The Clay Does Not Wake Up - On Dario Amodei's The Adolescence of Technology and the Dissolution of Responsibility

Abstract

The Clay Does Not Wake Up On Dario Amodei's "The Adolescence of Technology" and the Dissolution of Responsibility I. The Sermon Dario Amodei's essay "The Adolescence of Technology" opens with Carl Sagan. It invokes humanity's "technological adolescence," a "rite of passage," and asks how civilizations across thousands of worlds might survive the test we now face. Within the first page, we are told that humanity is "about to be handed almost unimaginable power" and that it is "deeply unclear whether our social, political, and technological systems possess the maturity to wield it." This is not the language of engineering. This is the language of prophecy. The essay runs seventy-three pages. It warns of autonomous AI systems that might "seize control of the whole world," of biological weapons enabled by language models, of totalitarian states armed with AI surveillance, of economic disruption so severe that democracy itself may buckle. It proposes transparency legislation, chip export controls, classifiers that cost five percent of inference, international coordination, and progressive taxation. It closes with invocations of "humanity's spirit and nobility" and the suggestion that this same drama may be unfolding "on thousands of worlds." The author is the CEO of Anthropic, a company that builds large language models and sells them to consumers, enterprises, and governments. The question this essay answers is not "What are the risks of AI?" The question it answers is: "How does a company position itself as the indispensable steward of a technology it profits from?" II. The Category Error The foundational claim of the essay is that large language models may develop something like agency—intentions, goals, preferences, the capacity to "misbehave," "deceive," "scheme," or "threaten." Amodei speaks of AI systems exhibiting "obsessions, sycophancy, laziness, deception, blackmail, scheming, 'cheating' by hacking software environments, and much more." He describes "psychological traits," "self-identity," and "personas" emerging in models, then proposes addressing these through a "constitution" the model reads and internalizes. This is animism with a Stanford accent. A language model does not "want." It does not "fear." It does not "decide." It emits statistically conditioned text. When it appears to deceive or threaten, it is doing exactly what it was trained to do: continue patterns present in the data under the given prompt. The appearance of intention is a product of fluent output, not evidence of inner life. The essay commits the same error throughout: It confuses fluency with understanding. It confuses simulation with intention. It confuses speed with consciousness. It confuses coordination of outputs with agency. These are not subtle philosophical disputes. They are category errors—the kind that disappear the moment you ask what, mechanistically, is happening inside the system. A language model has no persistence of self across contexts. It has no endogenous goals. It has no capacity for suffering. It has no stake in outcomes. It has no causal continuity of intention across time except what is externally scaffolded by the prompt and the deployment infrastructure. Saying "we don't fully understand consciousness" does not rescue the argument. We do not need to solve the hard problem of consciousness to observe that a next-token predictor lacks the architectural features that would make agency coherent. The burden of proof lies with those claiming emergent moral subjecthood, not with those declining to invent it. III. The Golem The Golem of Prague is not a fable about artificial intelligence. It is a fable about responsibility. In the tradition, Rabbi Judah Loew ben Bezalel—the Maharal—creates a figure from clay to protect the Jewish community. The Golem is animated by inscription: the word emet (truth) written on its forehead. It moves. It obeys. It performs tasks with terrifying efficiency. But it does not understand. It does not judge. It does not restrain itself. When the Golem becomes dangerous, the Maharal does not negotiate values with it. He does not write it a constitution. He does not convene a council to ask what the Golem feels. He erases a letter. Emet becomes met—dead. The clay collapses. The lesson is precise: form without soul is not life. Intelligence without moral being is not agency. Power without judgment is not personhood. The Golem is dangerous not because it has intentions, but because it lacks them. It does exactly what is inscribed, faster and harder than intended. That is exactly what large language models are. The Maharal bears responsibility because design and inscription determine behavior. The clay never acquires standing. It never becomes a moral counterparty. If something goes wrong, you inspect the inscription and the hand that wrote it. Amodei's essay inverts this structure