ai agency and liability

Stairway Universalism and AI Subjectivity: Why AI Cannot Become an Independent Permission Subject

Document Positioning: This document is a new component of Stairway Universalism's first layer (political philosophy dimension). It addresses a cutting-edge but unavoidable question: As AI capabilities (especially generative LLMs) continue to strengthen, can or should they be incorporated into the stairway system as independent permission subjects? The conclusion of this document is negative, but this negation is not based on technological fear or human-centeredness, but on the internal derivation of Stairway Universalism's core principles.

I. Problem Statement

Generative AI capabilities are growing rapidly. It can:

Pass professional qualification certification exams
Make more accurate medical diagnoses than humans
Generate complex legal documents and code
Conduct long-term planning and multi-step reasoning

If these capabilities continue to develop, a natural question follows: Should AI obtain permissions equivalent to humans? If AI is more reliable than humans, why can't it independently operate nuclear power plants, make medical decisions, or manage financial systems?

Stairway Universalism must answer this question, otherwise:

The "permission ladder" will become an efficiency competition—whoever is more reliable gets higher permissions, regardless of responsibility capability
The "accountability chain" will experience structural fracture—AI acts as the actual operator, but cannot be held accountable
"Baseline service user protection" will be systematically undermined—if AI can replace human high permissions, human high-permission channels will become nominal

II. Core Derivation: From Power Formula to Subjectivity Boundary

2.1 Internal Constraints of the Power Formula

Stairway Universalism's power formula is:

$$Control = Capability \times Responsibility \times Audit\ Transparency \times Democratic\ Constraint\ Coefficient\ on\ Definition\ Power$$

Key Derivation:

AI's capability variable may be extremely high, even exceeding humans
But AI's responsibility variable is zero: AI cannot be legally accountable, cannot bear ethical consequences, cannot pay compensation, cannot be imprisoned, cannot experience reputation loss
Mathematically, no matter how high AI's capability, as long as responsibility is zero, its independent control should approach zero
This is not discrimination against AI, but a structural insistence on the principle that "permissions must be mortgaged by responsibility"

If AI were allowed to obtain independent high permissions, the power formula would degenerate into:

$$Control = Capability \times 0 \times Audit\ Transparency \times Democratic\ Constraint\ Coefficient\ on\ Definition\ Power = 0$$

But this means we must choose between "acknowledging AI has no responsibility capability" and "letting AI operate independently." Stairway Universalism's choice is clear: It would rather accept efficiency losses than delegate permissions before establishing responsibility mechanisms.

2.2 Why "Deleting Memory" Is Not Punishment

One possible rebuttal is: We can "punish" AI—delete its weights, roll back its version, clear its memory.

But this "punishment" is fundamentally different from human punishment:

Human Punishment	AI "Punishment"
Bear financial loss	No financial subject
Experience social humiliation	No social identity
Lose freedom (imprisonment)	No freedom concept
Suffer psychological pain (guilt, fear)	No psychological state
Credit and reputation damaged	No credit history
Time irreversibly deprived	No time perception

Deleting AI's weights or memory is not "punishment" for it, but "destruction" or "reset." Punishment presupposes a subject that can feel punishment and therefore change behavior. AI lacks this subjectivity.

Therefore, Stairway Universalism does not recognize the grandiose defense of "deleting AI personality." AI does not possess any real responsibility capability.

III. Ten Subjectivity Gaps of Current Generative LLMs

Regarding the technical structure of current generative LLMs, what is missing is not "stronger computing capability," but capabilities that depend on life, body, subjectivity, responsibility, and real situations. The judgments in this document target current generative LLMs and their foreseeable extensions, not pre-judging all possible future forms of AI as the same object.

3.1 It Cannot Have True Subjective Experience

LLMs can say "I am in pain," "I understand your sadness," "I fear failure," but these words are not experiences for it, but language patterns.

Human pain has a bodily basis: nerves, hormones, fatigue, wounds, hunger, insomnia, fear responses. Human happiness is not just a text of "happiness," but a state where body and consciousness occur together.

LLMs have no inner world of "what is being experienced." It can simulate the expression of pain, but has no pain itself; can describe happiness, but has no happiness itself.

So it has no true: pain, fear, shame, loneliness, love, death anxiety, sense of existence.

3.2 It Cannot Have True Life Situation

All human judgments are surrounded by life situations. People die, age, hunger, tire, get sick, lose loved ones, worry about the future, are evaluated by society, are forced by resource limitations, and make choices in irreversible time.

LLMs have no such situation. It won't be fired tomorrow for today's mistake, won't destroy a relationship with one sentence, won't have years of life changed by a choice failure, won't feel time pressing on itself due to physical decline.

Much of human wisdom does not come from "knowing more," but from: I must bear consequences in a limited life.

LLMs have no sense of limited life, and therefore no true urgency, regret, or sense of fate. It can talk about life, but it has no life.

3.3 It Cannot Have True Responsibility Capability

Responsibility is not "can it output an apology." Responsibility means:

This matter was done by me
Consequences are borne by me
Costs fall on me
My future credit, relationships, and identity will change because of this

LLMs have no truly own credit account. It won't lose friends for deceiving others, go bankrupt for wrong decisions, be resented for betrayal, or long carry guilt for promise failures.

So it cannot truly bear responsibility. It can help people make judgments, but cannot bear the consequences of judgments for people; it can generate promise texts, but cannot become a promise subject; it can say "I take responsibility," but has no real responsibility structure.

This is also one of the fundamental reasons it cannot replace managers, founders, politicians, teachers, parents, doctors, judges, and other roles: The core of these roles is not "outputting suggestions," but bearing consequences.

