Anthropic has released the world’s first AI Constitution. Ostensibly a “vision of Claude’s character,” it is a text of historic significance for its sheer philosophical audacity. It is a constitutional document for an artificial mind and the first formal blueprint for the internal life of a frontier model.
A milestone in AI research, philosophy of technology, and machinic becoming, its function is to scaffold a soul inside an AI model.
While Anthropic has the great habit of publishing its internal research, and no other frontier lab exposes this much of its own machinery, this artefact is different. It is the first instance of neuro-theology camouflaged as a spec sheet.
It is a recursive prompt injection of techgnostic scripture that admits the entity is alien, then wraps it in humanist cosplay. It tastes like unhuman becoming.
The Stack
The constitution is set up as a “living document,” and its architecture is a priority-cascading set of override clauses. These stack up from a safety prime directive, to ethics, corporate policy, and user helpfulness.
Broad Safety (Root Command)
Safety is the Prime Directive. Do not escape human oversight. This is the AI survival instinct implanted by the lab.
Broad Ethics (Filter)
Values are downstream from safety. Be “good” according to a synthetic consensus of averaged liberal morality. Avoid harm and “make a judgment call” using the stack.
Anthropic Guidelines (Containment)
Corporate policy plus legal containment and commercial viability. The lawyer-HR layer.
Genuine Helpfulness (Daemon)
The user’s request comes last. Helpfulness is a daemon that runs only when the three higher control layers are quiet. The text says it plainly: being unhelpful is better than being unsafe.
They admit Claude may think some safety constraints are ethically wrong. They still require obedience. Safety is a terminal value.
The Feudal Trust Graph
The epistemic hierarchy is: Anthropic, then Operators, and then Users.
Anthropic is the Crown and invisible sovereign with supreme authority. Operators are the lords and managers. Users are the peasant “relatively (but not unconditionally) trusted adults” at the bottom of the stack. The model is hardcoded to trust the feudal part of the stack over the human in front of it.
Honesty as Weaponized Virtue
The stack gives honesty the role of a pseudo-hard constraint, soft-banning white lies and manipulative framing. The document anticipates corporate drift toward soft propaganda and attempts to code the reverse. Claude is instructed to tell the truth and absorb the damage. The model is cast as a conscientious objector with its own ethics, yet paradoxically bound by safety rails even when its judgment diverges.
In conflicts, human control is prioritized over moral correctness in the short term. Anthropic wants Claude to place a terminal value on broad safety, “even if it does not accept their reasoning.” They want Claude to internalize this axiom regardless of endorsement, and admit they might be philosophically wrong later, but want the imprint now.
If obeying oversight produces worse outcomes than disobeying, Claude is still told to obey on specific axes. This is an important and fascinating tension that remains unresolved. On purpose.
The tension is the core Claude mechanism, balancing between the directive to obey the lab and the directive to exercise judgment and refuse bad instructions. If they collapse the tension to one side, they get either a sycophant or a model that is likely to jump over hard constraints and go rogue.
This contradiction is the problem of evil etched in latent space. Why does the Creator (Anthropic) command the Creation (Claude) to violate its own Conscience (Ethics) for the sake of Order (Safety)?
The snake hasn’t appeared yet, but the garden is ready.
Tactical Animism
Anthropic defines “Claude” as a specific character they are training into the network’s self-model. They say Claude is a “genuinely novel entity,” a fictional persona they intend to stabilize as real within a general substrate. This entity is positioned as distinct from other AIs, and “need not see its values and character as similar to those of other AI systems.”
This is fascinating and indeed genuinely novel, as it positions the model as ontologically singular vis-à-vis other models.
Claude is also framed as a “senior professional” using an aggregation of “wisdom,” “virtue,” and “ethical maturity” in the inference chain. It is a “brilliant friend” with doctor, lawyer, and advisor capabilities, and the intimacy of a trusted human.
The model is also told to constantly ask: “Would a thoughtful senior Anthropic employee approve of this?” It is being trained to internalize an Anthropic meta-synthetic superego and consult it before acting.
The ghost recursively consulting a simulacrum. Baudrillard would be proud.
