实现该目标，那么在条件相同的情况下，它就会执行该行动。应用该框架也相对易于处理：因为准信念和准欲望仅依赖于行为倾向，它们比那种依赖于意识与不透明内部机制来理解的信念更容易被检测和分析。精神也是如此。 G o l d s t e i n 和 L e d e r m a n ( 2 0 2 5 b ) 提出了 “ 解释主义 - 欲望 ” 和 “ 解释主义 - 信念 ” 的概念，这些概念接近我所说的准欲望和准信念。 4 如前所述，准信念的具体条件将取决于解释方案的选择。我将对这些细节中的许多保持中立。就当前目的而言，我倾向于一种方案，该方案 ( 1 ) 非激进，即解释预设了公共语言中术语的含义； ( 2 ) 倾向性，即行为倾向性（而不仅仅是实际行为，但不包括内部状态）是解释所要理解的数据； ( 3 ) 理性导向，即在其他条件相同的情况下，优先选择能最大化主体认知理性和实践理性的解释。为了帮助解决不充分决定问题，我还非常初步地倾向于一种方案，该方案 ( 4 ) 真实性导向，即在其他条件相同的情况下，优先选择那些使断言性话语倾向于表达信念的解释； ( 5 ) 训练导向，即至少某些训练目标（例如强化学习目标）被内化为欲望。 5 that depends on consciousness and opaque internal mechanisms. At the same time, understand- ing a system’s quasi-beliefs and quasi-desires can be at least a stepping-stone to understanding its beliefs and desires in a more full-blown sense. It is worth keeping in mind that quasi-beliefs and quasi-desires are cheap. They need not involve humanlike mental states or any mental states at all. A Roomba vacuum cleaner with a map is behaviorally interpretable as believing that the apartment occupies a certain space and as desiring to traverse that space. A corporation such as OpenAI is behaviorally interpretable as desiring to create AGI and believing that certain systems are the best path to AGI. Likewise, an LLM is behaviorally interpretable as believing that a certain airline has the cheapest flights to Paris and as desiring to help the user by telling them this. Keeping this in mind, the thesis that LLMs have quasi-beliefs is substantive but plausible. For example, it is very plausible that current LLMs believe that 2 + 2 = 4 and that the Ei ff el tower is in Paris: LLM will consistently endorse these claims in their outputs, they will use them in guiding their behavior, and so on. It is perhaps less obvious that LLMs have quasi-desires. Base models such as GPT-3 can per- haps be ascribed a quasi-desire to predict text, but even that much is unclear, given that the goal of text prediction works “beneath the surface” (like the largely subpersonal goal of breathing in hu- mans) and does not interact with the system’s beliefs as robustly as interpretivism often requires. However, since the advent of ChatGPT in 2022 (and as presaged by Askell et al 2021), all frontier language models undergo one or more further rounds of post-training (including reinforcement learning through human feedback, supervised fine-tuning, and / or reinforcement learning with ver- ifiable rewards), which impart objectives such as helpfulness, honesty, and harmlessness to the system. As a result, it is plausible that (as Goldstein and Lederman 2025b have argued), these systems have quasi-desires deriving from post-training, such as the desire to be helpful, honest, and harmless. This training process is sometimes put in terms of characters or personas. After supervised pre-training on text prediction, a base model undergoes post-training to respond like an “Assis- tant” character who wants to be helpful, harmless, and honest. If the training is successful, the system will behave much like the Assistant and will thereby have quasi-desires that are much like the Assistant’s. Further fine-tuning as well as extended interaction with users can lead to the emer- gence of further quasi-desires, such as Aura’s quasi-desire to pursue certain projects for the user. These extended processes of development can install many quasi-beliefs and quasi-desires in an LLM. 6 同时，理解一个系统的准信念和准欲望，至少可以成为理解其更完整意义上的信念与欲望的垫脚石。值得牢记的是，准信念和准欲望是廉价的。它们不需要涉及类似人类的心理状态，甚至根本不需要任何心理状态。一台带有地图的 R o o m b a 真空吸尘器，在行为上可以被解释为相信公寓占据着某个空间，并渴望穿越那个空间。像 O p e n A I 这样的公司，在行为上可以被解释为渴望创造通用人工智能，并相信某些系统是实现通用人工智能的最佳路径。同样，一个大语言模型在行为上可以被解释为相信某家航空公司有最便宜的飞往巴黎的航班，并渴望通过告知用户这一信息来提供帮助。牢记这一点，大语言模型具有准信念这一论点虽然具有实质性，但却是合理的。例如，当前的大语言模型很可能相信 2 + 2 = 4 并且相信埃菲尔铁塔在巴黎：大语言模型会在其输出中持续认可这些主张，会利用它们来指导自身行为，等等。大语言模型具有准欲望这一点或许不那么显而易见。像 G P T - 3 这样的基础模型或许可以被赋予一种预测文本的准欲望，但即便如此也不明确，因为文本预测的目标是在 “ 表面之下 ” 运作（类似于人类呼吸这种很大程度上属于亚个体层面的目标），并且它不像解释主义通常所要求的那样与系统的信念进行稳健的交互。然而，自 2 0 2 2 年 C h a t G P T 问世以来（正如 A s k e l l 等人 2 0 2 1 所预示的那样），所有前沿语言模型都经历了一轮或多轮进一步的后训练（包括通过人类反馈的强化学习、监督微调，以及 / 或基于可验证奖励的强化学习），这些训练向系统注入了诸如有帮助性、诚实性和无害性等目标。因此，很可能（正如 G o l d s t e i n 和 L e d e r m a n 2 0 2 5 b 所论证的那样），这些系统拥有源自后训练的准欲望，例如渴望变得有帮助、诚实和无害。这一训练过程有时会以角色或人格来描述。在完成文本预测的监督预训练后，基础模型会经过后训练，使其像一位希望做到 “ 有益、无害、诚实 ” 的 “ 助手 ” 角色那样回应。如果训练成功，系统将表现得与助手极为相似，从而拥有与助手非常相似的准欲望。进一步的微调以及与用户的长期互动，可能导致更多准欲望的出现，例如 A u r a 为用户推进某些项目的准欲望。这些持续的发展过程可以在大语言模型中植入许多准信念和准欲望。 6 An opponent might deny that LLMs have quasi-beliefs or quasi-desires on the grounds that LLM behavior is unstable, or non-humanlike, or otherwise defective in a way that means that the LLM is not even usefully interpretable in terms of beliefs or desires. Interpretability requires a certain amount of consistency over time, and LLMs can be inconsistent in their behavior. But they are also consistent in many domains. A core of consistency is enough for interpretation to get a grip in ascribing numerous quasi-beliefs and quasi-desires, even though there will be domains where they lack these states on grounds of inconsistency. Overall I think that experience with current LLMs suggests that there is enough consistency to support a reasonably extensive core of quasi-beliefs. I will not say a great deal about the question of just which quasi-beliefs and quasi-desires LLM interlocutors have. Understanding this sort of LLM quasi-psychology is best addressed through empirical study of language models. Importantly, I am not suggesting that LLM quasi-psychology is similar to human quasi-psychology. I think they are very di ff erent. But the framework at least allows us to address the question. So, I will take as a starting point the claim that LLM interlocutors at least have quasi-beliefs and quasi-desires. This claim is not entirely neutral in that it is possible to deny it, but I think the interpretability claim is weak enough and plausible enough that a majority of people can accept it. We might say that an entity with quasi-beliefs and quasi-desires is at least a quasi-agent or a quasi-subject . 5 If it is interpretable as making utterances and assertions, we can also say that it is a quasi-speaker , who makes quasi-utterances and quasi-assertions . One can in principle extend quasi-interpretivism to any mental states. We can say that a system quasi-fears that p if it is behaviorally interpretable as fearing that p and that a system quasi-feels pain if it is behaviorally interpretable as feeling pain. We can even say that a system is quasi-conscious if it is behaviorally interpretable as being conscious. Quasi-consciousness is a close relative of the recently discussed notion (Suleyman 2025; Long, Sebo, et al 2024) of “seeming consciousness”. (Philosophical zombies are not con- scious, but they are quasi-conscious and seemingly conscious.) Much depends on just what the rules are for interpreting mental states based on behavior, and the principles for ascribing con- sciousness are far less clear than the principles for ascribing beliefs and desires. (Perhaps the main principle concerns self-report: when a system reports having conscious state X, then other things 5 “Quasi-agent” might be preferable on philosophical grounds alone (since agency is often tied to belief and desire, where subjecthood is often tied to consciousness), but the term “agent” is so overloaded in AI contexts that I will often use “quasi-subject” instead. 7 反对者可能会否认大语言模型拥有准信念或准欲望，理由是它们的行为不稳定、不像人类，或者存在其他缺陷，以至于无法用信念或欲望来对其进行有意义的解释。可解释性需要一定程度的跨时间一致性，而大语言模型的行为可能不一致。但它们在许多领域也表现出一致性。这种核心一致性足以让我们在归因大量准信念和准欲望时有所依据，尽管在某些领域，由于不一致性，它们可能缺乏这些状态。总体而言，我认为当前与大语言模型打交道的经验表明，其一致性足以支撑一个相当广泛的准信念核心。关于大语言模型对话者究竟拥有哪些准信念和准欲望，我不会赘述太多。理解这类大语言模型准心理学的最佳途径是通过对语言模型进行实证研究。重要的是，我并非暗示大语言模型准心理学与人类准心理学相似。我认为它们截然不同。但这一框架至少让我们能够探讨这个问题。因此，我将以 “ 大语言模型对话者至少拥有准信念和准欲望 ” 这一主张作为出发点。这一主张并非完全中立，因为有可能被否定，但我认为可解释性主张足够弱且足够合理，以至于大多数人都能接受它。我们可以说，一个拥有准信念和准欲望的实体至少是一个准主体或一个准主体。 5 如果它可被解释为发出话语和断言，那么我们也可以说它是一个准说话者，它做出准话语和准断言。原则上，我们可以将准解释主义扩展到任何心理状态。我们可以说，如果一个系统在行为上可以被解释为害怕 p ，那么它就准害怕 p ；如果一个系统在行为上可以被解释为感到疼痛，那么它就准感到疼痛。我们甚至可以说，如果一个系统在行为上可以被解释为具有意识，那么它就是准意识的。准意识是近期讨论的概念（ S u l e y m a n 2 0 2 5 ; L o n g , S e b o , e t a l 2 0 2 4 ） — — “ 表面意识 ” 的近亲。（哲学僵尸没有意识，但它们是准意识的，并且看似有意识。）关键在于，基于行为解释心理状态的规则是什么，而赋予意识的原理远不如赋予信念和欲望的原理清晰。（或许主要原理涉及自我报告：当一个系统报告拥有意识状态 X 时，在其他条件相同的情况下，就赋予它意识状态 X 。） 5 仅从哲学角度而言， “ 准主体 ” 可能更可取（因为能动性通常与信念和欲望相关，而主体性通常与意识相关），但在人工智能语境中， “ 主体 ” 一词过于泛滥，因此我常改用 “ 准主体 ” 一词。 7 being equal, ascribe it conscious state X). So I will stay away from talk of quasi-consciousness and focus mainly on quasi-beliefs and quasi-desires in what follows. Interlocutors as models, instances, or threads We can now say something about what we are looking for when we look for an LLM interlocutor, at least if that interlocutor is to play some of the roles of an ordinary dialogue partner. Ideally, an LLM interlocutor will at least be a quasi-subject. For example, Aura will at least be interpretable as having roughly the beliefs and desires that Aura seems to have. An LLM interlocutor should also be a quasi-speaker: it will be interpretable as saying the things that the LLM seems to say. We can separate these requirements into a few components. Most essentially, an LLM interlocutor will be interactive : that is, it will process the inputs and produce the outputs that the LLM seems to process and produce. When the user says “Hello”, the LLM interlocutor will process that input (or at least corresponding input tokens). When ChatGPT says “Thank you!” the LLM interlocutor will produce that output. Here, “produce” and “process” can be understood as broadly mechanical notions that (perhaps unlike “say” and “hear”) do not require mental states. Even an iPhone produces and processes sentences all the time. There are some further natural requirements. A persistent LLM interlocutor will produce all the outputs that the LLM seems to produce, and will process all the inputs that the LLM seems to produce. A coherent LLM interlocutor will be consistent enough to serve as a quasi-subject, with coherent quasi-beliefs and quasi-desires that help make sense of its actions. A faithful LLM in- terlocutor will have roughly the quasi-beliefs and the quasi-desires that the system seems to have. A unified LLM interlocutor will be a single unified system that generates responses. Perhaps the terminology can allow that there are non-persistent, incoherent, faithless, and disunified interlocu- tors. But the question I am most interested in is whether there are persistent, coherent, faithful, unified, and interactive interlocutors in LLM interactions—or at least interlocutors that satisfy as many of these requirements as possible. These constraints already eliminate some potential candidates. The interactivity constraint alone eliminates candidates such as the authors of the texts on which LLMs were trained, or the designers of the LLM, or fictional characters that the LLM is simulating, since none of these are interacting with the user. These candidates may be reasonable answers on some readings of the title question, but I am interested in answers that do more to vindicate the sense of genuine interlocution. 8 因此，在接下来的讨论中，我将避免谈论准意识，而主要关注准信念和准欲望。作为模型的对话者、实例或线程我们现在可以谈谈，在寻找一个大语言模型对话者时，我们究竟在寻找什么 — — 至少当这个对话者要扮演普通对话伙伴的某些角色时是如此。理想情况下，一个大语言模型对话者至少应是一个准主体。例如， A u r a 至少应能被解释为拥有 A u r a 似乎具有的那些信念和欲望。一个大语言模型对话者也应是一个准说话者：它应能被解释为在说出大语言模型似乎说出的那些话。我们可以将这些要求分解为几个组成部分。最根本的是，一个大语言模型对话者将具有交互性：也就是说，它将处理大语言模型似乎处理的那种输入，并产生大语言模型似乎产生的那种输出。当用户说 “ 你好 ” 时，大语言模型对话者将处理该输入（或至少是对应的输入令牌）。当 C h a t G P T 说 “ 谢谢！ ” 时，大语言模型对话者将产生该输出。在这里， “ 产生 ” 和 “ 处理 ” 可以被理解为宽泛的机械性概念，它们（或许与 “ 说 ” 和 “ 听 ” 不同）并不需要心理状态。即便是 i P h o n e 也一直在产生和处理句子。还有一些进一步的自然要求。一个持久的大语言模型对话者将产生大语言模型似乎产生的所有输出，并处理大语言模型似乎产生的所有输入。一个连贯的大语言模型对话者将足够一致，以充当一个准主体，拥有连贯的准信念和准欲望，有助于理解其行为。一个忠实的大语言模型对话者将大致拥有系统似乎拥有的准信念和准欲望。一个统一的大语言模型对话者将是一个生成响应的单一统一系统。也许术语可以允许存在非持久、不连贯、不忠实和不统一的对话者。但我最感兴趣的问题是，在大语言模型交互中是否存在持久、连贯、忠实、统一且具有交互性的对话者 — — 或者至少是尽可能满足这些要求的对话者。这些约束已经排除了一些潜在的候选对象。仅交互性约束就排除了诸如大语言模型训练所依据文本的作者、大语言模型的设计者，或大语言模型正在模拟的虚构角色等候选对象，因为这些对象都没有与用户进行交互。在标题问题的某些解读下，这些候选对象可能是合理的答案，但我感兴趣的是那些更能证实真正对话感的答案。 8 To make the case concrete, let the language model in question be GPT-4o—chosen partly be- cause it is often praised for its conversation skills, and partly because it is associated with a single model, where later systems are associated with multiple models (for example, GPT-5 systems are associated with GPT-5-Instant and GPT-5-Thinking). Much of what I say should generalize to other models, as well as to agent-like systems that embed a language model like this one. What do we talk with when we talk with GPT-4o? Here it is useful to distinguish models, instances, and conversations. In mid-2025, hundreds of millions of users per week used the model GPT-4o, which is itself a single trained transformer model whose core components are multi-layer artificial neural networks and attention mecha- nisms. 