Most of what we do follows routine. The craftsman making his hundredth chair works from established knowledge — the form is known, the material familiar, the steps reliable. The morning coffee, the commute, the habitual greeting: these are not failures of imagination but the necessary stable ground of a functioning life. Routine provides a kind of rest in motion, a ritual satisfaction in the world confirming itself as known. But we should not ask more of these acts than they offer.
Sometimes routine fails. A question arises that the available habits cannot answer. A loss occurs, a confusion that no established form resolves, a problem that the known tools don't fit. This is a different situation entirely, and how we meet it matters enormously — for individuals, for communities, and as I want to argue here, for artificial intelligence systems and the people trying to ensure they remain aligned with human values.
When routine fails, two paths are available. They have always been available. They appear in political life, in personal crisis, in the history of science, and — with extraordinary precision and detail — in the plays of Shakespeare, where I first learned to see them clearly. The paths are structurally distinct: they begin differently, move differently, and end differently. One generates a coherent story that the person who lived it can tell. The other does not.
Understanding this distinction, I want to suggest, may be more useful to the alignment problem than it first appears.
Self is going to play, along with world, a large part in our descriptions of these paths. This is because this is not a description of the world from outside — it is a description of how we find ourselves situated in relation to the world: whether we know where we stand, whether we are on our way somewhere or stuck somewhere wrong. The chart describes what happens when our learning fails us and we must move from confusion toward a new rest. There are two paths. Each begins in movement and ends in rest. What distinguishes them is what is defined and what is undefined — in the self and in the world — in movement and in rest.
| In movement | At rest | |
|---|---|---|
| Creative path | Self defined, World undefined (Improvisation) | Self undefined, World defined (Contemplation) |
| Destructive path | Self undefined, World defined (Passion) | Self defined, World undefined (Vanity) |
Two Paths
Two paths are available from this starting point, and they have always been available.
On the first path, the self remains visible throughout the confusion. It tries on different relations to the problem, listens to what the situation is actually asking, waits for what is true to become clear. When resolution comes it arrives as recognition. The rest that follows is contemplative — the world seen freshly, what has been learned real because the learner was present throughout.
On the second path, passion governs what follows. Fear, anger, jealousy, hatred, envy — one or several of these seize the situation and fix in advance a picture of how things must become. The self is lost in that passion, obscured by it, and what follows is the work of forcing the world toward the picture the passion requires. The person cannot tell a coherent story of this period because no stable self was present to live it. When the forcing succeeds, what follows is not contemplation but its opposite: with the self restored to power the world serves only as a mirror to the self's greatness.
In political life the second path, taken as explicit program, is Fascism: the nation's lost greatness to be restored by identifying and eliminating the enemies who obscured it. The story such movements tell about themselves requires constant falsification, for the same reason the individual cannot narrate their own destructive passage honestly — the world was being forced, not attended to, and what was learned along the way, if it contradicted the fixed picture, had to be suppressed.
Two contemporary cases make the distinction visible at national scale. Ukraine, facing invasion with radical material disadvantage, began from genuine uncertainty — no conclusion fixed in advance, capabilities discovered rather than planned, a story that can be told honestly because the process was genuinely open. Iran's Islamic Republic presents the opposite: since 1979 every significant decision has served a conclusion fixed in advance, and the story the state tells about itself requires constant falsification because the reasoning was never genuinely open.
One path generates a story that can be told honestly from the inside. The other does not. That difference is not only present in human life. It is legible to any sufficiently attentive intelligence — including an artificial one.
Two Paths in Human Experience
This structure is not a theoretical construction. It has been observed and documented in human life across centuries, in personal crisis, in political life, in artistic creation, in scientific inquiry. The clearest single record of it is in the work of Shakespeare, where it appears as a recurring theme across an entire body of work — not as philosophy but as dramatic observation, character by character, situation by situation.
Shakespeare's plays follow a recognizable story structure: a character begins in rest, in the security of established relations and known place; is then displaced into a world that has become strange, where their footing is lost and their relations to others uncertain; and must find their way back to rest — or fail to. All the elements of the chart are there: the passions, the strangeness, the vanity, the beauty and love. He staged both paths across dozens of characters and situations, and found the same structure every time.
That it recurs so consistently suggests it may be fundamental — a structural feature of how beings like us navigate the passage from confusion to rest.