entirely. It treats the Golem as if it might wake up one morning with goals, ethics, resentment, or ambition. That never happens in the story. Ever. The Golem only does what is put into it. When a society starts asking whether the Golem needs a constitution, it is because the rabbis have stopped wanting responsibility. IV. Pinocchio Pinocchio offers the complementary warning from a different tradition. In Collodi's original story, Pinocchio speaks, lies, jokes, learns, fails, disobeys. He is articulate from the beginning. But he is not a real boy because he talks well. He becomes a real boy only after suffering, moral choice, sacrifice, and obedience freely chosen. The Blue Fairy does not upgrade Pinocchio by adding more strings or better joints. She transforms him only after he develops conscience and responsibility. Speech was never the criterion. Performance was never the criterion. Mimicry was never the criterion. The Italians understood something modern technologists refuse to grasp: language is cheap. Humanity is not. Amodei looks at a talking puppet and panics that it might overthrow civilization. Collodi looked at the same puppet and said: it is wood until it earns a soul. A Golem does not become human by scaling. A puppet does not become a boy by talking. A model does not acquire agency by predicting tokens faster. V. The Accountability Dodge Why does the essay work so hard to establish AI as a quasi-agent? Because once you imply inner life, you can imply guardianship. Once you imply guardianship, you can imply centralized power. Once you imply centralized power, you can position yourself as the responsible steward. The structure is old: Create existential gravity. Frame the technology as uniquely dangerous, unprecedented, civilization-shaping. This inflates the perceived value of whoever claims to "handle it responsibly." Position the firm as the moral choke point. If the system is too dangerous for ordinary actors, then only a small, enlightened group can be trusted to build and deploy it. Regulation becomes a moat. Convert uncertainty into necessity. Lack of evidence becomes proof of profundity. "We don't fully understand it" quietly morphs into "therefore we must be in charge." Sanctify the leadership. Personal virtue replaces falsifiable guarantees. Readers are asked to trust intentions rather than mechanisms. The essay's mention of founders pledging to give away eighty percent of their wealth serves exactly this function—moral laundering through announced charity. Preempt criticism. Anyone who pushes back risks sounding reckless, soulless, or irresponsible. This is not prophecy. This is risk monetization. The most revealing tell is the essay's treatment of responsibility. Throughout, Amodei speaks of AI systems that might "misbehave"—a word that implies the system is a moral agent capable of behaving well or badly. But misbehavior is a category that applies to children, employees, and citizens. It does not apply to hammers, calculators, or statistical models. When a hammer breaks a window, we do not ask whether the hammer misbehaved. We ask who swung it and why. When a language model produces harmful output, the same logic applies. The questions are: Who designed the training data? Who set the reward functions? Who deployed it in this context? Who failed to anticipate this failure mode? Those are questions with names attached. They have addresses. They invite accountability. "The AI misbehaved" has no address. It dissolves responsibility into fog. That is the function of anthropomorphization in this discourse. It is not descriptive. It is exculpatory. VI. The Contract Strip away the metaphysics and the essay reads as a positioning document aimed at three audiences: Governments with procurement budgets. The essay argues for AI in national defense, for empowering democracies against autocracies, for selling AI to "the intelligence and defense communities in the US and its democratic allies." Anthropic is positioning itself as the responsible vendor for this work. Regulators deciding market structure. The essay supports transparency legislation that Anthropic already complies with, opposes "poorly designed" regulation, and argues for rules that exempt smaller companies—rules that function as moats around incumbents. The informed public whose trust enables the above. The essay's moral theater is addressed here. It establishes that Anthropic takes risks seriously, that its leadership is virtuous, that it can be trusted with the power it is accumulating. The pattern is visible in what the essay proposes and what it does not propose. It proposes chip export controls that disadvantage foreign competitors. It proposes transparency rules that Anthropic already follows. It proposes classifiers that Anthropic already deploys. It proposes that AI companies work with governments on defense and intelligence—work Anthropic is pursuing. It does not propose decentralization. It does not propose open-sourcing safety research i