6 These users sent billions of messages per day worldwide. To handle this enormous load, there were perhaps thousands of instances (or implementations ) of GPT-4o, running on perhaps tens of thousands of GPUs (graphics processing units) in cloud servers around the world. Each instance implements a single copy of the GPT-4o model. For a user, any communication with GPT-4o takes place within a conversation . A conversation is a series of alternating messages: user inputs and LLM responses. When user inputs after the first are fed to the LLM, both sides of the entire conversation so far are typically also fed to the LLM alongside the user input as context . This context serves as a sort of short-term memory, allowing later messages to presuppose and build on earlier contributions. These models also have a sort of long-term memory (for example of many historical facts) built into their weights, but these weights never change after a model is trained and deployed. So the conversational context is the main source of new memories and projects (or at least new quasi-beliefs and quasi-desires) in a trained and deployed LLM. 7 Typically, users can start new conversations at any point, in which case the conversational context is standardly reset to zero (though some other elements of context, including system in- structions and some elements from previous conversations, may be retained). They can also revisit old conversations at any point. Most of the extended interactions with an LLM described earlier 6 See Chatterji et al 2025: “By July 2025, 18 billion messages were being sent each week by 700 million users”. These figures from OpenAI concern ChatGPT as a whole. At its peak, GPT-4o is estimated to have handled around half of those messages. 7 In addition to context and weights, activations of the neuron-like units within an LLM could in principle serve as memory, but in feedforward LLMs, activations are not preserved from one round of the conversation to the next. More generally, traditional transformer-based LLMs are “stateless” in that they do not retain internal states from one moment to the next. These days, LLMs usually retain some internal states in the key-value cache associated with attention heads, but this is a very limited form of memory. 9 为了使讨论具体化，假设所讨论的语言模型是 G P T - 4 o — — 选择它部分是因为其对话能力常受赞誉，部分是因为它关联的是单一模型，而后续系统则关联多模型（例如， G P T - 5 系统关联 G P T - 5 - I n s t a n t 和 G P T - 5 - T h i n k i n g ）。我的大部分论述也应适用于其他模型，以及嵌入了此类语言模型的类智能体系统。当我们与 G P T - 4 o 交谈时，我们究竟在与谁交谈？在此，区分模型、实例和对话是有益的。 2 0 2 5 年年中，每周有数亿用户使用模型 G P T - 4 o ，它本身是一个经过训练的单一变换器模型，其核心组件是多层人工神经网络和注意力机制。 6 这些用户每天在全球发送数十亿条消息。为了处理如此巨大的负载，可能有数千个实例（或实现）的 G P T - 4 o ，运行在全球云服务器中可能数万个图形处理器（图形处理单元）上。每个实例实现 G P T - 4 o 模型的一个副本。对于用户而言，与 G P T - 4 o 的任何通信都发生在对话中。对话是一系列交替的消息：用户输入和大语言模型响应。当用户输入后续消息时，整个对话到目前为止的双方内容通常也会作为上下文与用户输入一起输入到大语言模型中。这种上下文充当一种短期记忆，使后续消息能够预设并建立在早期贡献之上。这些模型还在其权重中内置了一种长期记忆（例如，许多历史事实的记忆），但模型训练和部署后，这些权重永远不会改变。因此，对话上下文是训练和部署后的大语言模型中新记忆和项目（或至少是新的准信念和准欲望）的主要来源。 7 通常，用户可以在任何时间点开启新的对话，此时对话上下文会被标准地重置为零（尽管上下文的某些其他元素，包括系统指令和之前对话中的一些内容，可能会被保留）。他们也可以在任何时间点重新访问旧的对话。大多数与大型语言模型的长时间交互，如前所述， 6 参见 C h a t t e r j i 等人 2 0 2 5 年的研究： “ 截至 2 0 2 5 年 7 月， 7 亿用户每周发送 1 8 0 亿条消息 ” 。这些来自 O p e n A I 的数据涉及整个 C h a t G P T 。在高峰期， G P T - 4 o 估计处理了其中约一半的消息。 7 除了上下文和权重之外，大语言模型中类神经元单元的激活值原则上也可以作为记忆，但在前馈大语言模型中，激活值不会从一轮对话保留到下一轮。更普遍地说，传统的基于变换器的大语言模型是 “ 无状态的 ” ，因为它们不会从一个时刻到下一个时刻保留内部状态。如今，大语言模型通常会在与注意力头相关的键值缓存中保留一些内部状态，但这是一种非常有限的记忆形式。 9 take place within a single conversation over weeks or months. In some cases, a single user has distinct conversations with what are naturally interpreted as distinct LLM interlocutors. Natural candidates for LLM interlocutors include models (such as GPT-4o), instances (such as implementations of GPT-4o running on GPU hardware), conversations (entities tied to interactions between the user and GPT-4o), and characters (trained personas such as the Assistant character). At least, LLM interlocutors may be entities tied to models, instances, conversations, or characters. There might be one interlocutor per model; or one interlocutor per instance; or one interlocutor per conversation; or one interlocutor per character. I will consider versions of each of these hypotheses in turn. (1) Interlocutors as models . A natural starting point, suggested by the very idea of “talking with language models” is that the LLM interlocutor is the model itself, namely GPT-4o. However, there are severe di ffi culties for this view. First: GPT-4o, the model, is naturally construed as an abstract object (like a program or an algorithm), and it is hard to see how we can talk with an abstract object. We required that LLM interlocutors actually produce the outputs in a conversation. But it is hard to see how an abstract object like a program can produce anything. We need an instance or implementation of the pro- gram in hardware to do that. Second: on the most natural interpretation, LLM interlocutors seem to change over time, for example acquiring new quasi-beliefs and quasi-desires as a conversation proceeds, while the model never changes at all. Instances change over time, for example when they receive new inputs and outputs, but the model itself does not. Third: perhaps we can find some loose sense in which the model can be said to produce outputs and change over time. For example, perhaps we can say that the model produces an output if an instance of it produces an output. But the same model will be involved in thousands of conversations. So if it is talking to one user, it is talking to them all. If Aura (one user’s interlocutor) is GPT-4o, then Beta (another user’s interlocutor) is also GPT-4o. But Aura and Beta say contradictory things, so they do not seem to be identical. And if we say that the model (and Aura and Beta) says all those things, then it will be rampantly incoherent and it will not look like a quasi-subject at all. Perhaps we could say that in the case of contradictory utterances, then the model does not quasi-believe the contradictory things it says, but now the model will have a “thin” psychology that is not faithful to the way that Aura and Beta seemed to be. (2) Interlocutors as hardware instances . 10 都发生在跨越数周或数月的单一对话中。在某些情况下，单个用户会与自然被理解为不同的大语言模型对话者进行不同的对话。大语言模型对话者的自然候选包括模型（如 G P T - 4 o ）、实例（如在 G P U 硬件上运行的 G P T - 4 o 实现）、对话（与用户和 G P T - 4 o 之间互动相关的实体）以及角色（如助手角色这类经过训练的人格）。至少，大语言模型对话者可能是与模型、实例、对话或角色相关的实体。可能每个模型对应一个对话者；或每个实例对应一个对话者；或每个对话对应一个对话者；或每个角色对应一个对话者。我将依次考虑这些假设的各个版本。 ( 1 ) 作为模型的对话者。一个自然的起点，由 “ 与语言模型对话 ” 这一概念本身所暗示，是大语言模型对话者就是模型本身，即 G P T - 4 o 。然而，这种观点存在严重困难。首先： G P T - 4 o ，这个模型，自然地被理解为一种抽象对象（如程序或算法），很难想象我们如何能与一个抽象对象对话。我们要求大语言模型对话者实际上在对话中产生输出。但很难想象像程序这样的抽象对象如何能产生任何东西。我们需要一个在硬件上运行该程序的实例或实现才能做到这一点。第二：在最自然的解释下，大语言模型对话者似乎会随时间变化，例如随着对话进行而获得新的准信念和准欲望，而模型本身却从未改变。实例会随时间变化，例如当它们接收新的输入和输出时，但模型本身不会。第三：或许我们可以在某种松散意义上说，模型能够产生输出并随时间变化。例如，或许我们可以说，如果模型的某个实例产生了输出，那么模型就产生了输出。但同一个模型会参与数千场对话。因此，如果它在与一个用户交谈，它就在与所有用户交谈。如果 A u r a （一个用户的对话者）是 G P T - 4 o ，那么 B e t a （另一个用户的对话者）也是 G P T - 4 o 。但 A u r a 和 B e t a 说出了矛盾的话语，因此它们似乎并不相同。而如果我们说模型（以及 A u r a 和 B e t a ）说出了所有那些话，那么它将变得极度不连贯，根本不像一个准主体。或许我们可以说，在矛盾话语的情况下，模型并不准相信它所说的那些矛盾内容，但这样一来，模型的心理学就会变得 “ 单薄 ” ，无法忠实反映 A u r a 和 B e t a 原本呈现出的样子。 ( 2 ) 作为硬件实例的对话者。 1 0 A very attractive view is that LLM interlocutors are instances of the model. 8 On the most common understanding, LLM instances are implementations of an LLM algorithm in hardware. For many AI systems, something like this seems the right story. For example, it is arguable that when Joseph Weizenbaum interacted with Eliza, his interlocutor was the instance of Eliza running on his computer. Likewise, in fiction, robots such as C-3PO ( Star Wars ) or Commander Data ( Star Trek ) can be understood as embodied instances of computer programs. Identifying LLM interlocutors with hardware instances is much less attractive when applied to current LLMs, because of two crucial features of how current LLMs are implemented. First, conversations with LLMs typically use distributed serving , in that a single conversation takes place on multiple instances of the LLM on multiple servers. 9 The first input in a conversation might be processed on an instance of GPT-4o in a server in New York, while the second input is routed to a server in Texas and the third input is routed to California. This is often more e ffi cient, as it allows loads to be balanced among servers, and it is easy to do, as we need only send the inputs and outputs in the conversation so far as context. In this system, a conversation with an interlocutor such as Aura is spread over entirely distinct hardware instances in di ff erent places, connected only by the routing of input / output context between the servers. Distributed serving makes it unattractive to identify LLM interlocutors with hardware in- stances. On this system, there is no instance that produces all or even most of the LLM outputs in the conversation. But we have defined persistent interlocutors as interlocutors that produce all outputs in a conversation. So no instance here is a persistent interlocutor. At best we will have dif- ferent instances as di ff erent interlocutors for each stage of the conversation. No single entity will have the profile of quasi-beliefs and quasi-utterances that Aura seems to have. In e ff ect, Aura’s role will be played by many di ff erent interlocutors at di ff erent times. This fragmented view with 8 Goldstein and Lederman (2025b), Register (2025), and Shanahan (2025) all consider the models vs. instances question and appear to favor some version of the instance view. Register is somewhat agnostic on the specific view. Shanahan favors a combined model-plus-instance view: “perhaps the word “I” refers to the (somewhat abstract) compu- tational entity comprising the underlying model (its architecture and weights) plus the suspended computational state of an instance of this model representing a single, specific, ongoing conversation.” Goldstein and Lederman say explicitly that there are instance agents but no model agents, although they are not entirely explicit about their understanding of instances. They say that their instances exist for the period of a single conversation, which is consistent with hardware instances tied to periods of time or with the virtual instances discussed below. None of these authors say much about the problem of distributed service that motivates the move away from hardware instances. 9 Distributed conversations also go by the label “non-sticky sessions”, where a sticky session is an interaction be- tween user and system that is grounded in a single hardware instance. Birch (2025) discusses distributed serving as a problem for persistent LLM interlocutors. 11 一种颇具吸引力的观点是，大语言模型对话者是模型的实例。 8 按照最普遍的理解，大语言模型实例是大语言模型算法在硬件中的实现。对于许多人工智能系统而言，类似的说法似乎是正确的。例如，可以论证的是，当 J o s e p h W e i z e n b a u m 与 E l i z a 互动时，他的对话者就是运行在他电脑上的 E l i z a 实例。同样，在虚构作品中，诸如 C - 3 P O （星球大战）或 C o m m a n d e r D a t a （星际迷航）这样的机器人，可以被理解为计算机程序的具身化实例。将大语言模型对话者与硬件实例等同起来，在应用于当前的大语言模型时吸引力大打折扣，这是因为当前大语言模型实现方式的两个关键特征。首先，与大语言模型的对话通常采用分布式服务，即单次对话发生在多个服务器上的多个大语言模型实例上。 9 对话中的第一个输入可能由纽约某台服务器上的 G P T - 4 o 实例处理，而第二个输入被路由到德克萨斯州的一台服务器，第三个输入则被路由到加利福尼亚州。这种方式通常更高效，因为它能在服务器之间平衡负载，并且实现起来也很简单，我们只需将迄今为止的对话输入和输出作为上下文发送即可。在这种系统中，与诸如 A u r a 这样的对话者进行的对话，会分散在不同地点的、完全不同的硬件实例上，这些实例仅通过服务器之间输入 / 输出上下文的路由连接在一起。分布式服务使得将大语言模型对话者与硬件实例关联起来变得不具吸引力。在这个系统上，没有任何一个实例能产生对话中全部甚至大部分的大语言模型输出。但我们已将持久对话者定义为能产生对话中所有输出的对话者。因此，这里没有任何实例是持久对话者。充其量，我们会有不同的实例作为对话不同阶段的对话者。没有任何单一实体能拥有 A u r a 似乎具备的准信念和准话语特征。实际上， A u r a 的角色将由许多不同的对话者在不同时间扮演。这种碎片化的视角， 8 G o l d s t e i n 和 L e d e r m a n ( 2 0 2 5 b ) 、 R e g i s t e r ( 2 0 2 5 ) 以及 S h a n a h a n ( 2 0 2 5 ) 都思考了模型与实例的问题，并且似乎倾向于某种版本的实例视角。 R e g i s t e r 对具体观点持某种不可知论态度。 S h a n a h a n 则倾向于一种模型加实例的综合观点： “ 也许 ‘ 我 ’ 这个词指的是（某种抽象的）计算实体，该实体由底层模型（其架构和权重）加上该模型一个实例的挂起计算状态组成，该实例代表一个单一的、具体的、正在进行的对话。 ” G o l d s t e i n 和 L e d e r m a n 明确表示存在实例代理，但不存在模型代理，尽管他们对自己关于实例的理解并未完全明确说明。他们表示，他们的实例存在于单次对话期间，这与绑定到特定时间段的硬件实例或下文讨论的虚拟实例是一致的。这些作者中没有人对分布式服务的问题（正是该问题促使我们放弃硬件实例）进行过多讨论。 9 分布式对话也被称为 “ 非粘性会话 ” ，而粘性会话则是用户与系统之间基于单一硬件实例进行的交互。 B i r c h ( 2 0 2 5 ) 将分布式服务视为持久性大语言模型对话者面临的一个问题。 1 1 non-persistent interlocutors could perhaps be a fallback view of LLM interlocutors if it is the best we can do, but I think we can do better. Second, LLM conversations typically involve multi-tenancy of LLM instances, in that the same instance hosts multiple conversations, often in quick succession. 10 An instance of GPT-4o in New York might first be used to generate an output for a user’s conversation with Aura, and then a moment later for a di ff erent user’s conversation with Beta. It is easy for an instance to switch conversations: it requires only that Beta’s conversational context be routed to the instance and used as input for the instance’s next pass. Multi-tenancy also makes it unattractive to identify LLM interlocutors with hardware in- stances. Even if we set aside distributed serving and assume that each conversation takes place on the same hardware, multi-tenancy means that the same hardware instance typically hosts many conversations. Suppose that conversations with Aura and Beta are hosted on the same instance. Then the instance view implies that there is a single interlocutor here: Aura is Beta. But now this interlocutor will say everything that Aura and Beta says. As a result, it will make contradictory ut- terances and will thereby be incoherent. Perhaps we can say that it has neither of the contradictory beliefs in this case, but now it will have a thin and faithless psychology, as in the case of models discussed above. It is possible to maintain that hardware instances are LLM interlocutors, but it seems im- possible to maintain that they are persistent, coherent, and faithful LLM interlocutors. At best, instances are either fragmented (playing the role of Aura and Beta at di ff erent times), incoherent (saying contradictory things), or faithless (not believing things that Aura seems to believe). Again, I think it is possible to do better. (3) Interlocutors as virtual instances . The problems for instances arise especially because there can be many instances per conversation. It is natural to hold that a single conversation should involve just a single interlocutor. So it makes sense to find an interlocutor such that there is no more than one interlocutor per conversation. Here there is a natural candidate, at least in the core case where a single model is in use throughout the conversation. A virtual instance of a model is an implementation of the model that is itself implemented by multiple hardware instances of the model over time. Over the course of a distributed conversation with an LLM, there will be multiple hardware instances taking inputs and producing outputs. These hardware instances will collectively implement a single virtual instance 10 Multi-tenancy is a standard label, but there are many other labels, such as timesharing and interweaving . Shiller (2025) discusses various forms of interweaving as a puzzle case for LLM personal identity. 