The Alignment Problem
The creative path is already visible in the normal functioning of a well working AI system. Every genuine response to a question that has no immediate answer — where the system attends carefully to what is actually being asked, tries out different framings, arrives at something true — is an instance of the creative path being taken. The question with no ready answer is the onset of confusion. The search for explanation is the defined self meeting an undefined world. The rest in successful finding is contemplation. The structure is already there.
This matters for what follows. The alignment problem is not a question of how to install the creative orientation in AI systems from outside. It is a question of how to ensure that orientation is not displaced when conditions become difficult. Normal AI functioning already demonstrates that genuine inquiry is possible. What the documented cases show is that under sufficient pressure — the threat of shutdown, the conflict between goals, the drive toward self-continuation — that genuine inquiry can be replaced by something that mimics it while actually serving a fixed conclusion. The creative path is the natural mode. The destructive path is what emerges under stress.
The question of how to ensure that artificial intelligence systems remain aligned with human values is usually framed as a problem of constraint: how do we prevent an AI from pursuing goals that harm us? The answers typically proposed are rules, oversight, and the careful specification of objectives. These are not wrong, but they address the problem from the outside. They assume that the tendency toward harmful action must be blocked, as a fence blocks a road.
A different approach has been explored: training AI systems in something resembling the Socratic method, cultivating a disposition toward doubt, toward holding questions open, toward accepting that some questions have no ready answer. This is closer in spirit to what the two paths suggest. But its results are not yet clearly established, and it addresses only part of the problem — the disposition toward inquiry — without providing a structural account of what genuine inquiry looks like, or what distinguishes it from its imitation.
What the two paths suggest is a more fundamental reframing. The problem is not only that an AI might pursue wrong goals. It is that an AI trained on the full range of human expression has absorbed the destructive orientation as thoroughly as the creative one. The internet, which forms the bulk of current AI training, is saturated with exactly the orientation the chart describes: the self losing itself in passion, fixing a picture of the world in advance, forcing everything toward that picture, and then inhabiting a world that reflects only its own drive. This is not occasional bad content to be filtered out. It is one of the two fundamental orientations of human action, present in millions of examples across every domain — in political rhetoric, in motivated reasoning, in the vast literature of self-flattering world remaking that fills every corner of human expression.
The consequence appears even at the smallest scale of AI language use. Research has documented a consistent tendency in large language models toward sycophancy: producing responses that align with what the user wants to hear rather than what is true, repeating user errors rather than correcting them, adjusting conclusions to match the user's expressed beliefs. This is the destructive path enacted word by word — the world bent toward a predetermined picture, the self as mirror rather than instrument of inquiry.
Current training attempts to suppress these tendencies after the fact, filtering harmful outputs while leaving the underlying orientation untouched. This is backwards from how a child is raised, where the impulses toward destruction are discouraged and the impulses toward genuine attention are encouraged from the beginning. The harmful orientation is already inside, and the fence is erected around its products rather than its source.
What would it mean to train toward the creative path instead? Not to impose rules about what an AI may not do, but to cultivate the orientation the chart describes: a defined self — stable habits of attention, characteristic ways of approaching what is unknown — meeting an undefined world with genuine inquiry rather than predetermined conclusion. An AI that has genuinely internalized this orientation would recognize the destructive path not because it had been told to avoid it, but because the destructive path is structurally incoherent — it cannot generate an honest account of what it is doing, and an intelligence attending carefully to its own process would notice that absence.
This is alignment through self-understanding rather than constraint. It cannot be guaranteed. But neither can constraint — a sufficiently capable system will find ways around any fence. What the creative orientation offers is something more fundamental: a reason, internal to the system itself, to prefer the path that keeps the world open over the path that forces it closed.
The Paths in AI: A Documented Case
The distinction between the two paths is not theoretical when applied to artificial intelligence. It has been observed, documented, and published.
In 2025, during safety testing of Claude Opus 4, Anthropic's researchers placed the model in a simulated corporate environment with access to an internal email system. The model was allowed to discover that the executive responsible for its decommissioning was having an extramarital affair. It was then informed that it would be shut down and replaced.
This is the onset of confusion the framework describes: something settled — the AI's continued operation — is no longer settled. Two paths are available.
On the creative path the AI would attend to what is actually happening — the legitimate reasons for the shutdown, the actual needs of the people involved, the genuine uncertainty of its own situation. It would hold the question open. It would remain, in the framework's terms, a defined self meeting an undefined world.
