What If Your Life Had Multiple Possible Versions? Building an Auditable Counterfactual Self Architecture

本文为原文前 6,000 字符的节选翻译，完整内容请查看原文。

Can AI simulate alternative versions of a person without confusing historical evidence, plausible assumptions, and fictional narratives? Most AI systems are designed to predict what happens next. But there is another, more difficult problem: What could have happened if one important decision had been different? What if Alan Turing had accepted John von Neumann’s offer and stayed at Princeton in 1938? What if he had remained at the National Physical Laboratory and continued working on the ACE computer? What if the prosecution in 1952 had never happened? These are not ordinary prediction questions. They involve incomplete evidence, alternative causal paths, uncertain outcomes, and the danger of turning plausible fiction into apparent historical fact. This is the problem behind Counterfactual Self Architecture (CSA). I recently published the research framework on PhilPapers: “Read the CSA paper on PhilPapers” (https://philpapers.org/rec/ALHCSA)

人工智能能否在不混淆历史证据、合理假设和虚构叙事的情况下，模拟一个人的不同版本？大多数人工智能系统旨在预测接下来会发生什么。但还有一个更困难的问题：如果一个重要的决定有所不同，会发生什么？如果艾伦·图灵在1938年接受了约翰·冯·诺依曼的邀请并留在普林斯顿会怎样？如果他留在国家物理实验室并继续研究ACE计算机又会怎样？如果1952年的起诉从未发生呢？这些不是普通的预测问题。它们涉及不完整的证据、替代的因果路径、不确定的结果，以及将合理的虚构转化为表面历史事实的风险。这就是“反事实自我架构”（CSA）背后的问题。我最近在PhilPapers上发表了该研究框架：“在PhilPapers上阅读CSA论文”(https://philpapers.org/rec/ALHCSA)。

The problem: AI is good at storytelling, but storytelling is not evidence. Give a generative AI model a person’s biography and ask it to imagine an alternative life. It will probably produce a convincing narrative. But what does that narrative actually represent? A historically supported possibility? A plausible extrapolation? An unsupported assumption? Or simply a coherent story generated by a language model? These categories are fundamentally different. A narrative can be internally consistent while being historically or causally unjustified. CSA starts from this distinction: counterfactual coherence must not be confused with counterfactual warrant.

问题在于：人工智能擅长讲故事，但讲故事并不等于证据。给生成式人工智能模型一个人的传记，让它想象另一种生活，它很可能会产生一个令人信服的叙事。但这个叙事到底代表什么？是历史支持的可能性？合理的推断？未经证实的假设？还是仅仅由语言模型生成的连贯故事？这些类别有着本质的区别。一个叙事可以在内部保持一致，但在历史或因果关系上却是站不住脚的。CSA正是从这一区别出发：反事实的连贯性绝不能与反事实的依据相混淆。

CSA: A different way to structure possible selves. Counterfactual Self Architecture is a framework for representing alternative versions of a person’s trajectory. Instead of generating one unrestricted alternative biography, CSA organizes the process around explicit reference information, interventions, and uncertainty. Its architecture distinguishes several important elements: Reference Record (A structured representation of what is known about the subject), Identity and Continuity Constraints (A way to distinguish characteristics that remain stable from those that may change across alternative scenarios), Divergence Points (Explicit interventions that define where an alternative trajectory separates from the reference history), Counterfactual Branches (Alternative trajectories generated from those interventions), and Evaluation and Audit (A mechanism for examining whether the generated scenario is supported, coherent, or speculative). The objective is not to produce the most entertaining alternative life. It is to make the assumptions behind each alternative visible and auditable.

CSA：构建可能自我的不同方式。反事实自我架构是一个用于表示个人轨迹替代版本的框架。CSA不是生成一个不受限制的替代传记，而是围绕明确的参考信息、干预措施和不确定性来组织这一过程。其架构区分了几个重要要素：参考记录（对已知主体信息的结构化表示）、身份与连续性约束（区分在不同场景中保持稳定与可能发生变化的特征的方法）、分歧点（定义替代轨迹从参考历史中分离出来的明确干预）、反事实分支（由这些干预产生的替代轨迹），以及评估与审计（检查生成的场景是否得到支持、是否连贯或是否属于推测的机制）。其目标不是为了产生最有趣的替代生活，而是为了使每个替代方案背后的假设变得可见且可审计。

The critical technical component: ICC. One of the central challenges is deciding what the system actually knows about a person. CSA uses an evidence-status approach called ICC: Status (Preserved: Supported information retained across the scenario; Variable: Information that may change under the intervention; Unknown: Insufficient evidence; Disputed: Credible evidence or interpretations remain in disagreement). The distinction between Unknown and Disputed is especially important. Unknown means that the available information is insufficient. Disputed means that there is a genuine disagreement that should remain visible rather than being silently resolved by the model. This creates an important engineering principle: The system must be allowed to abstain. A model should not be forced to invent certainty simply because a user expects a complete narrative.