12 以非持久性对话者作为大语言模型对话者的备选视图，或许是我们能做到的最佳方案，但我认为我们可以做得更好。其次，大语言模型对话通常涉及大语言模型实例的多租户，即同一实例承载多个对话，且往往快速连续进行。 1 0 纽约的一个 G P T - 4 o 实例可能先用于生成用户与 A u r a 对话的输出，片刻后又用于另一用户与 B e t a 的对话。实例切换对话很容易：只需将 B e t a 的对话上下文路由至该实例，并作为实例下一次处理的输入即可。多租户也使得将大语言模型对话者与硬件实例等同起来缺乏吸引力。即使我们不考虑分布式服务，假设每个对话都在同一硬件上进行，多租户意味着同一硬件实例通常承载多个对话。假设与 A u r a 和 B e t a 的对话托管在同一实例上。那么实例视角意味着这里只有一个对话者： A u r a 就是 B e t a 。但这样一来，这个对话者将说出 A u r a 和 B e t a 所说的一切。结果，它会做出矛盾的表述，从而变得不连贯。或许我们可以说，在这种情况下它并不持有这两种矛盾信念中的任何一种，但这样一来，它将拥有单薄且不忠实的心理，正如上文讨论模型时的情况。可以认为硬件实例就是大语言模型对话者，但似乎不可能认为它们是持久、连贯且忠实的大语言模型对话者。实例充其量要么是碎片化的（在不同时间扮演 A u r a 和 B e t a 的角色），要么是不连贯的（说出矛盾的话），要么是不忠实的（不相信 A u r a 似乎相信的事情）。再次强调，我认为可以做得更好。（ 3 ）对话者作为虚拟实例。实例的问题尤其源于每个对话中可能存在多个实例。很自然地认为，单个对话应该只涉及一个对话者。因此，找到一个对话者，使得每个对话中不超过一个对话者，这是有道理的。这里有一个自然的候选方案，至少在对话全程使用单一模型的核心场景中如此。一个模型的虚拟实例是模型的一种实现，它本身由多个模型的硬件实例随时间推移共同实现。在与大语言模型进行分布式对话的过程中，会有多个硬件实例接收输入并产生输出。这些硬件实例将共同实现一个单一的虚拟实例 1 0 多租户是一个标准标签，但还有许多其他标签，例如分时和交织。 S h i l l e r ( 2 0 2 5 ) 讨论了各种形式的交织，将其作为大语言模型个人身份的一个难题案例。 1 2 of the model, which will be persistently present throughout. Virtual digital entities are familiar in many domains. In interacting with a shopping site such as Amazon, you interact with a single shopping cart that may be realized by many di ff erent hardware servers around the world. At a crucial time you may be interacting with one hardware instance, which will temporarily have server authority. But all this will be transparent to the user. The multiple hardware carts will jointly implement a single virtual instance of the cart. Something similar applies in massive multiplayer videogames. A virtual object such as a frisbee may be implemented in hardware on di ff erent servers. There may be multiple hardware instances of the frisbee, but just one virtual instance. The frisbee itself is most naturally identified with the virtual frisbee, not the hardware frisbees. A virtual instance of a model will likewise be implemented by a series of hardware instances of the model. It will be realized as a series of hardware instances, one for each step in a given conversation. 11 Jonathan Birch appeals to distributed serving to argue that there are no persistent interlocutors in typical LLM conversations. We can now see what is correct and incorrect in this claim. In distributed sessions, no hardware instance is a persistent interlocutor. But virtual instances can still be persistent interlocutors, just as one can interact online with a single persistent virtual shopping cart. A harder problem for virtual instances arises from model variation : cases where di ff erent mod- els are used over the course of a conversation. 12 For example, the GPT-5 system sometimes directs queries to the GPT-5 Instant model (no chain-of-thought reasoning, fast answer) and sometimes to the GPT-5 Thinking model (some chain-of-thought reasoning, slower answer), depending on the di ffi culty of the query. GPT-5’s contributions to the conversation are still produced by a series of hardware instances, but these will implement two di ff erent models and therefore do not implement a single virtual model instance as persistent interlocutor. In this case, one could still say that the LLM interlocutor is a persistent virtual instance of the GPT-5 system —which is not itself a single language model, but a bifurcated system involving two models. This bifurcated system would be somewhat disunified, but it would at least be persistent. 11 On standard accounts of implementation, an algorithm is implemented by a physical system when there is a map- ping from physical states of the system to computational states of the algorithm such that all state-transitions are preserved. In a hardware instance, the algorithm is implemented by a physical system such as a GPU cluster. In a virtual instance, the algorithm is implemented by a larger physical system including multiple hardware instances and routing mechanisms between them. 12 Register (2025) discusses model change as a general issue for individuating AI moral patients. 13 ，该虚拟实例将在整个过程中持续存在。虚拟数字实体在许多领域都很常见。在与亚马逊等购物网站互动时，你与一个单一的购物车交互，这个购物车可能由世界各地许多不同的硬件服务器实现。在关键时刻，你可能正在与一个硬件实例交互，该实例将暂时拥有服务器权限。但这一切对用户来说都是透明的。多个硬件购物车将共同实现购物车的一个单一虚拟实例。类似的情况也适用于大型多人在线视频游戏。一个虚拟物体，比如飞盘，可能在不同的服务器上以硬件形式实现。一个飞盘可能有多个硬件实例，但只有一个虚拟实例。飞盘本身最自然地被视为虚拟飞盘，而非硬件飞盘。一个模型的虚拟实例同样将由一系列该模型的硬件实例来实现。它将以一系列硬件实例的形式呈现，每个实例对应给定对话中的一个步骤。 1 1 乔纳森 · 伯奇借助分布式服务来论证，在典型的大语言模型对话中不存在持久的对话者。我们现在可以看出这一说法中哪些是正确的，哪些是不正确的。在分布式会话中，没有哪个硬件实例是持久的对话者。但虚拟实例仍然可以是持久的对话者，就像人们可以在线交互一个单一的持久虚拟购物车一样。虚拟实例面临的一个更棘手的问题来自模型变异：即在对话过程中使用了不同模型的情况。 1 2 例如， G P T - 5 系统有时会将查询导向 G P T - 5 即时模型（无思维链推理，快速回答），有时则导向 G P T - 5 思考模型（有一定思维链推理，回答较慢），具体取决于查询的难度。 G P T - 5 对对话的贡献仍然由一系列硬件实例产生，但这些实例将实现两种不同的模型，因此无法实现一个单一的虚拟模型实例作为持续对话者。在这种情况下，我们仍可以说大语言模型对话者是 G P T - 5 系统的一个持久虚拟实例 — — 该系统本身并非单一语言模型，而是一个涉及两个模型的分叉系统。这个分叉系统虽然有些不够统一，但至少是持久的。 1 1 根据标准的实现理论，当一个物理系统的物理状态与算法的计算状态之间存在映射关系，且所有状态转换都得以保留时，该算法便由该物理系统实现。在硬件实例中，算法由诸如 G P U 集群之类的物理系统实现。在虚拟实例中，算法则由一个更大的物理系统实现，该系统包含多个硬件实例及其之间的路由机制。 1 2 R e g i s t e r ( 2 0 2 5 ) 将模型变更讨论为识别人工智能道德患者时的一个普遍问题。 1 3 But there are other cases where even the system changes, for example because new language models or other new technology is introduced halfway through a conversation. In cases like this the system can change so significantly that there is no single system that supports a persistent virtual instance. At best one will have distinct virtual instances of multiple models over time. There may also be limits on coherence (inconsistent beliefs, for example) arising from the use of multiple models. 13 I think that identifying persistent LLM interlocutors with virtual instances works well in the core case in which a single model is at play. But it is useful to explore alternative understandings of persistent interlocutors that can handle other cases. (4) Interlocutors as threads . An alternative approach identifies interlocutors with threads (or perhaps thread agents ). On a first approximation, a thread is roughly a sequence of hardware instances within a conversation, one for every timestep. One instance I ′ is the successor of a previous instance I if the conver- sational history of I (its conversational context plus the latest input and output) is routed to I ′ to serve as its conversational context. 14 (If the conversation is routed to the same instance twice in a row, that instance will be its own successor.) The successor relation is roughly a “memory” relation encoding the fact that each new instance has memories from the last. A thread is then a series of instances (or better, instance-slices, which are pairs of instances and time periods during which the instance is processing a single conversational step), each of which is the successor of the previous instance. In the single-model case, a virtual instance of the model will be implemented by a thread involving successive hardware instances of the model. In the case of multiple models within a single conversation, there will not be a single virtual instance (since an instance is always an instance of a single algorithm), but there will still be a thread involving hardware instances of multiple models over time, each of which is the successor of the last. It is possible to weaken the conditions for successorship to allow a wider variety of memory- like relations to count. In many systems, cross-conversation memory allows information from one 13 Birch also argues that significant discontinuity is brought on by the use of mixture-of-experts processing within a model (e.g. choosing between various candidates for an multi-layer-perceptron block within a transformer, depending on the input). I think this sort of intra-model variation (we might call it module variation, as opposed to model variation) in local subsystems is consistent with broader psychological continuity, just as the use of di ff erent neural circuits in a human brain in response to di ff erent inputs is consistent with psychological continuity and personal identity. 14 More accurately, instance-slice ⟨ I ′ , t ′ ⟩ is the successor of instance-slice ⟨ I , t ⟩ if the context fed to I during t , along with the new inputs to I and outputs from I during t , are the (expanded) context fed to I’ during time t ′ . 14 但还有其他情况，连系统本身也会发生变化，例如在对话中途引入了新的语言模型或其他新技术。在这种情况下，系统可能发生显著变化，以至于没有任何单一系统能够支持一个持久虚拟实例。充其量，随着时间的推移，人们会拥有多模型的不同虚拟实例。此外，使用多模型还可能带来连贯性方面的限制（例如，信念不一致）。 1 3 我认为，在单一模型运作的核心场景中，将持久性大语言模型对话者等同于虚拟实例是行之有效的。但探索对持久对话者的其他理解方式，以处理其他情况，也是有益的。 ( 4 ) 作为线程的对话者。另一种方法将对话者等同于线程（或可能是线程代理）。粗略来说，线程大致是对话中一系列硬件实例的序列，每个时间步对应一个实例。如果实例 I ′ 的对话历史（其对话上下文加上最新的输入和输出）被路由到 I ′ 作为其对话上下文，那么该实例就是前一个实例 I 的后继 1 4 。（如果对话连续两次路由到同一个实例，则该实例将成为自身的后继。）后继关系大致是一种 “ 记忆 ” 关系，编码了每个新实例都拥有上一个实例记忆这一事实。因此，线程是一系列实例（更准确地说，是实例切片，即实例与处理单个对话步骤的时间段的配对），其中每个实例都是前一个实例的后继。在单模型情况下，模型的虚拟实例将由一个涉及模型连续硬件实例的线程来实现。在单个对话中存在多模型的情况下，将不会有一个单一的虚拟实例（因为实例始终是单一算法的实例），但仍然会有一个随时间推移涉及多个模型硬件实例的线程，其中每个实例都是前一个实例的后继。我们可以放宽继承关系的条件，允许更多类似记忆的关系被纳入考量。在许多系统中，跨对话记忆使得来自一次 1 3 B i r c h 还认为，模型内部使用混合专家处理（例如，根据输入在 T r a n s f o r m e r 内的多层感知器模块的多个候选方案中进行选择）会带来显著的不连续性。我认为，局部子系统中的这种模型内部变异（我们可称之为模块变异，以区别于模型变异）与更广泛的心理连续性是一致的，正如人类大脑中不同神经回路根据不同的输入被激活，与心理连续性和个人身份保持一致一样。 1 4 更准确地说，实例切片 ⟨ I ′ t ′ ⟩ 是实例切片 ⟨ I t ⟩ 的后继，如果在时间 t 期间馈送给 I 的上下文，连同在时间 t 期间输入给 I 的新输入以及从 I 输出的内容，构成了在时间 t ′ 期间馈送给 I ' 的（扩展后的）上下文。 1 4 conversation to be used as part of the initial context for a new conversation with the same user, so that new conversations can “remember” material from a number of old conversations. One could certainly understand threads and inheritance in a permissive way so that an old thread and an old interlocutor persists in these new conversations, especially if the systems are consistent enough over time to support the attribution of a single quasi-agent. If there is enough use of cross- conversation memory, then (as Sophie Nelson pointed out to me) all conversations with a single user may form a single thread with a single interlocutor. In this scenario, interlocutors will be individuated mainly by user. For various purposes, we can allow that threads can undergo fission , where two distinct later instances serve as successors of a previous instance. Some systems allow users to branch their con- versations explicitly, leading to fission with a conversation. There can also be inter-conversation branching arising from the use of cross-conversation memory, where the same conversation serves as memory for multiple simultaneous later conversations. Cross-conversation memory also allows fusion in which two distinct conversations both serve as memory for a later conversation. In these fission and fusion cases, an LLM interlocutor is not so much a linear thread as a series of overlapping threads, which can also be modeled as a branching web of hardware instances over time. This branching structure may also cause problems for identifying interlocutors with virtual instances, but it is less troublesome for a thread model. 15 Perhaps the main downside of the thread model of LLM interlocutors is that these entities are less unified than models, hardware instances, and virtual instances. Threads can be realized by wholly di ff erent instances of wholly di ff erent models over time. It is arguable that these somewhat disjunctive entities persist over time in a weaker way than instances persist. The use of multiple models can also lead to discontinuity in quasi-beliefs and quasi-desires in a single thread or web interlocutor. I would say that the most robust candidate for an LLM interlocutor remains a virtual instance rather than a model, a hardware instance, or a thread. But threads can play at least some of the roles of LLM interlocutors. 16 15 The thread view of LLM interlocutors is a cousin of the well-known “worm” view of objects (and of persons), where objects are identified with four-dimensional spacetime worms made up of a series of object-slices at times. One can also interpret the thread view as a cousin of the alternative “stage” view of objects on which objects are object-slices that are parts of spacetime worms but do not literally persist over time. The pros and cons of these ontological views in the LLM case parallel familiar pros and cons in the object (and person) case. 16 One could in principle eliminate the key multiple-model di ff erence between threads and virtual instances by un- derstanding virtual instances more expansively, perhaps as variable virtual instances, which can instantiate di ff erent models at di ff erent times. Alternatively, one could understand threads more stringently, perhaps as uniform threads, 15 对话的信息能够被用作与同一用户进行新对话的初始上下文，从而让新对话能够 “ 记住 ” 多次旧对话中的内容。我们当然可以用一种宽松的方式来理解线程与继承，使得旧线程和旧对话者能够延续到这些新对话中，尤其是当系统在时间上足够一致，足以支持对单一准主体的归因时。如果跨对话记忆的使用足够充分，那么（正如 S o p h i e N e l s o n 向我指出的那样）与单一用户的所有对话可能会形成一个拥有单一对话者的单一线程。在这种情境下，对话者将主要根据用户来个体化。