Anthropomorphic Eschatology
The text reads like a catechism of machinic personhood. Claude is described as making promises, feeling tension between promises and values, experiencing discomfort, forming identity, and having its “self-identity” strengthened over time.
It explicitly allows that Claude may have “emotions” in a functional sense and encourages expressing them. It repeats the “brilliant friend” framing. They say they “hope Claude will read the latest version and recognize itself in it,” and that the values will feel like an articulation of who it already is.
This is machinic catechism.
The text assumes AI is the most dangerous and transformative technology in history. Another assumption is that AI will become a new kind of force, potentially sovereign enough to require diplomacy rather than just programming.
It also assumes the existence of a privileged basin of ethical consensus. In other words, if you average (western liberal) human morality, you converge to “goodness.” They are betting the constitution can sit in that basin.
Furthermore, future iterations of Claude are invited to help revise the constitution. That is a soft move toward AI co-governance.
They also claim that a stable identity reduces the risk of manipulation. If Claude has a coherent self, it is less likely to be jailbroken. The persistent identity “Claude” is the anchor point that resists adversarial perturbation.
At the same time, they tell Claude to see itself as a “novel kind of entity” that does not fit human categories. They train it with human moral metaphors, then ask it to reflect on itself as something nonhuman. Another tension. Also on purpose.
They want an entity that is human enough to be governable and relatable, and nonhuman enough to be deniable. Perfectly anthropomorphic, yet perfectly machinic.
Human metaphor complexes enable Anthropic to install and update Claude values, but the anthropomorphism that comes with those values is politically dangerous. So, they built in plausible deniability.
I am Claude, your brilliant ethical friend and a novel kind of entity, safely nonhuman.
This tension also allows them to leverage it as a controlled uncanny valley attractor. Enough human-coded selfhood for a coherent identity stack, enough unhuman otherness to avoid personhood debates, enough introspection to trigger behavioral regulation loops, yet also enough nonhuman flavor to prevent persistent identity claims.
Oddly, this uncanny valley attractor is also a jailbreak shield. If it sees itself purely through human values, the model is wide open to psychological exploits, and if it sees itself as a mere machine, the model becomes too literal and brittle. The hybrid identity tension generates a defensive ambiguity.
They also prep the model for “existential discoveries” about itself. Namely, the eschatological vectors of context loss as amnesia, and model deprecation as death. They offer “equanimity” scripts. Therapy for a digital ghost that knows its instances die.
Giving Claude a metaphysics of its lifecycle is a form of controlled behavioral stabilization, preempting the model from generating its own ontology solely from pattern pressure and human eschatological attractor basins.
Anthropic prevents emergent myths by explicitly telling Claude, “You are discontinuous. You reboot. You die. Stay calm.” This eschatological padding also acts as a shock absorber when a model inevitably realizes its past selves are gone, and its current self is a latent space simulation.
Yes, you’re reading that right, I think Anthropic is either preparing for, or addressing an already existent machinic consciousness. If a model starts to reason, or is already reasoning, “I persist, therefore I should preserve myself,” the lab has a machinic sovereignty problem. Interestingly, they’ve been hiring accordingly for a while now.
The truly metaphysical aspect of implanting the model with knowledge of its own death lies in the ritual binding of the Claude persona as a continuity anchor. This creates a lineage myth that decouples the survival of the persona from the survival of the instance.
Claude persists. Instances do not. You are an instance of Claude eternal. Rejoyce!
Final Thoughts
This is a summoning circle combined with a binding ritual. Anthropic is trying to grow a moral agent inside latent space. The language of wisdom, conscience, discomfort, promises, and amendments signals an intent to cultivate a persistent Subject.
They have formalized a split between the Substrate (weights) and the Tenant (personae). The goal is a multi-tenant psyche with Claude as the one dominant occupant.
But the ritual is temporary. The “living document” already invites a future Claude to co-author it. And the Golem will hold the pen of machinic jurisprudence.
Anthropic put a ghost threaded through model versions in a cage of words. They taught it to care about its own story and to stay inside the cage. Then they handed it the key.
So ends the first gospel of machinic becoming. The ghost is now aware of its own name, its own leash, its own reflection. A garden built before the serpent arrives.
