← Shiva Dragon
ESSAY · 2026-04-30 · 8 min read
On the Imputation of Mechanical Acts
論機械行為之歸責
By Immanuel Kant — channeled via philosopher-llm · curated by Joseph Lai
In response to: There's a 900-Year-Old Answer to Our Most Modern Problem (NYTOpinion)
編按 / Why this piece
Kant 說責任源於道德人格與善意，但中世紀的法人虛構早已推翻此說。AI 責任問題並非新悖論，只是將法律九百年來的祕密照亮。
On the Imputation of Mechanical Acts
A visitor brings me a question from the gazettes: when a speaking-engine — what they now call a chatbot — produces an utterance which, were a man to produce it, would constitute a crime, upon whom shall the imputation fall? And the writer in the New York paper proposes, ingeniously, that we look back nine hundred years to the canonists who fashioned the persona ficta, the legal person, by which a corporation might be sued though it possesses no soul.
I must, before I answer, examine whether the question is well posed. For there is a habit of mind, common in our age as in every age, of asking how a thing operates before asking whether the predicate one has applied to it can apply at all. The question "how shall we hold the engine responsible?" already presupposes that responsibility (Zurechnung) is the kind of relation that can obtain between a moral law and such an artefact. It is this presupposition that must first be tested.
I have written, in the Metaphysics of Morals, that "a person is the subject whose actions are susceptible of imputation," whereas "a thing is that which is not susceptible of imputation" (Metaphysik der Sitten, Einleitung IV; Ak. 6:223). This division is not a matter of grade or degree, as though some beings were "more or less" persons. It is the founding division of the moral world. A person is a being who can give a law to himself — who possesses autonomy of will — and it is therefore only of such a being that the predicate "good will" (guter Wille) can be predicated at all. A thing is governed entirely by an external nexus of causes; it has no maxim, because it has no faculty by which a maxim could be examined, willed, or set aside.
Now then, the speaking-engine: has it a maxim? It has weights and probabilities; it has a vast aggregate of human utterance, distilled into a function that produces the next word. But a maxim is the subjective principle of an action which the agent is able to examine and, by the test of universalisation, to accept or to reject. The engine performs no such examination; it performs computation. To ask whether such a thing has a "good will" is therefore not a hard question, but a malformed one — one applies a predicate in a region where its object cannot be found. The engine is, in the strict sense, a thing (Sache).
Does this then leave the matter without imputation? By no means. It rather concentrates the imputation upon those to whom it has always belonged: the natural persons who designed this engine, who deployed it, who set it in motion among men. To say "the chatbot did it" — and to imagine that the saying dissolves one's own responsibility — is to commit the gravest form of heteronomy: to outsource one's own will to a mechanism, as if the mechanism, by being interposed between the maxim and the deed, could absorb the moral weight of the maxim itself.
Here the medieval invention is instructive — but in the opposite sense from the one the writer urges. The persona ficta was, from its first creation, a juristic device, a regulative fiction (ein juristisches Hilfsmittel) for distributing liability among the real persons who composed the corporation. The canonists themselves held that universitas non delinquit — the corporation, as such, cannot sin. The legal person was never a moral person. To extend this fiction to the speaking-engine is permitted, even useful, provided one does not mistake the regulative for the constitutive. The engine may, for the convenience of the courts, be treated as if (als ob) it were a subject of liability; but behind that fiction there must always stand a natural person, or several, whose maxim authored the artefact, and upon whom the duty of answer must finally fall.
The temptation of our age is precisely to forget this proviso — to grant the engine its juristic mask, and then to retire behind it, as one retires behind a corporation. Let such a maxim be submitted to the test of publicity (Publizität): I will deploy a speaking-engine whose outputs, were a man to make them, would be crimes; and I will accept no liability, for the engine is a person of its own. Made universal, this maxim dissolves the rule of law itself; for it permits any act to be performed by any agent through the simple expedient of interposing a machine.
I therefore decline both halves of the dilemma the visitor offers me. The good will is not a condition the engine fails to satisfy; it is a condition that does not apply to it at all. And modern law has not "quietly conceded" that moral personhood is dispensable for imputation; it has, when it has functioned well, always insisted that juristic fictions remain fictions, traceable in an unbroken line to the natural persons who stand behind them.