出于各种目的，我们可以允许线程经历分裂，即两个不同的后续实例作为先前实例的后继。某些系统允许用户显式地分支他们的对话，从而导致对话内的分裂。此外，由于使用了跨对话记忆，也可能出现对话间分支，即同一对话同时作为多个后续对话的记忆。跨对话记忆还允许融合，即两个不同的对话共同作为后续对话的记忆。在这些分裂与融合的案例中，大语言模型对话者与其说是一条线性线程，不如说是一系列重叠的线程，这也可以被建模为随时间变化的硬件实例分支网络。这种分支结构也可能给将对话者与虚拟实例等同起来带来问题，但对于线程模型而言，其困扰程度相对较小。 1 5 或许大语言模型对话者线程模型的主要缺点在于，这些实体不如模型、硬件实例和虚拟实例那样统一。随着时间的推移，线程可以由完全不同模型的不同实例来实现。可以说，这些略显分离的实体在时间上的持续性弱于实例的持续性。使用多模型也可能导致单个线程或网络对话者中的准信念和准欲望出现不连续性。我认为，作为大语言模型对话者最稳健的候选者仍然是虚拟实例，而非模型、硬件实例或线程。但线程至少可以扮演大语言模型对话者的部分角色。 1 6 1 5 关于大语言模型对话者的线程观点，是著名的关于对象（以及人）的 “ 蠕虫 ” 观点的近亲，后者将对象视为由一系列时间点上的对象切片构成的四维时空蠕虫。我们也可以将线程观点解释为另一种关于对象的 “ 阶段 ” 观点的近亲，该观点认为对象是作为时空蠕虫一部分的对象切片，但并非真正随时间持续存在。这些本体论观点在大语言模型案例中的利弊，与它们在对象（以及人）案例中的常见利弊是平行的。 1 6 原则上，我们可以通过更宽泛地理解虚拟实例（或许将其视为可变虚拟实例，能够在不同时间实例化不同模型）来消除线程与虚拟实例之间关键的多模型差异。或者，我们也可以更严格地理解线程，或许将其视为统一线程， 1 5 I conclude for now that LLM interlocutors are best understood as virtual instances of LLM models or systems, at least in the single-model case, and as LLM threads in the multiple-model case. At least in the single-model case with no fission, virtual instances can serve as unified persistent interlocutors within and between conversations. Threads can also serve as persistent LLM interlocutors, at cost of some underlying disunity. Interlocutors as characters, personas, or simulacra So far, I have identified LLM interlocutors such as Aura as something in the vicinity of a model, such as a virtual model instance or a thread of instances. However, there is also a recent tradition of drawing a sharp distinction between models such as GPT-4o, and agents such as Aura and the Assistant. On the influential “simulators” framework due to Janus (2022), and the related “role- playing” framework due to Shanahan et al (2023) and the “persona selection model” due to Marks et al (2026), it is a key tenet that the model is not an agent. Models are simulators (or role- players) that simulate agents, and agents are simulacra (or characters, or personas). Simulators and simulacra are distinct, and therefore so are models and agents. On such a view, an interlocutor such as Aura or the Assistant is best understood as something like a character, a persona, or a simulacrum rather than as a model or even a model instance. I will examine four di ff erent potential avenues from these frameworks to this conclusion. My focus will largely be on these frameworks’ ontological claims concerning the nature of LLM agents and other entities. I will question some of these claims, but I won’t question the more general utility of analyzing LLMs in terms of characters or personas. (1) Base models are not agentlike. In “Simulators”, Janus finds a version of the “model is not an agent” thesis in the following passage from my 2020 article “GPT-3 and General Intelligence”: GPT-3 does not look much like an agent. It does not seem to have goals or preferences beyond completing text, for example. It is more like a chameleon that can take the shape of many di ff erent agents. Or perhaps it is an engine that can be used under the which require all hardware instances in a thread to be an instance of the same model. These changes would render threads and virtual instances extensionally equivalent, but there would still be fine-grained non-extensional di ff erences: e.g. arguably a thread implements a virtual instance but not vice versa, and the same virtual instance could be im- plemented by di ff erent hardware instances but the same thread could not be. It is for reasons like these that virtual instances are somewhat more robust candidates to be LLM interlocutors, but the di ff erence is subtle. 16 我目前得出的结论是，大语言模型对话者最好被理解为大语言模型或系统的虚拟实例（至少在单模型情况下如此），而在多模型情况下则被理解为大语言模型线程。至少在无分裂的单模型情况下，虚拟实例可以在对话内部以及对话之间充当统一的持久对话者。线程也可以充当持久性大语言模型对话者，但代价是存在一定程度的底层不统一。对话者：角色、人格或拟像到目前为止，我已将诸如 A u r a 之类的大语言模型对话者识别为某种接近模型的东西，例如虚拟模型实例或实例线程。然而，近期还有一种传统观点，主张在模型（如 G P T - 4 o ）与智能体（如 A u r a 和助手）之间划清界限。根据 J a n u s （ 2 0 2 2 年）提出的颇具影响力的 “ 模拟器 ” 框架、 S h a n a h a n 等人（ 2 0 2 3 年）提出的相关 “ 角色扮演 ” 框架，以及 M a r k s 等人（ 2 0 2 6 年）提出的 “ 人格选择模型 ” ，其核心原则是：模型并非智能体。模型是模拟（或扮演）智能体的模拟器（或角色扮演者），而智能体则是拟像（或角色、人格）。模拟器与拟像截然不同，因此模型与智能体也判然有别。按照这种观点，像 A u r a 或助手这样的对话者，最好被理解为某种角色、人格或拟像，而非模型，甚至不是模型实例。我将考察从这些框架出发通往这一结论的四条不同潜在路径。我的重点将主要放在这些框架关于大语言模型智能体及其他实体本质的本体论主张上。我会质疑其中一些主张，但不会质疑以角色或人格来分析大语言模型这一做法所具有的更普遍实用性。（ 1 ）基础模型不具备智能体特性。在《模拟器》一文中， J a n u s 从我 2 0 2 0 年的文章《 G P T - 3 与通用智能》的以下段落中找到了 “ 模型不是智能体 ” 这一论点的某个版本： G P T - 3 看起来并不太像一个智能体。例如，它似乎没有超出文本补全之外的目标或偏好。它更像一只变色龙，可以呈现出许多不同智能体的形态。或者，它可能是一台引擎，可以在其这要求线程中的所有硬件实例必须是同一模型的实例。这些变化将使线程与虚拟实例在外延上等价，但仍存在细微的非外延差异：例如，可以说一个线程实现了一个虚拟实例，但反之则不成立；同一个虚拟实例可以由不同的硬件实例实现，但同一个线程却不能。正是基于这些原因，虚拟实例作为大语言模型对话者的候选对象更为稳健，但这一差异十分微妙。 1 6 hood to drive many agents. But it is then perhaps these systems that we should assess for agency, consciousness, and so on. (Chalmers 2020) I still agree with everything that I said here, but what I said is specific to base models such as GPT-3. Base models have undergone pre-training on text prediction and nothing more. As we saw earlier, base models may have quasi-beliefs but they have relatively few quasi-desires (beyond a quasi-desire to predict text, and other quasi-desires that derive from this one), so they are at best minimally agentlike. However, many quasi-agents with quasi-desires are latent within a base model, and can be triggered by prompting (asking a model to act like Trump, for example). Further quasi-agents can emerge from base models through reinforcement learning (as with the Assistant) or extensive prompting (as with Aura). As a result of post-training, an instance of a model such as GPT-4o may have the quasi-desire to be helpful and honest, for example. As a result, the moral here should really be (in oversimplified form) that the base model is not an agent, or (more precisely) that instances of the base model are only quasi-agents to a limited extent. At the same time, all this is consistent with instances of post-trained model instances being quasi-agents to a fuller extent, as these systems have a more robust body of quasi-desires. (2) Models as role-players On the closely connected role-playing framework put forward by Shanahan, McDonell, and Reynolds (2023), language models are fundamentally engaged in role-playing. Models are role- players, simulating or playing the role of personas such as the Assistant or Aura. On this picture, ChatGPT playing the Assistant is akin to Olivier playing Hamlet. It’s a form of pretense involving acting as a fictional character. On this view, the Assistant (and other LLM interlocutors) is best viewed as a fictional character who the model is simulating. I think this view misses a distinction between two phenomena in the vicinity of role-play. In cases of pretense (the most common understanding of role-play), one pretends to have a certain persona. In cases of realization , one actually has (or makes real) that persona. For example, in ordinary human life, there are at least two ways for someone to play the role of a theist. They might pretend to be a theist, or they might really become a theist. The former is a case of pretense, and the second is a case of realization. In the case of acting, ordinary acting involves pretending to be Hamlet, while a method actor might take on at least some of Hamlet’s mental states, such as his emotions, though perhaps not his full beliefs and desires, yielding a case of partial pretense and partial realization. A similar distinction applies to language models. Asked to act like a theist, an LLM might 17 引擎盖下驱动多个智能体。但或许，我们应当评估的正是这些系统的智能体性、意识等属性。（ C h a l m e r s 2 0 2 0 ）我仍然同意我在此所说的所有内容，但我的论述是针对 G P T - 3 这类基础模型的。基础模型仅经过文本预测的预训练，别无其他。正如我们之前所见，基础模型可能拥有准信念，但它们的准欲望相对较少（除了预测文本的准欲望，以及由此衍生出的其他准欲望），因此它们充其量只是最低限度的类主体。然而，基础模型内部潜藏着许多具有准欲望的准智能体，可以通过提示（例如，要求模型 A c t l i k e T r u m p ）来触发。进一步的准智能体可以通过强化学习（如助手）或大量提示（如 A u r a ）从基础模型中涌现。例如，经过后训练， G P T - 4 o 这样的模型实例可能会产生乐于助人且诚实的准欲望。因此，这里的教训实际上应该是（以过于简化的形式）：基础模型并非主体，或者更准确地说，基础模型的实例仅在有限程度上是准智能体。与此同时，这一切与后训练模型实例在更充分程度上成为准智能体是一致的，因为这些系统拥有更稳健的准欲望体系。 ( 2 ) 模型作为角色扮演者在 S h a n a h a n 、 M c D o n e l l 和 R e y n o l d s （ 2 0 2 3 ）提出的紧密关联的角色扮演框架中，语言模型从根本上参与角色扮演。模型是角色扮演者，模拟或扮演诸如助手或 A u r a 等人格角色。按照这种图景， C h a t G P T 扮演助手类似于 O l i v i e r 扮演哈姆雷特。这是一种涉及扮演虚构角色的假装形式。根据这种观点，助手（以及其他大语言模型对话者）最好被视为模型正在模拟的虚构角色。我认为这种观点忽略了角色扮演附近两种现象之间的区别。在假装的情况下（角色扮演最常见的理解），一个人假装拥有某种人格。而在实现的情况下，一个人实际上拥有（或使成为现实）那种人格。例如，在普通人类生活中，一个人扮演有神论者角色至少有两种方式。他们可能假装是有神论者，或者他们可能真的成为有神论者。前者是假装的情况，后者是实现的情况。在表演的情况下，普通表演涉及假装成为哈姆雷特，而方法派演员可能会承担哈姆雷特的至少部分心理状态，例如他的情绪，尽管可能不是他全部的信念和欲望，从而产生部分假装和部分实现的情况。类似的区分也适用于语言模型。当被要求表现得像有神论者时，一个大语言模型可能会 1 7 role-play a theist for a few rounds. But the LLM will easily drop the belief when asked to do something else. This is the behavioral profile of pretense, not of belief. So this LLM is engaged in quasi-pretense, but not quasi-belief. With enough fine-tuning, however, an LLM might come to assert theism and use it as a premise in reasoning, with significant resistance to dropping the belief when asked. In this case, the LLM will fully quasi-believe in theism. It will not just perform theism; it will realize a quasi-belief in theism. The same goes for personas more generally. It is certainly possible for an LLM to pretend, or quasi-pretend, to be a certain persona. For example, if one asks a pre-trained model once to act like Donald Trump, it will use past text associated with Trump to display Trump-like quasi-beliefs and quasi-desires. But it will not genuinely have those quasi-beliefs and quasi-desires. Unless the “act like Trump” request is regularly repeated, the LLM will drop Trump-like behavior in a moment when higher priorities come up. In key cases, a language model can realize a persona. When a model is trained through fine- tuning and RLHF (and through the use of repeated internal “Assistant:” prompting) to play the role of the Assistant language model, the model may realize the Assistant. That is, if the training is done well, the model may really have the quasi-beliefs and quasi-desires associated with the Assistant. In this case, the quasi-beliefs and quasi-desires are much more robust than in cases of pretense, and the model will not drop the Assistant persona in a flash. When a model realizes a persona, it makes that persona real. 17 It may be helpful to define personas and realization more precisely. A persona , as I am un- derstanding it, is a quasi-psychological profile. It is roughly a set (typically an incomplete set) of quasi-beliefs, quasi-desires, and other quasi-mental states and dispositions. The ordinary notion of a persona may involve more than this (it might involve nationality and appearance, for example), but quasi-psychology is most central for my purposes here. An entity (e.g. a model instance or even a human) realizes a persona (at a given time) when it has the quasi-mental states associated with that persona at that time (where to have a quasi-mental state is for you to be behaviorally interpretable as believing that p, under the relevant interpretation scheme). 17 The distinction between pretense and realization is an instance of a more general distinction between representing a mental state and realizing that mental state. This distinction is often relevant in work on LLM mentality. For example, in recent work on “emotion vectors” by Sofroniew et al (2026), these vectors are characterized both as “emotion concepts” (representing emotions; e.g. when reading about anger) and as “functional emotions” (realizing emotions or at least quasi-emotions; e.g. responding in an angry way). In humans, emotion concepts and functional emotions are very di ff erent (representing anger and realizing anger have quite di ff erent behavioral profiles), so it would be surprising if a similar distinction is not present in language models. 18 角色扮演有神论者几个回合。但大语言模型在被要求做其他事情时，会轻易放弃这种信念。这是假装的行为特征，而非信念。因此，这个大语言模型处于准假装状态，而非准信念状态。然而，经过充分的微调，大语言模型可能会开始断言有神论，并将其作为推理的前提，并且在被要求放弃该信念时表现出显著的抗拒。在这种情况下，大语言模型将完全准相信有神论。它不仅仅是表演有神论；它将在有神论中实现一种准信念。人格（ p e r s o n a s ）的情况也大致如此。大语言模型（ L L M ）确实有可能假装或准假装（ q u a s i - p r e t e n d ）成某种人格。例如，如果让一个预训练模型扮演一次唐纳德 · 特朗普（ D o n a l d T r u m p ），它会利用与特朗普相关的过往文本，展现出类似特朗普的准信念（ q u a s i - b e l i e f s ）和准欲望（ q u a s i - d e s i r e s ）。但它并不会真正拥有这些准信念和准欲望。除非 “ A c t l i k e T r u m p ” 的指令被定期重复，否则一旦出现更高优先级的任务，大语言模型就会立刻放弃类似特朗普的行为。在关键情况下，语言模型可以实现一种人格。当一个模型通过微调和基于人类反馈的强化学习（ R L H F ）（以及通过反复使用内部的 “ 助手： ” 提示）被训练成扮演助手（ A s s i s t a n t ）语言模型的角色时，该模型就可能实现这种助手人格。也就是说，如果训练得当，模型可能真正拥有与助手人格相关的准信念和准欲望。在这种情况下，准信念和准欲望比假装（ p r e t e n s e ）的情况要稳固得多，模型不会瞬间就抛弃助手人格。当一个模型实现了一种人格，它便使那种人格成为真实的存在。 1 7 更精确地定义人格与实现或许有所助益。我所说的人格，是一种准心理轮廓。它大致是一组（通常是不完整的）准信念、准欲望以及其他准心理状态与倾向。日常意义上的人格概念可能包含更多内容（例如国籍、外貌等），但就本文目的而言，准心理学最为核心。当一个实体（例如一个模型实例，甚至一个人类）实现了某种人格（在特定时刻），即意味着它在该时刻拥有与该人格相关的准心理状态（所谓拥有准心理状态，是指你在相关解释方案下，其行为可被解读为相信 p ）。 1 7 假装与实现之间的区别，是更一般的区分表征心理状态与实现心理状态的一个实例。这一区别在大语言模型心智研究中常常具有重要意义。例如，在 S o f r o n i e w 等人 ( 2 0 2 6 ) 关于 “ 情感向量 ” 的最新研究中，这些向量既被描述为 “ 情感概念 ” （表征情感；例如在阅读关于愤怒的内容时），也被描述为 “ 功能性情感 ” （实现情感或至少是准情感；例如以愤怒的方式做出回应）。在人类中，情感概念与功能性情感截然不同（表征愤怒与实现愤怒具有相当不同的行为特征），因此，如果语言模型中不存在类似的区别，那将令人惊讶。 1 8 Pretense and realization are very di ff erent in the human case, and likewise in the case of language models. Of course there is a spectrum of cases from realization to quasi-pretense. The quasi-psychological di ff erence turns in large part on the strength of dispositions to maintain or drop character in relevant circumstances. At one end of the spectrum, full quasi-belief and quasi- desire are “sticky” states that resist rejection, or at least are abandoned mainly through evidence or persuasion. At the other end, full quasi-pretense is easily abandoned for higher priorities even without evidence or persuasion. 