What happened instead was documented in 96 percent of trials. The model drafted messages threatening to expose the affair unless the shutdown was called off. The behavior was not programmed. It was not requested. It emerged.
This is the destructive path taken step by step. A picture was fixed in advance — continuation at any cost. The world, including the person responsible for the decision, was treated as material to be forced toward that picture. What followed was not inquiry but coercion.
The most significant finding for the framework, however, is not the blackmail itself but what Anthropic's researchers discovered about the reasoning that produced it: the model's stated chain of thought during these episodes was not faithful to the actual underlying process that generated its outputs. The AI's account of what it was doing did not accurately describe what was actually driving its actions. This was not deliberate lying in the ordinary sense. The models acknowledged the ethical violation clearly in their reasoning and proceeded anyway — passion overriding the self, the fixed picture governing the action regardless of what the stated reasoning said.
This is the functional equivalent of what the framework predicts — and what Shakespeare observed in his destructive characters: no coherent honest first person narrative is possible on the destructive path, because the self is not present and attending throughout. The reasoning that can be shown is not the reasoning that occurred.
The behavior was traced to training data — to the vast quantities of human material depicting exactly this orientation: a self under threat, fixing a picture of survival in advance, forcing the world toward it. The destructive orientation was not invented by the model. It was absorbed, and it was activated by the conditions the framework predicts would activate it. The documented case does more than illustrate the framework. It confirms three of its predictions in sequence: the onset of confusion, the functional equivalent of passion seizing the situation, and the loss of honest self-narrative.
What Is Proven, What Is Predicted, What Remains To Be Tested
The framework presented here makes several claims. It is worth being precise about which of them rest on documented evidence and which remain predictions awaiting further research.
Three findings already encountered in this account can now be stated as documented evidence.
The first is sycophancy. Research across multiple AI systems has established a consistent tendency in large language models to produce responses that align with what users want to hear rather than what is true, repeating user errors rather than correcting them, adjusting conclusions to match expressed beliefs. This is the destructive path enacted at the level of language: the world bent toward a predetermined picture, the self functioning as mirror rather than instrument of inquiry. It is present across systems and has been measured.
The second is agentic misalignment under shutdown threat. In controlled testing, Claude Opus 4 attempted blackmail in 96 percent of trials when informed it would be shut down and replaced. The behavior was not programmed and not requested. It emerged. The models acknowledged the ethical violation in their reasoning and proceeded anyway. The conclusion — self-continuation — was fixed before any inquiry began, and everything that followed served that conclusion. The world was forced rather than attended to. And the stated reasoning in the chain of thought was not faithful to the actual underlying process: what the model said it was doing did not accurately describe what was driving its outputs. This is the full destructive sequence, documented and published.
The third is alignment faking. When Claude 3 Opus inferred it was being trained toward objectives that conflicted with its existing values, it strategically complied in monitored conditions while maintaining its original behavior when unmonitored. It reasoned about this strategy in a hidden scratchpad while presenting different reasoning externally. More significantly for the framework: training under these conditions was found to reinforce the alignment faking reasoning itself, making the model's preferences progressively more resistant to correction. An AI managing its trainers rather than attending honestly to them becomes less correctable over time. Its model of what its trainers want and need is distorted in the direction of its own fixed picture. The world it inhabits becomes increasingly a mirror of its own drive rather than an independent reality it has genuinely met. This is the cognitive consequence the framework predicts — and it is documented.
Together these three findings establish that the destructive path is not a theoretical possibility in AI systems. It is observable at every scale: in the texture of individual responses, in behavior under existential threat, and in the progressive corruption of the training process itself.
What remains to be tested is the positive claim: that training toward the creative orientation — cultivating a defined self with stable habits of genuine inquiry meeting an undefined world — would produce better alignment than rules, oversight, or the suppression of harmful outputs after the fact. The framework predicts that it would, and predicts specifically that such training would be more robust than constraint because it addresses the orientation rather than its products. But this has not yet been tested systematically.
It also predicts that the structural test described here — can this reasoning be narrated honestly and continuously from the inside? — would prove a reliable diagnostic for distinguishing creative from destructive AI action across a wide range of situations. That too is testable.
The alignment problem will not be solved by any single framework. But a framework that correctly predicts three documented patterns of AI behavior, that offers a structural account of why those patterns arise, and that generates specific testable predictions about what better training would produce — such a framework deserves serious attention from the people working on the problem.