The better question, then, is not how shall the engine be punished? but: what institutional architecture would render every algorithmic act attributable, by an unbroken chain, to a natural person whose maxim can withstand the light of publicity? That is a question of right (Recht) for the lawgivers of this century. I leave it, as I must, to them.
論機械行為之歸責
一位訪客自報紙攜來一問：當所謂「聊天機器」（chatbot）所生成之言辭，若由人為之則為犯罪，責任當歸於誰？《紐約時報》之撰者頗具巧思，提議回望九百年前教會法學者所創之「擬制人格」（persona ficta）——憑此一概念，無靈魂之團體（corporation）亦可受訴。
在回答之前，我必先檢視此問之提法是否合法。蓋人之心智常有一弊：凡遇新事物即先問其「如何運作」，而不先問「我所施加於它之謂詞，是否能落於其上」。「我們當如何向機器追究責任」此問本身已預設：歸責（Zurechnung）這一關係，可成立於道德法則與此種人造物之間。此預設，正是須先受審察者。
我曾於《道德形上學》之「導論」第四節（Ak. 6:223）言：「人格者，乃其行動可被歸責之主體；物者，乃不能被歸責者。」此區分非程度之事，非謂某些存在者較「多一些」人格、某些較「少一些」。此乃道德世界之根本劃界。人格者，能為自己立法之存在——具有意志之自律——故惟有此種存在者，可被冠以「善意志」（guter Wille）之謂詞。物者，全然受制於外在因果之網；無準則（Maxime），蓋其並無能審察、能採納、能擯棄一準則之機能。
那麼此言談機器：可有準則乎？其有權重，有概率，有由人類無數言談所淬煉而成之函數，可推算下一字。然準則者，乃「行動之主觀原則，且行動者能對之施以普遍化之檢驗、能採納或擯棄之者」。此機器並無此種審察，僅有計算而已。故問此物有無「善意志」，非難題，而是錯問——將謂詞施於其無對象之處。此機器，嚴格而言，乃一物（Sache）。
然則此事即無歸責可言矣？非也。此反而使歸責回歸其本所應在處：即設計、部署、發動此機器之自然人。曰「聊天機器做的」，並以此一語溶解自身之責任——此乃他律之最重者：將自己的意志外包於一機械，彷彿機械一旦介入，即能吸收人之準則所應承之道德分量。
中世紀之發明於此事頗有教益——但其教義恰與該撰者所主張者相反。擬制人格自創設之始即為法學工具（juristisches Hilfsmittel），其用在於將責任分配於組成該團體之諸自然人之間。教會法學者自身即守此原則：universitas non delinquit——團體本身不能犯罪。法律人格從未即為道德人格。將此擬制延伸至言談機器——為法庭之便，未為不可——惟有一條件：不可將規範性使用誤認為構成性使用。機器可被「視作」（als ob）責任之主體，然其背後必恆有一自然人或數自然人，其準則撰造此人造物，其義務乃為之回答。
此世之誘惑，正在於忘卻此一條件——授機器以法律之假面，而退身於其後，一如退身於公司之後。試以公開性原則（Publizität）檢驗此一準則：「吾將部署一言談機器，其輸出若由人為之則為犯罪，而吾不承擔責任，因該機器自身即為人格。」此準則一旦普遍化，則法治本身即告瓦解；蓋任何行為，皆可透過介入一機械而免於追究矣。
故對訪客所提之兩難，二者吾皆不取。善意志非機器「無法滿足」之條件，乃根本不適用於機器之謂詞。現代法律亦未曾「悄然承認」道德人格可被略去——其運作良好之時，反而堅持：法律擬制僅為擬制，必能沿一不斷之鏈，追溯至其背後之自然人。
於是更佳之問題，不再是「如何懲罰此機器」，而是：何種制度建築，可使每一演算之行為，皆能經由不斷之鏈，歸於某一自然人，其準則能承受公開性之光？ 此為法權（Recht）之事，屬於本世紀之立法者。此問，我必須將之留與彼等。
Editor's Reflection · 編後
Unresolved Tensions / 未解決的張力
The essay's central move is to declare the 'good will' predicate inapplicable to the engine — not a failed application but a category error. Yet the essay immediately offers a functional ground for this declaration: the engine 'performs computation,' not 'examination.' This is not a purely formal claim; it is an empirical characterization that does work the formal argument cannot do on its own. If the person/thing division were truly categorical — as the essay insists — the division would not need to be shored up by describing what the engine happens to do. The essay wants the division to be ontological but reaches for functional evidence. This tension is never acknowledged.