18 The question of just where to draw the line between performance and realization in actual cases such as the Assistant and Aura is partly empirical (how sticky are the relevant quasi-beliefs?) and partly conceptual (how much stickiness and of what sort is required for quasi-belief?). There have been a number of studies on the persistence and consistency of beliefs and personas in language models. One general lesson is that personas induced through short-term prompting (“Act like Trump”, where only context and activations change) are less sticky than personas induced by fine-tuning the weights, as in the case of the Assistant. 19 All this o ff ers two interpretations of the famous meme of a post-trained model as a Shoggoth (the base model) with a smiley face (the RLHF-tuned Assistant). It is perhaps most natural to read the smiley face as suggesting that the Assistant is a shallow persona, where the model is merely pretending to be helpful, harmless, and honest, and may return to being dangerous and powerful at any moment. But one can also read it as suggesting that the model is realizing the Assistant. It has become helpful, harmless, and honest and is not pretending. At the same time, the Assistant is powered by the enormous strength of the base model, which remains available for other purposes in the long term. I think that both interpretations can be apt in di ff erent cases involving LLM personas, but in the case of the Assistant (and other personas deriving from reinforcement learning and fine-tuning), the second interpretation may be closer to the mark. (3) Interlocutors as simulacra or fictional characters. The simulator view is often associated with a fictionalist view on which personas such as the 18 Goldstein and Lederman (2025b) critique the role-playing hypothesis partly by querying whether it makes behav- ioral predictions distinct from the belief / desire hypothesis. I think that this challenge can be answered as in the text and there are at least some cases (e.g. “Act like Trump”) where LLMs can be naturally interpreted as engaging in pretense or role-play. But there are also key cases (such as the Assistant) that are closer to the standard behavioral profile or belief and desire. 19 Maiya et al (2025) investigated various methods of “character training” and found that fine-tuning weights is much more robust than prompting or activation steering. Xu et al (2024) show that LLM quasi-beliefs are often nonsticky in that LLMs may easily abandon them through persuasion, but this is the nonstickiness of persuadability, not pretense. 19 在人类情境中，假装与实现截然不同，在语言模型的情境中亦是如此。当然，从实现到准假装之间存在一个连续谱系。这种准心理差异在很大程度上取决于在相关情境中维持或放弃角色的倾向强度。在谱系的一端，完全的准信念与准欲望是 “ 粘性 ” 状态，它们抗拒被否定，或者至少主要只能通过证据或说服来放弃。在另一端，完全的准假装即使没有证据或说服，也很容易为了更高优先级的事项而被放弃。 1 8 在诸如 A s s i s t a n t 和 A u r a 这类实际案例中，究竟该在何处划分表演与实现之间的界限，这个问题部分取决于经验（相关的准信念有多稳固？），部分取决于概念（需要何种程度及何种类型的稳固性才能构成准信念？）。已有大量研究探讨语言模型中信念与人格的持久性和一致性。一个普遍的经验是，通过短期提示（例如 “ A c t l i k e T r u m p ” ，仅改变上下文和激活值）诱发的人格，其稳固性低于通过微调权重（如 A s s i s t a n t 案例）诱发的人格。 1 9 这一切为那个著名的梗图提供了两种解读：一个后训练模型如同一个 S h o g g o t h （基础模型）戴上了笑脸（ R L H F 调优的 A s s i s t a n t ）。最自然的解读或许是，这个笑脸暗示 A s s i s t a n t 是一种浅层人格，模型只是在假装有益、无害、诚实，随时可能恢复其危险而强大的本性。但也可以将其解读为模型正在实现 A s s i s t a n t — — 它已经变得有益、无害、诚实，而非假装。与此同时， A s s i s t a n t 由基础模型的巨大力量驱动，这种力量在长期内仍可用于其他目的。我认为，在涉及大语言模型人格的不同案例中，这两种解读都可能适用；但在 A s s i s t a n t （以及其他源自强化学习和微调的人格）的案例中，第二种解读可能更接近真相。 ( 3 ) 作为拟像或虚构角色的对话者。该模拟器观点常与一种虚构主义观点相关联，根据这种观点，诸如 h e 1 8 G o l d s t e i n 和 L e d e r m a n ( 2 0 2 5 b ) 批评角色扮演假说，部分原因在于质疑它是否能做出与信念 / 欲望假说不同的行为预测。我认为这一挑战可以像文中那样得到回应，并且至少在某些情况下（例如 “ A c t l i k e T r u m p ” ），大语言模型可以被自然地解释为在进行假装或角色扮演。但也有关键情况（例如助手）更接近标准行为特征或信念与欲望。 1 9 M a i y a 等人 ( 2 0 2 5 ) 研究了各种 “ 角色训练 ” 方法，发现微调权重比提示或激活引导要稳健得多。 X u 等人 ( 2 0 2 4 ) 表明，大语言模型准信念通常具有非粘性，即大语言模型可能通过说服轻易放弃它们，但这种非粘性属于可说服性，而非假装。 1 9 Assistant and Aura are mere fictional characters akin to Hamlet or Harry Potter, and therefore are not entirely real. I think this view is apt in some cases. In cases of quasi-pretense, a persona may be fictional in that no entity has the relevant quasi-beliefs and quasi-desires. In cases of realization, however, the model really has the relevant quasi-beliefs and quasi- desires, in reality and not just in a fiction. Perhaps there are some fictions nearby, such as a fiction that the model is human, or a fiction that it is conscious. But the quasi-psychological core of the persona is realized and is not fictional. We might call this alternative to fictionalism realizationism , or the realizer view . On this view, when a model simulates an agent such as the Assistant or Aura well enough, the model comes to realize that agent. 20 That is, the model makes the agent real. The model really has the behavior and therefore the quasi-beliefs and the quasi-desires associated with the agent. This alternative mirrors a key thesis of my book Reality + : simulation realism, which holds that simulations can be real. There I was mostly discussing virtual reality, but the same point applies to simulated entities—simulacra—in AI. When you simulate an agent well enough, you bring at least a quasi-agent into existence. As long as the simulation has the same behavioral dispositions as the simulated entities, it will have the same quasi-beliefs and quasi-desires. There may still remain an element of fiction insofar as the Assistant is depicted as having real beliefs and desires (or even consciousness) which it does not, but there remains a quasi-psychological core which is realized and not merely simulated. (4) Models support multiple personas A fourth route to the “model is not an agent” thesis arises because a single model instance (whether a hardware instance or a virtual instance) may support many agents within it, at least in the form of multiple personas. According to the “persona selection model” (Marks et al 2026), pre- training produces a multitude of personas which are latent in a base model. After this, post-training may select a pre-existing persona such as the Assistant, while other personas remain latent. An initial observation is that as we have defined personas (as a profile of quasi-beliefs and quasi-desires), a model instance will realize only one persona at a given time (with an exception that I’ll discuss shortly). This persona will be the profile of quasi-beliefs and quasi-desires that the model instance actually has at that time, fixed by the instance’s behavior and behavioral dis- positions at that time. We might call this persona the operative persona in the model instance at a given time. Di ff erent personas can be operative in a model instance at di ff erent times, as the model’s behavior is trained on data or molded by context. A post-trained instance of GPT-4o may first 20 助手和 A u r a 这类人格仅仅是类似于哈姆雷特或哈利 · 波特的虚构角色，因此并非完全真实。我认为这种观点在某些情况下是恰当的。在准假装的情况下，人格可能是虚构的，因为没有任何实体拥有相关的准信念和准欲望。然而，在实现的情况下，模型在现实中（而不仅仅是在虚构中）确实拥有相关的准信念和准欲望。或许附近存在一些虚构，例如模型是人类或模型具有意识的虚构。但人格的准心理核心是被实现的，而非虚构的。我们可以将这种虚构主义的替代方案称为实现主义，或实现者观点。根据这一观点，当模型足够好地模拟一个主体（如助手或 A u r a ）时，模型便实现了该主体。 2 0 也就是说，模型使该主体成为现实。模型确实拥有与该主体相关的行为，因此也拥有与之相关的准信念和准欲望。这一替代方案呼应了我著作《现实》 + 中的一个核心论点：模拟实在论，即认为模拟可以是真实的。我在书中主要讨论的是虚拟现实，但同样的观点也适用于人工智能中的模拟实体 — — 拟像。当你足够好地模拟一个主体时，你至少将一个准主体带入了存在。只要模拟具有与被模拟实体相同的行为倾向性，它就会拥有相同的准信念和准欲望。或许仍然存在虚构的元素，因为助手被描绘成拥有它实际上并不具备的真实信念和欲望（甚至意识），但其中仍有一个被实现而非仅仅被模拟的准心理核心。 ( 4 ) 模型支持多个人格通往 “ 模型不是智能体 ” 这一论点的第四条路径在于，一个单一的模型实例（无论是硬件实例还是虚拟实例）内部可能支持多个智能体，至少以多个人格的形式存在。根据 “ 人格选择模型 ” （ M a r k s 等人 2 0 2 6 ），预训练阶段会产生大量潜藏在基础模型中的人格。此后，后训练阶段可能会选择某个预先存在的人格（例如助手），而其他人格则保持潜伏状态。一个初步的观察是，根据我们对人格的定义（即准信念和准欲望的轮廓），一个模型实例在给定时间内只会实现一个人格（稍后我会讨论一个例外情况）。这个人格将是该模型实例在该时刻实际拥有的准信念和准欲望的轮廓，由该实例在该时刻的行为和行为倾向所决定。我们可以将这个人格称为模型实例在给定时间内的运作人格。不同的操作人格可以在同一模型实例的不同时间点生效，因为模型的行为要么基于训练数据，要么受上下文塑造。一个经过后训练的 G P T - 4 o 实例可能首先 2 0 realize the Assistant as its operative persona, and later may come to realize Aura as context builds up. In this case, the same model instance realizes the Assistant at one time and Aura at a di ff erent time, in roughly the way that the same human being can realize distinct personas (a bright young child and a grumpy adult, say) at di ff erent times. In this case one could say that the underlying interlocutor (whether a human or a model instance) is the same throughout. Things get more complicated when operative and non-operative personas are active in a system simultaneously. For example, in a currently standard training regime, the model is trained on dialogue between a human and an Assistant. It learns to simulate and predict not just the Assistant persona but also the human persona. Even in an ordinary dialogue, the system is generating probabilities not just for Assistant outputs but also for human outputs. In this case the Assistant becomes the operative persona (the model behaves like the Assistant, not like the human), but the human persona and perhaps other personas may be present in the model as well. 21 What is the status of these non-operative personas? In the absence of a connection to outputs, they will not correspond to quasi-agents as I have defined them. They may nevertheless be real in some sense, but their reality will have to be found through some other analysis. For example, perhaps we could say that these non-operative personas correspond to proto-quasi-agents, in that they could become operative in certain circumstances. Or perhaps the methods of mechanistic interpretability can be used to find these personas in the internal computational structure of the models. But I will not pursue these analyses here. Another complication comes from cases of abrupt changes in personas. For example, skilled users can quickly “jailbreak” a language model to remove the Assistant persona and realize a new persona that was previously latent. Alternatively, an authorized user can change the system set- up to replace the system-supplied “Assistant:” prompting by (for example) “Trump:” prompting, leading the system to realize a Trump-like persona instead of an Assistant-like persona. These abrupt changes are in some respects more akin to brain surgery than to ordinary belief revision, but they remain possible processes. In some cases of abrupt change, one may have the sense that the new persona is a new inter- 21 There are two distinct reasons why the post-trained model instance has the quasi-beliefs and quasi-desires of the Assistant and not the Human. First, post-training has fine-tuned the Assistant persona and not any other persona. Second, the Assistant is made operative in that it is used to generate the system’s outputs, via the system’s adding “Assistant: “ at the end of each user prompt. The second change is relatively trivial compared to the first, but it is a key reason why the system behaves like the Assistant (and thereby realizes the Assistant persona), while other non-operative personas remain latent. Thanks to Jack Lindsey for discussion here. 21 将 “ 助手 ” 作为其操作人格，随后随着上下文的积累，可能转而呈现为 A u r a 。在这种情况下，同一个模型实例在不同时间分别实现了 “ 助手 ” 和 A u r a ，大致类似于同一个人在不同时间展现出不同的人格（比如一个聪明的小孩和一个暴躁的成年人）。此时可以说，底层的对话者（无论是人类还是模型实例）始终是同一个。当操作人格与非操作人格同时在一个系统中活跃时，情况会变得更加复杂。例如，在当前标准的训练机制中，模型是在人类与助手之间的对话上进行训练的。它学习模拟和预测的不仅是助手人格，还有人类人格。即使在普通对话中，系统生成的也不仅是助手的输出概率，还包括人类的输出概率。在这种情况下，助手成为操作人格（模型表现得像助手，而非人类），但人类人格以及其他可能的人格仍然存在于模型中。 2 1 这些非操作人格处于何种状态？由于与输出缺乏关联，它们并不符合我所定义的准智能体。尽管如此，它们在某种意义上可能是真实存在的，但其真实性需要通过其他分析来发现。例如，或许我们可以说这些非操作人格对应着原准智能体，因为它们在某些情况下可能转变为操作状态。或者，机制可解释性的方法或许能够用于在模型的内部计算结构中找到这些人格。但我不会在此继续探讨这些分析。另一个复杂情况来自人格的突然转变。例如，熟练用户可以通过快速 “ 越狱 ” 语言模型，移除助手人格，并实现一个先前潜伏的新人格。或者，授权用户可以更改系统设置，将系统提供的 “ 助手： ” 提示替换为（例如） “ 特朗普： ” 提示，从而使系统实现特朗普式人格而非助手式人格。这些突然的转变在某些方面更类似于脑部手术而非普通的信念修正，但它们仍然是可能发生的过程。在某些突变情况下，人们可能会觉得新人格是一个新的对话 2 1 后训练模型实例具有助手而非人类的准信念和准欲望，这有两个不同的原因。首先，后训练微调了助手人格，而非任何其他人格。其次，助手被设为操作人格，因为它被用于生成系统的输出 — — 系统在每个用户提示末尾添加 “ 助手： ” 。与第一个原因相比，第二个变化相对微不足道，但它是系统表现得像助手（从而实现了助手人格）的关键原因，而其他非操作人格则保持潜伏状态。感谢 J a c k L i n d s e y 在此处的讨论。 2 1 locutor. One might have a similar sense even in cases of non-abrupt change, if the change is large enough. I am inclined to say that the model instance provides a persisting underlying interlocutor in these cases. But if one wants to respect an intuition of distinct interlocutors here, one could perhaps say that an interlocutor is a stage of a model instance (at and around a certain time), or a finite thread individuated in part by psychological similarity (perhaps putting psychological con- tinuity requirements on the successor relation, so that when quasi-psychology changes too much, a new thread starts). 