关键技术组件：ICC。核心挑战之一是确定系统对一个人到底了解多少。CSA使用了一种称为ICC的证据状态方法：状态（保留：在整个场景中保留的支持性信息；变量：在干预下可能发生变化的信息；未知：证据不足；有争议：可信的证据或解释仍存在分歧）。“未知”与“有争议”之间的区别尤为重要。“未知”意味着现有信息不足；“有争议”意味着存在真正的分歧，这种分歧应该保持可见，而不是由模型默默地解决。这产生了一个重要的工程原则：必须允许系统弃权。模型不应仅仅因为用户期望一个完整的叙事就被迫编造确定性。

Why Alan Turing is an interesting stress test. A natural next step is to examine CSA through a historical case study. Alan Turing offers an unusually interesting example because his biography contains several documented decision points. Consider three possible interventions: Scenario A — Princeton, 1938 (Turing accepts von Neumann’s offer instead of returning to Cambridge. The system can establish the historical decision and explore possible institutional consequences. However, it cannot confidently determine how his wartime contribution or later career would have changed). Scenario B — The ACE project (Turing remains at the National Physical Laboratory after 1948. This scenario raises questions about his potential role in the development of early computing. But the fact that he remained at NPL would not, by itself, establish that he would have completed the ACE project sooner or led it successfully). Scenario C — The 1952 prosecution (The prosecution does not occur. This intervention changes an important part of the historical record. But it does not prove that Turing would have lived longer or that every other institutional consequence would have disappeared). These scenarios reveal why counterfactual modeling needs more than a powerful language model. It needs explicit boundaries around what can and cannot be inferred.

为什么艾伦·图灵是一个有趣的压力测试。下一步自然是通过历史案例研究来检验CSA。艾伦·图灵提供了一个非常有趣的例子，因为他的传记包含几个有据可查的决策点。考虑三种可能的干预：场景A——1938年的普林斯顿（图灵接受了冯·诺依曼的邀请，而不是回到剑桥。系统可以确定历史决策并探索可能的机构后果。然而，它无法自信地确定他的战时贡献或后来的职业生涯会发生怎样的变化）。场景B——ACE项目（图灵在1948年后留在国家物理实验室。这个场景提出了关于他在早期计算发展中潜在作用的问题。但仅仅因为他留在NPL，并不能证明他会更快地完成ACE项目或成功领导该项目）。场景C——1952年的起诉（起诉没有发生。这一干预改变了历史记录的重要部分。但这并不能证明图灵会活得更久，也不能证明所有其他机构后果都会消失）。这些场景揭示了为什么反事实建模需要的不仅仅是一个强大的语言模型，它还需要在可推断和不可推断的内容之间设定明确的界限。

Three extensions worth exploring. The Turing case also points toward three possible extensions of the CSA framework. A. Archival Subject Mode: Most personal AI systems assume that the subject is alive and can provide additional information. Historical figures create a different problem. Their available information comes from archives, public records, correspondence, institutional documents, and historical research. Archival Subject Mode would explicitly distinguish historical subjects from living participants and from research involving deceased human participants. B. Archival Prospective Simulation: A particularly interesting possibility is to simulate a historical person’s future from a specific date using only the information available at that time. For example: What might Turing’s next decade have looked like from the summer of 1950? This requires a strict historical knowledge cutoff. The system must not use his 1952 morphogenesis publication as information available to a simulation supposedly operating in 1950. And if the scenario depicts him working in 1955, it must explicitly.

值得探索的三个扩展。图灵案例也指向了CSA框架的三个可能扩展。A. 档案主体模式：大多数个人人工智能系统假设主体还活着并能提供额外信息。历史人物则带来了不同的问题。他们可用的信息来自档案、公共记录、信件、机构文件和历史研究。档案主体模式将明确区分历史主体与在世参与者，以及涉及已故人类参与者的研究。B. 档案前瞻性模拟：一个特别有趣的可能是，仅使用当时可用的信息来模拟历史人物从特定日期开始的未来。例如：从1950年夏天看，图灵接下来的十年会是什么样子？这需要严格的历史知识截止点。系统不得将他1952年的形态发生论文作为模拟1950年情况的信息。如果场景描绘他在1955年工作，它必须明确说明。