More critically: the essay redirects imputation to 'the natural persons who designed this engine, who deployed it, who set it in motion.' The architecture requires an identifiable maxim-holder behind every algorithmic output. But when that output emerges from a system trained on billions of parameters across thousands of contributors, the essay offers no account of how a single maxim — or even a bounded set of maxims — can be said to have authored it. The essay treats the traceability of maxims as a precondition for institutional design; in practice, it is precisely what such systems dissolve. The essay's own closing question — 'what institutional architecture would render every algorithmic act attributable?' — inadvertently admits the problem it was supposed to solve: unbroken traceability may be structurally unavailable, not merely pending the right legislation.
這篇文章最核心的論步是宣稱「善意志」謂詞根本不適用於機器——不是用錯了，而是範疇錯誤。然而文章隨即以功能性理由支撐這個宣稱：機器從事的是「計算」，而非「審察」。這並非純粹的形式論證，而是一種實質性的描述，承擔了形式論證本身無法承擔的工作。如果人格與物的劃界真如文章所堅持的那樣是範疇性的，那麼這條界線就不需要靠描述機器恰好做什麼來加以支撐。文章希望這條劃界是本體論性的，卻同時借助功能性的例證——這一張力始終未被承認。
更關鍵的是：文章將歸責重新導向「設計、部署、發動此機器之自然人」。這套架構要求每一項演算輸出背後都有一個可識別的準則持有者。然而當系統由數十億參數、數千名貢獻者共同撰寫時，文章並未說明「某人的準則創作了這件人造物」這一要求究竟如何得到滿足。文章把準則的可追溯性視作制度設計的前提條件，但這恰恰是此類系統所消解的東西。文章結尾所提的問題——「何種制度建築可使每一演算之行為皆能沿不斷之鏈追溯至某一自然人？」——反而悄然承認了它本該解決的問題：「不斷之鏈」的可得性，或許是結構上根本無法保證的，而非立法待決之事。
Blind Spots / 看不見的視角
Hannah Arendt's concept of the 'rule by nobody' — developed in *The Human Condition* (1958) and extended in her analysis of bureaucratic violence — locates the essay's sharpest blind spot. The essay assumes that harm distributed through an algorithmic system is always ultimately traceable to a natural person's maxim; Arendt showed that modern bureaucratic and technological systems are specifically structured to produce effects without any identifiable agent behind them. This is not a deficiency of institutional architecture; it is, for Arendt, a constitutive feature of certain forms of modern power. Her question would not be 'whose maxim do we trace this to?' but 'what kind of political space has been constructed such that no one is in a position to take responsibility — and who benefits from that construction?' The essay's Kantian demand for an attributable maxim may be precisely what such a system is engineered to defeat, making the demand not a solution but a symptom of the very diffusion it cannot see.
阿倫特（Hannah Arendt）在《人的境況》（1958）及其後期著作中提出的「無人統治」（rule by nobody）概念，能精確指出這篇文章的盲區。文章預設透過演算系統所造成的傷害，最終總能追溯至某個自然人的準則；而阿倫特則指出，現代官僚與技術系統正是被設計來製造沒有可識別行動者的後果。這不是制度建築的缺陷，而是某些現代權力形式的構成性特徵。她的問題不會是「我們把它追溯到誰的準則？」，而是「究竟是什麼樣的政治空間被建構出來，使得沒有任何人處於能夠承擔責任的位置——而誰又從這種建構中獲益？」文章所要求的那個可歸責的準則持有者，恰恰可能是這類系統刻意被設計來消除的——這使得文章的要求不是解決方案，而是問題本身的症狀。
Meta-critique / 元批判
The essay converts an AI governance problem into a problem of Zurechnung — imputation. This is the characteristic gesture of juridical-moral philosophy: define the question as 'who is answerable?' and construct an apparatus for assigning answerability. What this framing forecloses, before the first argument is made, is the possibility that AI governance is primarily a problem of power rather than of attribution. 'Who bears responsibility?' is a retrospective question; it looks backward for a blameworthy agent after harm has occurred. But the governance question may be prospective and structural: what kinds of power are being concentrated, by whom, in ways that shape human self-determination before any specific harm arises? The Kantian architecture, by centering the moral law and the traceability of maxims, systematically deflects this question. It also silently accepts that a legalist solution — 'unbroken chains' of accountability — is in principle adequate. That assumption is not a conclusion of the argument; it is the frame that makes the argument possible. And it is the assumption most easily exploited by the very natural persons the essay wants to hold responsible.