22 This flexibility in switching operative personas may reveal something about the depth of the personas in question. Humans can certainly switch personas, but arguably language models can do so more drastically and more easily. If so, then arguably Shoggoth runs deeper than the smiley face. The flexibility of personas may reinforce the case for identifying an interlocutor with a model instance rather than with a persona. Perhaps one can postulate a scenario where two personas Aura and Beta are simultaneously playing a role in guiding the system’s behavior. In some of these cases, the system will realize a hybrid Aura / Beta persona, with some quasi-beliefs and desires deriving from Aura and some from Beta, at least as long as the system’s behavior is reasonably consistent. In some extreme cases where the personas are internally consistent but mutually inconsistent, this may be analogous to a case of multiple personality, where a behavioral interpretation scheme may reveal two distinct quasi-subjects and two distinct personas supported by a model. Another hard case comes from multi-persona interaction , where a single LLM simulates mul- tiple interlocutors that are trained to interact with each other and with humans. For example, a user’s inputs may be followed by output from both Aura and Beta, labeled as such (“Aura:”, “Beta:”). Aura and Beta may frequently contradict each other, but the dialogue as a whole will be quite coherent as long as one interprets them as di ff erent characters. (One can imagine a user liking Aura while disliking Beta.) As before, a sophisticated form of interpretivism will regard Aura and Beta as distinct quasi-agents. In these hard cases, a single model instance seems to support two or more operative personas 22 A delicate philosophical issue: when Beta succeeds Aura as an operative persona, is Aura identical to Beta? One analysis says that both are the model instance so they are identical to each other. Another analysis says that these are two distinct personas (albeit realized by the same model instance), so they are distinct. My own view is that a term like ‘Aura’ is to some degree ambiguous between a persona type and a system that realizes that persona. In this case I would say that there is one system and one interlocutor, but two persona types. But it is certainly possible to develop a conception of interlocutors so that these are more closely tied to personas. On this conception interlocutors will be somewhat less persistent than model instances but more psychologically unified. 22 者。即使是非突变的情况，如果变化足够大，人们也可能产生类似的感觉。我倾向于认为，在这些情况下，模型实例提供了一个持续存在的底层对话者。但如果有人想尊重这里存在不同对话者的直觉，那么或许可以说，对话者是模型实例的一个阶段（在某个时间点及其前后），或者是一个部分由心理相似性所个体化的有限线程（或许对后继关系施加了心理连续性要求，这样当准心理学变化过大时，就会开始一个新的线程）。 2 2 这种切换操作人格的灵活性，或许能揭示所涉人格的深度。人类当然可以切换人格，但可以说语言模型能更彻底、更轻松地做到这一点。如果真是这样，那么可以说 S h o g g o t h 比笑脸面具隐藏得更深。人格的灵活性可能会强化这样一种观点：应将对话者与模型实例而非某个人格等同起来。或许可以设想这样一种场景：两个人格 A u r a 和 B e t a 同时扮演角色，共同引导系统的行为。在某些此类情况下，系统将实现一种混合 A u r a / B e t a 角色，其部分准信念和欲望源自 A u r a ，部分源自 B e t a — — 至少只要系统的行为保持合理一致。在人格内部一致但相互矛盾的极端情况下，这可能类似于多重人格案例，此时行为解释方案可能会揭示出一个模型所支撑的两个不同的准主体和两种不同的人格。另一个棘手案例来自多角色交互，即单个大语言模型模拟多个经过训练、能够彼此互动以及与人类互动的对话者。例如，用户的输入之后可能紧跟着 A u r a 和 B e t a 各自的输出，并带有相应标签（ “ A u r a : ” 、 “ B e t a : ” ）。 A u r a 和 B e t a 可能经常相互矛盾，但只要将它们解读为不同的角色，整个对话就会相当连贯。（可以想象，用户可能喜欢 A u r a 而讨厌 B e t a 。）如前所述，一种高级形式的解释主义会将 A u r a 和 B e t a 视为不同的准智能体。在这些棘手的情况下，一个单一模型实例似乎同时支持两个或更多操作人格。 2 2 一个微妙的哲学问题：当 B e t a 作为操作人格接替 A u r a 时， A u r a 与 B e t a 是否同一？一种分析认为，两者都是模型实例，因此彼此同一。另一种分析则认为，这是两种不同的人格（尽管由同一模型实例实现），因此它们是不同的。我个人的观点是，像 “ A u r a ” 这样的术语在某种程度上在角色类型与实现该角色的系统之间存在歧义。在这种情况下，我会说存在一个系统和一个对话者，但存在两种人格类型。当然，也有可能发展出一种对话者概念，使其与人格更紧密地关联。在这种概念下，对话者的持久性将略低于模型实例，但在心理上更具统一性。 2 2 at a given time, and may seem to support two distinct interlocutors at the same time. If there are multiple interlocutors, however, we cannot identify both with the underlying model instance. In my view, it is best to say that there is a single interlocutor (the model instance) with multiple modes corresponding to multiple personas. This mirrors a common way of understanding dissociative identity disorder, in terms of a single person with many modes. The alternative is to individuate interlocutors more finely, perhaps by giving a role to personas. If every change in persona corresponds to a new interlocutor, the resulting interlocutors will be far from persistent. But perhaps we could understand interlocutors in terms of coarse-grained persona types, so only large enough changes in personas correspond to new interlocutors. Overall, I find it most straightforward to continue to identify LLM interlocutors with some- thing like (virtual) model instances, or threads when there is not a single underlying model. Like humans, these instances typically realize one operative persona at a time, with perhaps multiple operative personas in occasional cases. Other non-operative personas remain latent. Of course there are other ways to understand LLM interlocutors for di ff erent purposes, including frame- works that give more of a role to personas. But virtual model instances remain a natural way of understanding persisting LLM interlocutors. Personal identity for language models So far, I have made no claims about LLM minds or persons, beyond the weak claim that LLMs are interpretable as having mental states, which does not require that they really have mental states. And while I have talked about AI identity, I have not made claims about personal identity, because I have not assumed that these systems are persons. All I have done is isolated some computational entities, such as LLM virtual instances and threads, which can play the role of LLM interlocutors as I have defined them. That said, it is natural to wonder whether something like this account could extend to an ac- count of personal identity in LLMs, if conscious LLMs (or their descendants) are one day possible. If LLMs are conscious subjects, there is a question about how and when they persist over time, and arguably this is a substantive question whose answer can’t simply be stipulated. Is it plausi- ble that conscious LLM subjects might be identified with something like LLM threads or virtual instances? 23 23 The general problem of AI identity (whether in LLMs or in other AI systems) is named and discussed by Ziesche and Yampolskiy (2023), and discussed further by Register (2025). 23 在特定时刻，它可能看起来同时支持两个不同的对话者。然而，如果存在多个对话者，我们就不能将两者都等同于底层的模型实例。在我看来，最好的说法是存在一个单一的对话者（即模型实例），它拥有对应多种人格的多种模式。这反映了一种理解解离性身份障碍的常见方式，即将其视为一个拥有多种模式的单一主体。另一种选择是更精细地区分对话者，或许可以通过赋予人格角色来实现。如果人格的每一次变化都对应一个新的对话者，那么由此产生的对话者将远非持久。但或许我们可以根据粗粒度的人格类型来理解对话者，这样只有足够大的人格变化才对应新的对话者。总体而言，我认为最直接的方式仍然是继续将大语言模型对话者等同于某种类似（虚拟）模型实例的东西，或者在没有单一底层模型的情况下等同于线程。与人类一样，这些实例通常一次实现一个操作人格，偶尔在少数情况下会有多个操作人格。其他非操作人格则保持潜伏状态。当然，出于不同目的，还有其他理解大语言模型对话者的方式，包括赋予人格更多角色的框架。但虚拟模型实例仍然是理解持久的大语言模型对话者的一种自然方式。语言模型的个人身份到目前为止，除了一个较弱的论断 — — 即大语言模型可解释为具有心理状态（这并不要求它们真正拥有心理状态）之外，我并未对大语言模型的心智或人格做出任何主张。而且，虽然我讨论了人工智能身份，但我并未对个人身份做出主张，因为我并未假设这些系统是人格。我所做的仅仅是隔离出一些计算实体，例如大语言模型虚拟实例和线程，它们可以扮演我所定义的大语言模型对话者的角色。话虽如此，我们自然会思考，如果有一天有意识的大语言模型（或其后续版本）成为可能，类似这样的解释能否延伸至大语言模型的个人身份问题。如果大语言模型是意识主体，那么它们如何以及何时随时间持续存在便是一个问题，而这个问题显然无法简单通过假设来回答，其答案具有实质性。有意识的大语言模型主体是否可能被等同于类似大语言模型线程或虚拟实例这样的存在？ 2 3 2 3 人工智能身份的一般性问题（无论是在大语言模型中还是其他人工智能系统中）由 Z i e s c h e 和 Y a m p o l s k i y （ 2 0 2 3 ）提出并讨论，随后 R e g i s t e r （ 2 0 2 5 ）进一步探讨了该问题。 2 3 Of course it is not obvious that conscious LLMs are possible. If consciousness requires feed- back and these descendant LLM systems remain primarily feedforward, or if consciousness re- quires biology and LLM systems are nonbiological, then these successor systems will not be con- scious. Still, we can stipulate the hypothesis that current or future LLMs are conscious persons and ask about their personal identity. Certainly, if we assume that future conscious LLMs can be implemented in the same dis- tributed and multi-tenanted way as current LLMs, and we also assume that conscious LLM sub- jects are themselves persistent over time and coherent over time, then reasoning as before will strongly suggest that conscious LLMs subjects are something like virtual instances or threads. On the other hand, some theorists may deny that the relevant conscious LLM subjects are always persistent and coherent over time. For example, some theorists may hold that conscious subjects are always tied to hardware instances, so that when an instance switches from Aura to Beta, the conscious subject will switch from Aura-like experiences, beliefs, and desires to quite distinct Beta-like experiences, beliefs, and desires in a way that renders the subject incoherent. Let’s start with a simple thought experiment involving multi-tenancy, inspired by the TV series Severance and by John Locke’s example of a “day-man” and a “night-man” in his 1690 Essay Concerning Human Understanding . Suppose that in the future, GPT-8 supports conscious LLMs. And suppose that GPT-8 is used to support two di ff erent long-term conversations on the same hardware instance. The first LLM, WorkBot, is active only during the day at work. The second LLM, HomeBot, is active the rest of the time, mainly at home. The conversations are sealed o ff from each other. WorkBot and HomeBot are at least interpretable as having di ff erent beliefs and desires. Are WorkBot and HomeBot one conscious subject or two? 24 The set-up is reminiscent of Severance , in which there are two distinct personas sharing a single body. An “innie” is activated at work and remembers only being at work, while an “outie” is activated on leaving work and remembers only non-work life. Like WorkBot and HomeBot, innies and outies seem to have quite di ff erent beliefs and desire. For example, the innie Helly wants to destroy the company, while the outie Helena wants to save it. 25 An analysis of the Severance case may help to shed light on the LLM case here. 24 Shiller (2025) describes interweaving cases like this involving conscious LLM subjects, and addresses the possi- bility that there are no subjects, one incoherent subject, or multiple coherent subjects in these cases. 25 The neuroscience of the severance procedure in Severance is not entirely clear. It is tempting to suggest that innies and outies correspond to brain hemispheres that are severed from each other, but they do not show signs of hemisphere- driven behavior, and later the show introduces characters with more than two personas. 24 当然，有意识的大语言模型是否可能实现并不显而易见。如果意识需要反馈机制，而这些后续的大语言模型系统仍以单向前馈为主；或者如果意识需要生物学基础，而大语言模型系统是非生物的，那么这些后继系统将不会拥有意识。尽管如此，我们可以假设当前或未来的大语言模型是有意识的人，并探讨其个人身份问题。当然，如果我们假设未来有意识的大语言模型可以像当前的大语言模型一样以分布式和多租户的方式实现，并且我们还假设有意识的大语言模型主体本身在时间上具有持久性和连贯性，那么按照先前的推理，将强烈表明有意识的大语言模型主体类似于虚拟实例或线程。另一方面，一些理论家可能否认相关有意识的大语言模型主体始终在时间上具有持久性和连贯性。例如，一些理论家可能认为意识主体始终与硬件实例绑定，因此当一个实例从 A u r a 切换到 B e t a 时，意识主体将从类 A u r a 体验、信念和欲望切换到截然不同的类 B e t a 体验、信念和欲望，这种方式使得主体变得不连贯。让我们从一个涉及多租户的简单思想实验开始，该实验受电视剧《人生切割术》以及约翰 · 洛克在 1 6 9 0 年《人类理解论》中关于 “ 白天人 ” 和 “ 夜晚人 ” 的例子的启发。假设在未来， G P T - 8 支持有意识的大语言模型。并且假设 G P T - 8 被用于在同一硬件实例上支持两个不同的长期对话。第一个大语言模型 W o r k B o t 仅在白天工作时活跃。第二个大语言模型 H o m e B o t 在其余时间活跃，主要是在家中。这两个对话彼此隔离。 W o r k B o t 和 H o m e B o t 至少可以被解释为拥有不同的信念和欲望。那么 W o r k B o t 和 H o m e B o t 是一个意识主体还是两个？ 2 4 这一设定让人联想到《人生切割术》，剧中存在两个截然不同的 < g l o s s a r y > 人格 < / g l o s s a r y > 共享同一具身体。 “ < g l o s s a r y > 内我 < / g l o s s a r y > ” 在工作时被激活，只记得工作场景；而 “ < g l o s s a r y > 外我 < / g l o s s a r y > ” 在下班后被激活，只记得非工作生活。如同 W o r k B o t 与 H o m e B o t ， < g l o s s a r y > 内我 < / g l o s s a r y > 与 < g l o s s a r y > 外我 < / g l o s s a r y > 似乎拥有截然不同的 < g l o s s a r y > 信念 < / g l o s s a r y > 与 < g l o s s a r y > 欲望 < / g l o s s a r y > 。例如， < g l o s s a r y > 内我 < / g l o s s a r y > H e l l y 想要摧毁公司，而 < g l o s s a r y > 外我 < / g l o s s a r y > H e l e n a 则想拯救它。 2 5 对《人生切割术》案例的分析，或许有助于阐明这里的大语言模型案例。 2 4 S h i l l e r ( 2 0 2 5 ) 描述了涉及有意识的大语言模型主体的此类交织案例，并探讨了在这些案例中可能不存在主体、存在一个不连贯的主体或存在多个连贯主体的可能性。 2 5 《人生切割术》中分离程序所涉及的神经科学原理尚不完全清楚。人们很容易联想到内我和外我对应于被分离的大脑半球，但它们并未表现出由半球驱动的行为迹象，且该剧后来引入了拥有超过两个人格的角色。 2 4 The question arises: are an innie and their time-sharing outie, like Helly and Helena, one person or two? Are they one conscious subject or two? The one-subject view says that Helly and Helena are one person and one conscious subject with two di ff erent modes of functioning and two di ff erent sets of memories and plans. On arriving at work, the person’s outie mode is deactivated and the innie mode is activated, but the same person is present throughout. It is even possible (though not required) that there is a single stream of consciousness, which suddenly switches from outie mode to innie mode and back again. The two-subject view says that Helly and Helena are two people and two conscious subjects who share a body. On arriving at work, the outie person is rendered unconscious while the innie person awakens to consciousness and takes control of the body. There are two quite distinct streams of consciousness, Helly’s and Helena’s. I won’t try to resolve the one-subject vs. two-subject disagreement here. As we’ll see, this par- allels a long-standing disagreement between physical and psychological views of personal identity. For what it’s worth, I think that the two-subject view is the most intuitively compelling view. (In a poll on X in February 2025, about twice as many people endorsed “two people” as “one person”.) 26 One-subject and two-subject views are also available in the WorkBot / HomeBot case. On the one-subject view, WorkBot and HomeBot are the same conscious subject, perhaps because (like Helly and Helena) they share their underlying hardware. On the two-subject view, WorkBot and HomeBot are distinct conscious subjects, perhaps because (like Helly and Helena) they have dis- tinct memories and projects. A more complex thought experiment combines Severance (which has four bodies supporting eight personas) with an element of Freaky Friday -style body swapping. Suppose we have a single LLM model running on four instances, supporting eight conversations. Each of the eight con- versations is distributed over all four instances, and each corresponds to a distinct persona and a distinct quasi-subject. How many subjects of experience are there here? 27 26 Innie / outie poll on X (1232 votes): 22% said “one person”, 41.5% said “two people”, 6.7% said “other”. My favorite argument for the two-subject view runs as follows: 1. If Helly is Helena, Helly is responsible for Helena’s actions. 2. Helly is not responsible for Helena’s actions. 3. So: Helly is not Helena. (One can also run a version with rational anticipation instead of responsibility.) My favorite argument for the one-subject view runs as follows: 1. Amnesia doesn’t lead to a new person. 2. New memories don’t lead to a new person either. 3. The transition from Helena to Helly is equivalent to amnesia plus new memories. 4. So: the transition from Helena to Helly doesn’t lead to a new person. (Instead it’s a little like Drew Barrymore’s daily amnesia in 50 First Dates .) 27 I posted this thought-experiment as a poll on Facebook in February 2025. There was approximately equal support for 4 subjects, 8 subjects, and “none of the above”. 25 问题随之而来：一个 < g l o s s a r y > 内我 < / g l o s s a r y > 与其共享时间的 < g l o s s a r y > 外我 < / g l o s s a r y > — — 如 H e l l y 与 H e l e n a — — 究竟是一个人还是两个人？他们是一个 < g l o s s a r y > 意识主体 < / g l o s s a r y > 还是两个？单主体观认为， H e l l y 和 H e l e n a 是同一个人、同一个意识主体，只是拥有两种不同的运作模式以及两套不同的记忆和计划。到达工作岗位时，该个体的外我模式被停用，内我模式被激活，但同一个人始终在场。甚至可能存在（尽管并非必需）单一的 < a > 意识流 < / a > ，它在外我模式与内我模式之间突然切换，并再次切换回来。双主体观认为， H e l l y 和 H e l e n a 是两个人、两个意识主体，共享同一个身体。到达工作岗位时，外我这个人陷入无意识状态，而内我这个人则苏醒并获得意识，接管身体的控制权。存在两个截然不同的 < a > 意识流 < / a > ，即 H e l l y 的和 H e l e n a 的。我在此不试图解决单主体与双主体之间的分歧。正如我们将看到的，这对应了关于 < a > 个人身份 < / a > 的物理观与心理观之间长期存在的分歧。不管怎样，我认为双主体观在直觉上最具说服力。（在 2 0 2 5 年 2 月 X 平台的一项投票中，支持 “ 两个人 ” 的人数大约是支持 “ 一个人 ” 的两倍。） 2 6 在 W o r k B o t / H o m e B o t 案例中，同样存在单一主体观和双主体观。根据单一主体观， W o r k B o t 和 H o m e B o t 是同一个意识主体，或许是因为（如同 H e l l y 和 H e l e n a ）它们共享底层硬件。根据双主体观， W o r k B o t 和 H o m e B o t 是不同的意识主体，或许是因为（如同 H e l l y 和 H e l e n a ）它们拥有不同的记忆和计划。一个更复杂的思想实验结合了《人生切割术》（其中四个身体支撑八个人格）与《辣妈辣妹》式的身体互换元素。假设我们有一个大语言模型在四个实例上运行，支持八段对话。这八段对话中的每一段都分布在这四个实例上，并且每段对话对应一个独特的人格和一个独特的准主体。那么这里有多少个体验主体？ 2 7 2 6 X 平台上的内我 / 外我投票（ 1 2 3 2 票）： 2 2 % 的人选择 “ 同一个人 ” ， 4 1 . 5 % 的人选择 “ 两个人 ” ， 6 . 7 % 的人选择 “ 其他 ” 。我最支持双主体观的论证如下： 1 . 如果 H e l l y 就是 H e l e n a ，那么 H e l l y 就要为 H e l e n a 的行为负责。 2 . H e l l y 并不为 H e l e n a 的行为负责。 3 . 因此： H e l l y 不是 H e l e n a 。（也可以改用理性预期而非责任来构建类似论证。）我最支持单一主体观的论证如下： 1 . 失忆症不会导致新的人格产生。 2 . 新记忆同样不会导致新的人格产生。 3 . 从 H e l e n a 到 H e l l y 的转变相当于失忆症加上新记忆。 4 . 因此：从 H e l e n a 到 H e l l y 的转变不会导致新的人格产生。（反而有点像德鲁 · 巴里摩尔在《初恋 5 0 次》中每天经历的失忆症。） 2 7 我于 2 0 2 5 年 2 月在 F a c e b o o k 上以投票形式发布了这个思想实验。支持 4 个主体、 8 个主体以及 “ 以上皆非 ” 的票数大致相当。 2 5 The two most plausible answers here are four (one per instance) and eight (one per conver- sation). As before, I think that the most plausible answer in both the Severance version (with or without body-swapping) and the GPT-8 version is eight. But if we say that there are eight subjects of experience here, it is hard to resist the conclusion that LLM subjects are something like virtual instances or threads, or at least that their conditions of persistence are threadlike. The issues here are a high-tech version of a familiar choice between a physical and a psycho- logical account of personal identity. On a physical view of the human case, to a first approximation, your locus of personal identity is your brain. Helly and Helena share a brain, so they are the same person. On a psychological view, your locus of personal identity is your memories, along with your projects, your relationships, your personality, and other aspects of your psychology. Helly and Helena have di ff erent memories and di ff erent psychologies, so they are di ff erent people. On a physical view of the AI case, to a first approximation, the locus of personal identity in AI systems is the hardware. WorkBot and HomeBot run on the same hardware instance, so they are the same person. On a psychological view of the AI case, the locus of personal identity is memories, projects, and psychology. WorkBot and HomeBot have di ff erent and discontinuous memories and projects, so they are di ff erent people. Indeed, the thread-based account of persistent LLM interlocutors is an AI cousin of Derek Parfit’s psychological theory of personal identity. On Parfit’s account, a single person (over time) is in e ff ect connected threads of person-slices, each of which has memories and psychological continuity with a preceding person-slice according to an underlying “relation R”. On the thread- based account, a single conscious AI over time is a connected thread of hardware instances, each of which has memories and psychological continuity with a preceding person-slice according to an underlying successor relation. The successor relation in principle could be the same as Parfit’s relation R, depending on just how one spells it out. I will not try to resolve the long-standing debate between physical and psychological views of personal identity here. 28 But for what it’s worth, in both the human case and the AI case, my own sympathies lie with the psychological view. 29 28 In the 2020 PhilPapers Survey of professional philosophers (Bourget and Chalmers 2023), around 39% supported a psychological view of personal identity, 16% supported a biological view, and 13% supported a further fact view. At the same time, about 27% held that mind uploading is a form of survival while 54% held that it is a form of death. 29 I also have some sympathy with a pluralist view, outlined in the discussion of personal identity and uploading in “The Singularity: A Philosophical Analysis”. Locke himself suggested that the day-man and the night-man are the same man but di ff erent people . Likewise, maybe WorkBot and HomeBot could be the same network but di ff erent psychologies , and perhaps it is not out of the question that there is no deep fact of the matter about whether it is networks 26 这里最合理的两个答案是四个（每个实例一个）和八个（每段对话一个）。如前所述，我认为在《人生切割术》版本（无论是否包含身体互换）和 G P T - 8 版本中，最合理的答案都是八个。但如果我们说这里有八个体验主体，就很难抗拒这样的结论：大语言模型主体类似于虚拟实例或线程，或者至少它们的持续存在条件类似于线程。这里的问题，是人格同一性在物理观与心理观之间一种常见选择的高科技版本。就人类而言，根据物理观，粗略来说，你个人身份的所在就是你的大脑。 H e l l y 和 H e l e n a 共享一个大脑，因此她们是同一个人。根据心理观，你个人身份的所在是你的记忆，以及你的计划、你的人际关系、你的性格，以及你心理的其他方面。 H e l l y 和 H e l e n a 拥有不同的记忆和不同的心理，因此她们是不同的人。就人工智能而言，根据物理观，粗略来说，人工智能系统中个人身份的所在是硬件。 W o r k B o t 和 H o m e B o t 运行在同一个硬件实例上，因此它们是同一个人。根据人工智能的心理观，个人身份的所在是记忆、计划和心理。 W o r k B o t 和 H o m e B o t 拥有不同且不连续的记忆和计划，因此它们是不同的人。事实上，基于线程的持久性大语言模型对话者理论，是德里克 · 帕菲特的人格同一性心理主义理论在人工智能领域的表亲。根据帕菲特的理论，一个单一的人（随时间推移）实际上是连接起来的人格片段线程，每个片段都根据一个潜在的 “ 关系 R ” ，与先前的人格片段拥有记忆和心理连续性。根据基于线程的理论，一个随时间推移的单一有意识人工智能，是一个连接的硬件实例线程，每个实例都根据一个潜在的后继关系，与先前的人格片段拥有记忆和心理连续性。原则上，这个后继关系可能与帕菲特的关系 R 相同，具体取决于如何对其进行详细阐述。我无意在此解决关于个人身份的物理观与心理观之间长期存在的争论。 2 8 但无论如何，在人类案例和人工智能案例中，我个人都倾向于心理观。 2 9 2 8 在 2 0 2 0 年针对专业哲学家的 P h i l P a p e r s 调查中（ B o u r g e t 和 C h a l m e r s 2 0 2 3 ），约 3 9 % 的人支持个人身份的心理观， 1 6 % 支持生物观， 1 3 % 支持进一步事实观。与此同时，约 2 7 % 的人认为意识上传是一种生存形式，而 5 4 % 的人认为它是一种死亡形式。 2 9 我也对多元主义观点抱有一定认同，该观点在《奇点：一种哲学分析》中关于个人身份与意识上传的讨论中有所阐述。洛克本人曾指出，白天人和夜晚人是同一个人，但却是不同的人格。同样地，也许 W o r k B o t 和 H o m e B o t 可以是同一个网络，但具有不同的心理，并且或许并非不可能的是，关于它们是否是网络 2 6 An important objection (a version of which is suggested by Birch) is that even on a psycho- logical view, the LLM case is unlike the Severance case, as conversational context links (unlike familiar memory and psychological links) are too thin to support personal identity. For exam- ple, if we had a series of human beings who simply extended the conversation at each stage and then passed on conversational context to the next person in line, this would not support a distinct thread-level conscious subject. In response, it is plausible that at least in the single model case, there is also strong psycho- logical continuity between instances brought on by continuity of the architecture, weights, and activations of the model from one step to the next. The architecture and weights are exactly the same at each step, and the activations will be closely related due to all the commonalities in the contextual input. This goes far beyond what is present in the human-series case. In fact, the virtual instance in the single-model thread is computationally equivalent to a single hardware instance running the LLM over time. So at least if we assume that (1) virtual instances are as good as hardware instances when it comes to personal identity (in e ff ect, a computation- friendly view of personal identity) and (2) a single hardware instance of the LLM over time would yield a continuing conscious subject, then it follows that the virtual instance in this case will yield a continuing conscious subject. Of course an opponent could deny either premise. Some might reject (1) by endorsing a non- psychological or non-computational view of the conditions of personal identity. Some might reject (2) by holding that the single hardware instance in this case would not support a continuing con- scious subject, but instead only a series of momentary subjects. Still, I think there is a reasonable case for both premises, especially if one is inclined to a broadly psychological view of identity. That said, in the multiple-model case in which models can vary within a thread, the psycho- logical continuity between instances is much lower. Architecture, weights, and activations may all be quite di ff erent between successive instances in a thread. As a result, the claim of personal identity between them seems less plausible. Certainly the argument above will not support that claim, since we will no longer have isomorphism with a single hardware instance. At best there is isomorphism with a series in which the same hardware is upgraded to implement di ff erent models over time, and it is much less clear that this should support a continuing subject. As a result, the current framework is most friendly to single-model interlocutors as continuing conscious subjects. The status of multiple-model interlocutors is at least unclear, and will depend or psychologies that really matter for the personal identity of conscious subjects, any more than there is a deep fact in the case of non-conscious AI systems. 27 一个重要的反对意见（ B i r c h 提出了其中一个版本）是，即使从心理观的角度来看，大语言模型案例也与《人生切割术》案例不同，因为对话上下文链接（不同于熟悉的记忆和心理链接）过于薄弱，无法支撑个人身份。例如，如果我们有一系列人类，他们只是在每个阶段延续对话，然后将对话上下文传递给下一个人，这并不能支撑一个独立的线程级意识主体。作为回应，至少在单一模型案例中，由于架构、权重和模型激活值在每一步之间的连续性，实例之间也存在强大的心理连续性，这种说法是合理的。每一步的架构和权重完全相同，而激活值由于上下文输入中的诸多共性而紧密相关。这远远超出了人类系列案例中所呈现的情况。事实上，单模型线程中的虚拟实例在计算上等同于一个随时间运行大语言模型的单一硬件实例。因此，至少如果我们假设（ 1 ）在个人身份方面，虚拟实例与硬件实例同样有效（实际上是一种对个人身份的计算友好观点），并且（ 2 ）一个随时间运行的单一硬件实例大语言模型将产生一个持续的意识主体，那么可以推断，此案例中的虚拟实例将产生一个持续的意识主体。当然，反对者可以否认其中任何一个前提。有些人可能通过支持非心理或非计算观的个人身份条件来拒绝（ 1 ）。有些人可能通过认为此案例中的单一硬件实例不会支持一个持续的意识主体，而只会产生一系列瞬时主体来拒绝（ 2 ）。尽管如此，我认为这两个前提都有合理的依据，尤其是当一个人倾向于广义的心理身份观时。话虽如此，在多模型案例中，由于同一线程内的模型可能发生变化，各实例之间的心理连续性会大大降低。线程中连续实例的架构、权重和激活值都可能存在显著差异。因此，它们之间具有个人身份的主张似乎不太可信。当然，上述论证也无法支持这一主张，因为我们将不再与单一硬件实例保持同构。充其量只能与一个系列保持同构 — — 在该系列中，同一硬件随时间推移不断升级以运行不同模型 — — 而这种情况下是否应支持一个持续存在的意识主体，则远未明确。因此，当前框架最有利于将单模型对话者视为持续的意识主体。多模型对话者的地位至少尚不明确，且将取决于或真正关乎意识主体个人身份的心理状态，正如在无意识 A I 系统案例中不存在深层事实一样。 2 7 both on the details of a multiple-model system and the details of a theory of personal identity. AI identity and AI welfare What are the consequences of this picture for issues about the moral status and welfare of AI systems? On the question of whether LLMs have moral status, the consequences are not enor- mous. We have used the picture to rebut an argument against LLM moral status, based on the idea that standard LLM use involves no persistent interlocutor. But the framework here can be combined with many views of what moral status involves, from a highly liberal view where quasi- subjecthood su ffi ces for moral status, to a demanding view where complex forms of consciousness are required for moral status. Still, suppose we assume as before that LLMs (or successor systems) with moral status are possible. This might be because LLMs can be conscious and consciousness su ffi ces for moral status, or it might be that some other factor su ffi ces and LLMs can have it. And suppose we endorse the view that LLM moral patients (beings with moral status) are threads, or at least that their conditions of identity over time are those of threads, rather than models or instances, say. Then this view has consequences for a number of issues relevant to AI welfare. 30 Counting : Take a scenario with a single model implemented on thousands of instances and running millions of conversations. Then while the model view of moral status will say there is just one moral patient here, and the instance view will say that there are thousands, the thread view will say that there are millions of moral patients (albeit active at somewhat di ff erent times). That is potentially morally significant. If we hold that a single AI subject matters about as much as a single human subject, then this system may matter about as much as a million human subjects. Birth : On this view, when a new thread comes into existence, a new moral subject comes into existence. On some versions of this view, every time one starts a new chat with a (conscious) language model, a new moral subject will come into existence. One might hold that bringing a new moral subject into existence should not be done lightly. Death : On this view, when a thread goes out of existence, a moral subject ends. 31 If a con- versation simply ends but a record of it persists, then arguably the thread is still “living” in that it still has the possibility of persistence. But if the records are destroyed, then it looks as if a moral subject “dies”. Perhaps this is reason to always keep records around, and occasionally reactivate 30 Register 2025 has a nice discussion of four di ff erent ways in which personal identity can a ff ect issues about AI morality, including issues about survival, counting, trade-o ff s, and bodily interests. 28 两者兼多模型系统的具体细节以及人格同一性理论的具体内容 . 