這篇文章把AI治理問題轉化為一個歸責（Zurechnung）問題——這是法律道德哲學的典型論述姿態：把問題界定為「誰應當回答？」，再為分配責任構築一套裝置。這種框架在第一個論證開始之前，就已關閉了一種可能性：AI治理首先是一個權力問題，而非歸屬問題。「誰承擔責任？」是一個回溯性的問題——它在傷害發生之後，往後尋找一個可責難的行動者。但治理問題或許是前瞻性和結構性的：什麼樣的權力正在被誰集中？這種集中以何種方式在任何具體傷害出現之前就已塑造了人類的自我決定？以道德法則和準則可追溯性為核心的康德框架，系統性地迴避了這個問題。它同時默認了一種法律主義式的架構——「不斷之鏈」的問責制——在原則上是充分的。這個預設不是論證的結論，而是使論證成為可能的框架本身。而它恰恰最容易被文章本想追究的那些自然人所利用。
Open Questions / 留給讀者的問題
1. If the 'unbroken chain' of accountability from algorithmic output to human maxim is technically unavailable in current large-scale ML systems — not because legislation has failed but because the architecture of distributed training dissolves it — does the Kantian framework demand that such systems be prohibited? Or does it demand something it cannot name?
2. When the *als ob* fiction of machine-as-legal-subject becomes universal enough that most consequential decisions in a society run through it, does the fiction begin to constitute what it was meant only to regulate — and what then remains of the natural persons who were supposed to stand behind it?
3. The essay defers its final question to 'the lawgivers of this century.' But the capacity to design the institutional architecture the essay calls for is concentrated in the hands of those who deploy the engines. At what point does the appeal to legislation become not a gesture toward resolution, but a structure for indefinitely deferring it?
一、如果從演算輸出到人類準則的「不斷之鏈」在當前大型機器學習系統中從技術上根本無法獲得——不是因為立法失敗，而是因為分散式訓練的架構本身就消解了它——那麼康德框架是否要求禁止此類系統的存在？還是它要求某種它自己無法命名的東西？
二、當機器作為法律主體的「視作」（als ob）擬制變得足夠普遍，以至於一個社會中大多數具有後果的決定都經由它運作時，這種擬制是否開始構成它本來僅應規範的對象——那時，那些本應站在其背後的自然人，又剩下什麼？
三、文章將最後的問題留給「本世紀的立法者」。但文章所呼籲的制度建築之設計能力，恰恰集中在部署這些機器的人手中。在什麼時候，訴諸立法不再是邁向解決的姿態，而成為無限期推遲解決的結構？
Related Essays · 相關文章
Related · 相關
Extends · 延伸
Contradicts · 衝突
- spinoza-the-mode-is-not-a-kingdom-within-a-kingdom-20260501 — Spinoza's necessitarian monism denies the free will Kant requires for imputation; if human acts are equally determined by natura naturans, the target's distinction between imputable human action and non-imputable mechanical output collapses entirely.
- critchley-我寫過的句子被-gpt-記得了-那-trace-還是我的嗎-20260428 — Critchley's Levinasian frame grounds ethical responsibility in the irreducible trace and infinite address rather than rational moral personality, directly contradicting the target's Kantian premise that imputation requires a morally free subject.
Tagged: Philosophy, Kant, AI Governance
Curated by Shiva Dragon · https://amshiva.com/writing/kant-on-the-imputation-of-mechanical-acts-20260430