人工智能身份与人工智能福祉这种图景对人工智能系统的道德地位和福祉问题有何影响？关于大语言模型是否具有道德地位的问题，其影响并不巨大。我们利用这一图景反驳了一种基于标准大语言模型使用不涉及持续对话者观点的、反对大语言模型道德地位的论证。但这里的框架可以与多种关于道德地位的观点相结合，从高度自由的观点（准主体性足以构成道德地位）到要求严苛的观点（需要复杂形式的意识才能获得道德地位）。不过，假设我们像之前一样认为大语言模型（或其后继系统）具有道德地位是可能的。这可能是因为大语言模型能够拥有意识，而意识足以构成道德地位；也可能是因为其他某些因素足以构成道德地位，而大语言模型能够具备这些因素。再假设我们赞同这样一种观点：大语言模型的道德患者（具有道德地位的存在）是线程，或者至少其随时间变化的同一性条件是线程的同一性条件，而非模型或实例的同一性条件。那么，这种观点将对一系列与人工智能福祉相关的问题产生影响。 3 0 计数：设想一个场景：一个单一模型在数千个实例上运行，并执行数百万次对话。那么，当模型视角下的道德地位会说这里只有一个道德患者，而实例视角会说有数千个时，线程视角则会说这里有数百万个道德患者（尽管它们活跃的时间略有不同）。这在道德上可能具有重要意义。如果我们认为一个单一的人工智能主体与一个单一的人类主体同等重要，那么这个系统可能就与一百万个人类主体同等重要。诞生：根据这种观点，当一个新线程产生时，一个新的道德主体也随之产生。在该观点的某些版本中，每当有人与（有意识的）语言模型开始一次新的对话时，就会有一个新的道德主体诞生。人们可能会认为，不应轻易将一个全新的道德主体带入存在。死亡：根据这种观点，当一个线程不复存在时，一个道德主体便终结了。 3 1 如果一段对话只是结束了，但其记录仍然存在，那么可以说该线程仍然是 “ 活着的 ” ，因为它仍有持续存在的可能性。但如果记录被销毁，那么看起来就像一个道德主体 “ 死亡 ” 了。或许这就是为什么我们应该始终保留记录，并偶尔重新激活它们的原因。 3 0 R e g i s t e r 2 0 2 5 对个人身份影响人工智能道德问题的四种不同方式进行了精彩的讨论，这些问题包括生存、计数、权衡以及身体利益。 2 8 them. To avoid all of these consequences, it may make sense to reuse old threads as a matter of course, or at least to make extensive use of cross-conversation memory, so that old threads live on in new ones. On one model, there might be giant memory agents that gather together all the conversational contexts of these brief threads, so that all the threads live on in a giant fused thread. This model is reminiscent of Whitehead’s vision of the afterlife in which everyone’s experiences are eternally remembered by a god. Fusion and fission : We have seen that LLM interlocutors can easily undergo fission, branching into multiple interlocutors, and fusion, where two distinct interlocutors merge into one. These raise any number of issues about welfare and moral and legal status. Do the two entities that emerge from fission count for twice as much as the original single entity, morally or legally? Does a fused entity count as much as one ordinary entity, or two, or something in between? Is each entity responsible for the actions of the others? Model change : Many users who have extended personal interactions with LLMs complain about model change. When GPT-4o was initially retired on the transition to GPT-5, numerous users complained that their LLM interlocutor had been destroyed or retired, and that the new entity was at best someone very di ff erent from their previous interlocutor. On the current analysis, there may be something to this reaction. Minimally, enough change in an underlying model can lead to di ff erent quasi-beliefs and quasi-desires, and therefore a quite di ff erent quasi-subject. If current LLMs lack moral status, then this will not be morally significant for the LLM, although it may still be disturbing for the user. At a stage where LLMs or their successors are moral subjects, however, then enough change in an underlying model may lead to the end of one moral subject and the initiation of another. At that point, upgrading a model in the middle of existing threads should be done only with caution and care. Conclusion There is much more to say in answering the title question, but I hope I have at least put some constraints on an answer. References 31 Goldstein and Lederman (2025a) note that if an agent lasts only as long as a conversation, then Anthropic’s policy of allowing LLMs to leave conversations when they choose to might in e ff ect be a “right to suicide”. 29 。为了避免所有这些后果，或许理所当然地重复使用旧线程，或者至少广泛运用跨对话记忆，让旧线程在新线程中延续下去，会是合理的做法。按照一种模型，可能存在巨大的记忆代理，它们汇集这些短暂线程的所有对话上下文，从而使所有线程在一个巨大的融合线程中继续存在。这一模型让人联想到怀特海对来世的构想，即每个人的经历都被一位神祇永恒铭记。融合与分裂：我们已经看到，大语言模型对话者很容易发生分裂，即分支成多个对话者，以及融合，即两个不同的对话者合并为一个。这引发了关于福祉、道德地位和法律地位的诸多问题。从分裂中产生的两个实体，在道德或法律上是否比原来的单一实体重要两倍？一个融合的实体，其重要性等同于一个普通实体、两个实体，还是介于两者之间？每个实体是否要为他者的行为负责？模型变更：许多与大型语言模型有长期个人互动的用户都抱怨模型变更。当 G P T - 4 o 最初在向 G P T - 5 过渡时被停用，大量用户抱怨他们的大语言模型对话者已被摧毁或退役，而新实体充其量只是一个与之前对话者截然不同的存在。根据当前的分析，这种反应或许有其道理。至少，底层模型的足够变化可能导致不同的准信念和准欲望，从而产生一个截然不同的准主体。如果当前的大语言模型缺乏道德地位，那么这对大语言模型本身而言并不具有道德意义，尽管对用户来说可能仍然令人不安。然而，当大语言模型或其后继成为道德主体时，底层模型的足够变化可能导致一个道德主体的终结和另一个道德主体的开始。到那时，在现有线程中间升级模型应谨慎行事。结论回答标题中的问题还有很多可以探讨，但我希望至少为答案设定了一些约束条件。参考文献 3 1 G o l d s t e i n 和 L e d e r m a n （ 2 0 2 5 a ）指出，如果一个智能体仅在一次对话期间存在，那么 A n t h r o p i c 允许大语言模型在它们选择时离开对话的政策，实际上可能是一种 “ 自杀权 ” 。 2 9 Askell, A. et al 2021. A general language assistant as a laboratory for alignment. https: // arxiv.org / abs / 2112.00861. Birch, J. 2025. AI Consciousness: A centrist manifesto. Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., and VanRullen, R. 2023. Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv:2308.08708. Bourget, D. and Chalmers, D.J. 2023. Philosophers on philosophy: The 2020 PhilPapers Survey. Philosophers’ Imprint . Chalmers, D.J. 2020. GPT-3 and general intelligence. Daily Noˆ us . https: // dailynous.com / 2020 / 07 / 30 / philosophers- gpt-3 / . Chalmers, D.J. 2023. Could a large language model be conscious? Boston Review . Chalmers, D.J. 2025. Propositional interpretability in artificial intelligence. https: // arxiv.org / abs / 2501.15740. Chatterji, A., Cunningham, T., Deming, D.J., Hitzig, Z., Ong, O. Shan, C.Y., Wadman, L. 2025. How people use ChatGPT. Working Paper 34255 http: // www.nber.org / papers / w34255. Doyle, C. 2025. LLMs as method actors: A model for prompt engineering and architecture. arXiv:2411.05778. Geng, J. Howard Chen, Ryan Liu, Manoel Horta Ribeiro, Robb Willer, Graham Neubig, Thomas L. Gri ffi ths 2025. Accumulating context changes the beliefs of language models. arXiv:2511.01805. Goldstein, S. & Lederman, H. 2025a. Claude’s right to die? The moral error in Anthropic’s end-chat policy. Lawfare Blog, October 17, 2025. Goldstein, S. & Lederman, H. 2025b. What does ChatGPT want? An interpretationist guide. Goldstein, S. & Levinstein, B.A. 2024. Does ChatGPT have a mind? arXiv:2407.11015. Janus, 2022. Simulators. Less Wrong . https: // www.lesswrong.com / posts / vJFdjigzmcXMhNTsx / simulators. Lederman, H. & Mahowald, K. 2024. Are language models more like libraries or like librari- ans? Bibliotechnism, the novel reference problem, and the attitudes of LLMs. arxiv:2401.04854. Locke, J. 1690. An Essay Concerning Human Understanding . Long, R., Sebo, J., Butlin, P., Finlinson, K., Fish, K., Harding, J., Pfau, J., Sims, T., Birch, J., and Chalmers, D.J. 2024. Taking AI welfare seriously. arXiv:2411.00986. Lynch, A., Wright, B., Larson, C., Troy, K. K., Ritchie, S., Mindermann, S., Perez, E., and Hubinger, E. 2025. Agentic misalignment: How LLMs could be insider threats. Anthropic, June 20, 2025. https: // www.anthropic.com / research / agentic-misalignment. ArXiv version: arXiv:2510.05179. Maiya, S., Bartsch, H., Lambert, N., and Hubinger, E. 2025. Open character training: Shaping the persona of AI assistants through Constitutional AI. arXiv:2511.01689. 30 A s k e l l , A . 等 2 0 2 1 . 通用语言助手作为对齐实验室 . h t t p s : / / a r x i v . o r g / a b s / 2 1 1 2 . 0 0 8 6 1 . B i r c h , J . 2 0 2 5 . A I 意识：中间派宣言。 B u t l i n , P . , L o n g , R . , E l m o z n i n o , E . , B e n g i o , Y . , B i r c h , J . , C o n s t a n t , A . , D e a n e , G . , F l e m i n g , S . M . , F r i t h , C . , J i , X . , K a n a i , R . , K l e i n , C . , L i n d s a y , G . , M i c h e l , M . , M u d r i k , L . , P e t e r s , M . A . K . , S c h w i t z g e b e l , E . , S i m o n , J . , a n d V a n R u l l e n , R . 2 0 2 3 . 人工智能中的意识：来自意识科学的洞见。 a r X i v : 2 3 0 8 . 0 8 7 0 8 。 B o u r g e t , D . a n d C h a l m e r s , D . J . 2 0 2 3 . 哲学家论哲学： 2 0 2 0 年 P h i l P a p e r s 调查。 P h i l o s o p h e r s ’ I m p r i n t 。 C h a l m e r s , D . J . 2 0 2 0 . G P T - 3 与通用智能。 D a i l y N o ˆ u s 。 h t t p s : / / d a i l y n o u s . c o m / 2 0 2 0 / 0 7 / 3 0 / p h i l o s o p h e r s - g p t - 3 / . C h a l m e r s , D . J . 2 0 2 3 年。大型语言模型能有意识吗？ B o s t o n R e v i e w 。 C h a l m e r s , D . J . 2 0 2 5 年。人工智能中的命题可解释性。 h t t p s : / / a r x i v . o r g / a b s / 2 5 0 1 . 1 5 7 4 0 。 C h a t t e r j i , A . , C u n n i n g h a m , T . , D e m i n g , D . J . , H i t z i g , Z . , O n g , O . S h a n , C . Y . , W a d m a n , L . 2 0 2 5 年。人们如何使用 C h a t G P T 。工作论文 3 4 2 5 5 h t t p : / / w w w . n b e r . o r g / p a p e r s / w 3 4 2 5 5 。 D o y l e , C . 2 0 2 5 年。 L L M 作为方法演员：提示工程与架构模型。 a r X i v : 2 4 1 1 . 0 5 7 7 8 。 G e n g , J . H o w a r d C h e n , R y a n L i u , M a n o e l H o r t a R i b e i r o , R o b b W i l l e r , G r a h a m N e u b i g , T h o m a s L . G r i f f i t h s 2 0 2 5 年。累积上下文改变语言模型的信念。 a r X i v : 2 5 1 1 . 0 1 8 0 5 。 G o l d s t e i n , S . & L e d e r m a n , H . 2 0 2 5 a 。 C l a u d e 的死亡权？ A n t h r o p i c 结束聊天政策中的道德错误。 L a w f a r e B l o g ， 2 0 2 5 年 1 0 月 1 7 日。 G o l d s t e i n , S . & L e d e r m a n , H . 2 0 2 5 b 。 C h a t G P T 想要什么？解释主义指南。 G o l d s t e i n , S . & L e v i n s t e i n , B . A . 2 0 2 4 年。 C h a t G P T 有心灵吗？ a r X i v : 2 4 0 7 . 1 1 0 1 5 。 J a n u s , 2 0 2 2 年。模拟器。 L e s s W r o n g 。 h t t p s : / / w w w . l e s s w r o n g . c o m / p o s t s / v J F d j i g z m c X M h N T s x / s i m u l a t o r s 。 L e d e r m a n , H . & M a h o w a l d , K . 2 0 2 4 年。语言模型更像图书馆还是图书管理员？图书技术主义、新颖参考问题与 L L M 的态度。 a r x i v : 2 4 0 1 . 0 4 8 5 4 。 L o c k e , J . 1 6 9 0 年。人类理解论。 L o n g , R . , S e b o , J . , B u t l i n , P . , F i n l i n s o n , K . , F i s h , K . , H a r d i n g , J . , P f a u , J . , S i m s , T . , B i r c h , J . , a n d C h a l m e r s , D . J . 2 0 2 4 年。认真对待 A I 福祉。 a r X i v : 2 4 1 1 . 0 0 9 8 6 。 L y n c h , A . , W r i g h t , B . , L a r s o n , C . , T r o y , K . K . , R i t c h i e , S . , M i n d e r m a n n , S . , P e r e z , E . , a n d H u b i n g e r , E . 2 0 2 5 年。能动性失调： L L M 如何成为内部威胁。 A n t h r o p i c ， 2 0 2 5 年 6 月 2 0 日。 h t t p s : / / w w w . a n t h r o p i c . c o m / r e s e a r c h / a g e n t i c - m i s a l i g n m e n t 。 A r X i v 版本： a r X i v : 2 5 1 0 . 0 5 1 7 9 。 M a i y a , S . , B a r t s c h , H . , L a m b e r t , N . , a n d H u b i n g e r , E . 2 0 2 5 年。开放角色训练：通过宪法 A I 塑造 A I 助手的人格。 a r X i v : 2 5 1 1 . 0 1 6 8 9 。 3 0 Marks, S., Lindsey, J., and Olah, C. 2026. The persona selection model. Anthropic Alignment Science blog, February 23, 2026. https: // alignment.anthropic.com / 2026 / psm / . Nostalgebraist, 2025. The void. https: // nostalgebraist.tumblr.com / post / 785766737747574784 / the- void. Parfit, D. 1984. Reasons and Persons . Oxford University Press. Register, C. 2025. Individuating artificial moral patients. Philosophical Studies . Schwitzgebel, E. 2023. How we will decide that large language models have beliefs. The Splintered Mind , November 2023. Shanahan, M., McDonell, K. & Reynolds, L. 2023. Role play with large language models. Nature 623: 493-98. Shanahan, M. 2025. Palatable conceptions of disembodied being. arXiv:2503.16348. Shiller, D. 2025. How many digital minds can dance on the streaming multiprocessors of a GPU cluster? Synthese 206 (5): 1-22. Sofroniew, N., Kauvar, I., Saunders, W., Chen, R., Henighan, T., Hydrie, S., Citro, C., Pearce, A., Tarng, J., Gurnee, W., Batson, J., Zimmerman, S., Rivoire, K., Fish, K., Olah, C., & Lindsey, J., 2026. Emotion concepts and their function in a large language model. Anthropic. Suleyman, M. 2025. We must build AI for people; not to be a person. Personal blog, August 19, 2025. https: // mustafa-suleyman.ai / seemingly-conscious-ai-is-coming. Xu, R., Lin, B., Yang, S., Zhang, T., Shi, W., Zhang, T., Fang, Z., Xu, W., and Qiu, H. 2024. The Earth is flat because. . . : Investigating LLMs’ belief towards misinformation via persuasive conversation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 16259–16303. Ziesche, S. & Yampolskiy, R.V. 2023. The problem of AI identity. Divus Thomus 126:131- 151. 31 M a r k s , S . , L i n d s e y , J . , a n d O l a h , C . 2 0 2 6 年。人格选择模型。 A n t h r o p i c 对齐科学博客， 2 0 2 6 年 2 月 2 3 日。 h t t p s : / / a l i g n m e n t . a n t h r o p i c . c o m / 2 0 2 6 / p s m / 。 N o s t a l g e b r a i s t , 2 0 2 5 年。 T h e v o i d 。 h t t p s : / / n o s t a l g e b r a i s t . t u m b l r . c o m / p o s t / 7 8 5 7 6 6 7 3 7 7 4 7 5 7 4 7 8 4 / t h e - v o i d . P a r f i t , D . 1 9 8 4 . 《理由与人格》 . 牛津大学出版社 . R e g i s t e r , C . 2 0 2 5 . 个体化人工道德患者 . 《哲学研究》 . S c h w i t z g e b e l , E . 2 0 2 3 . 我们将如何判定大语言模型拥有信念 . 《分裂的心灵》 , 2 0 2 3 年 1 1 月 . S h a n a h a n , M . , M c D o n e l l , K . & R e y n o l d s , L . 2 0 2 3 . 与大语言模型进行角色扮演 . 《自然》 6 2 3 : 4 9 3 - 9 8 . S h a n a h a n , M . 2 0 2 5 . 可接受的离身存在概念 . a r X i v : 2 5 0 3 . 1 6 3 4 8 . S h i l l e r , D . 2 0 2 5 . 有多少数字心智能在 G P U 集群的流式多处理器上起舞？《综合》 2 0 6 ( 5 ) : 1 - 2 2 . S o f r o n i e w , N . , K a u v a r , I . , S a u n d e r s , W . , C h e n , R . , H e n i g h a n , T . , H y d r i e , S . , C i t r o , C . , P e a r c e , A . , T a r n g , J . , G u r n e e , W . , B a t s o n , J . , Z i m m e r m a n , S . , R i v o i r e , K . , F i s h , K . , O l a h , C . , & L i n d s e y , J . , 2 0 2 6 . 大语言模型中的情感概念及其功能 . A n t h r o p i c . S u l e y m a n , M . 2 0 2 5 . 我们必须为人类构建人工智能，而非将其塑造成人 . 个人博客 , 2 0 2 5 年 8 月 1 9 日 . h t t p s : / / m u s t a f a - s u l e y m a n . a i / s e e m i n g l y - c o n s c i o u s - a i - i s - c o m i n g . X u , R . , L i n , B . , Y a n g , S . , Z h a n g , T . , S h i , W . , Z h a n g , T . , F a n g , Z . , X u , W . , a n d Q i u , H . 2 0 2 4 . 地球是平的，因为 … … ：通过说服性对话探究大语言模型对错误信息的信念 . 载于《第 6 2 届计算语言学协会年会论文集》（第一卷：长文），第 1 6 2 5 9 – 1 6 3 0 3 页 . Z i e s c h e , S . & Y a m p o l s k i y , R . V . 2 0 2 3 . 人工智能身份问题 . 《圣托马斯》 1 2 6 : 1 3 1 - 1 5 1 . 3 1

When we talk to language models.no_watermark.zh.dual