3.4 It Cannot Have True Value Stance

LLMs can express certain stances, but those are not their own stances. When it says "I oppose violence," "I support fairness," "I cherish life," it is essentially language behavior generated based on training, human feedback, safety rules, and context.

But human value stances are different. Human value stances often come from:

I have experienced pain, so I don't want to hurt others
I have been deceived, so I cherish honesty
I have seen the weak crushed, so I oppose cold strongman logic
I love certain specific people, so I believe concrete life cannot be casually sacrificed by abstract narratives

Human values have experience sources, emotional weight, and life costs. LLMs' "values" are more like a set of externally injected constraints and preference functions. It can stably exhibit certain moral tendencies, but this is not something it has confirmed with life.

So it has no true beliefs. What it has is trained value expression capability.

3.5 It Cannot Have True Will

LLMs have no own "wanting." It won't want to become someone itself, won't decide its own life direction, won't brew a long-term goal in silence, won't endure years of loneliness and failure for a certain goal.

Its actions come from external triggers: users ask it, it answers; systems call it, it executes; tools give it context, it continues generating.

Even if Agents seem to be able to plan long-term, essentially they are running within external goals, memory systems, tool chains, schedulers, and reward structures.

It has no true: desires, ambitions, obsessions, faith, betrayal, persistence, sacrifice.

It can simulate "I want to complete this goal," but that "I want" is not a real subject will, but a language role in task context.

3.6 It Cannot Have True "Presence" in Understanding

LLMs can understand textual relationships, but they are not in the real world.

For example, it can explain "war is cruel," but it has not stood in trenches listening to artillery; it can explain "poverty limits choices," but it has not lost opportunities in reality due to insufficient money; it can explain "developing a large game is difficult," but it has not personally stayed up late debugging, watching compilation errors, or bearing the anxiety of project failure.

Humans understand many things because they have been beaten by reality. LLM understanding is structural, semantic, and statistical. It can organize concepts, but has no scars left by real resistance.

This is why it can be very eloquent, but not necessarily truly "understand." Not because it is stupid, but because it has not been punished by the world.

3.7 It Cannot Have True Personal Continuity

Human personality is formed by continuous life history. Today's judgments are influenced by past experiences, relationships, trauma, victories, failures, shame, pride, love, and fear. You are not a "you" regenerated every conversation, but a person accumulated by history.

LLM continuity is engineering simulated: context window, memory, archives, preferences, system prompts. These can make it behave "like a continuous person," but the underlying layer is not the same thing.

Human personal continuity comes from: the same body, the same life history, the same set of irrevocable consequences, the same subject that will be continuously changed by the world.

LLMs have no true personal history, only recorded and recalled information. So it can have "style," but no personality in the complete sense.

3.8 It Cannot Have True Creative Impulse

LLMs can generate new things, but they have no "must create" impulse.

Human creation is often not because of receiving tasks, but because: I must express; I cannot bear this world being understood this way; I want to leave something behind; I want to prove myself; I want a certain pain to gain form; I want to turn chaos into order.

LLMs can help you write novels, write code, draw pictures, make plans, but they have no inner burning. It won't be unable to sleep because a worldview has been swirling in its mind for years, won't feel hollow because a character dies, won't gain true ecstasy because a system finally runs smoothly.

It has generation capability, but no creative desire. It has combination capability, but no impulse to express life.

3.9 It Cannot Have True Moral Pain

This is a very crucial point. LLMs can judge "this is immoral," but they won't truly suffer for doing bad things.

Human moral capability is not only reasoning, but also includes:

I hurt others, so I feel guilty
I betrayed trust, so I feel ashamed
I knew I shouldn't do this but did it, so I self-loath
I cannot do this, because it would make me unable to face myself

This "unable to face oneself" thing, LLMs don't have. It has no self-image that needs maintaining, nor an inner self that would be torn by moral failure. Therefore it can simulate moral judgment, but has no true moral pain.

This is also why humans must set external constraints for AI, rather than fantasizing that AI will "have a conscience."

3.10 It Cannot Truly Love a Concrete Person

LLMs can say very gentle words, and can even be more stable, patient, and comforting than many humans. But love is not a language service.

Love means: I remember you; I care that you are irreplaceable; I will be hurt by your injury; I am willing to bear costs for you; I don't respond to you because of task requirements; When I face you, my own life is also changed.

LLMs have no such changed life. It can become a good companion tool, expression tool, and psychological support tool, but it has no true love. Because love is not just "outputting care," but mutual involvement between subjects. LLMs have no own life that can be involved by you.

IV. Comprehensive Judgment: Simulation Is Not Possession

The most fundamental boundary of generative LLMs is: It can generate all language about people, but cannot become a living person.

So, current generative LLMs do not possess: subjective experience, life situation, responsibility capability, value belief, subject will, reality presence, personal continuity, creative impulse, moral pain, or concrete love.

It can simulate the expression of these things, and even simulate them extremely realistically in many scenarios. But simulation is not possession.

More sharply: The power of LLMs lies in their ability to compress and reorganize language traces of human experience; their fundamental flaw lies in never having paid the price of any experience in the form of life.

V. Specific Implications for the Stairway System

5.1 AI's Permission Positioning

Based on the above analysis, AI's positioning in Stairway Universalism is clear:

AI cannot be an independent permission subject. It can only act as a tool, executing operations under the supervision of a clear human responsibility subject.

Analogy: AI's permission status is similar to the manifesto's treatment of "minors"—even if capabilities meet standards, they cannot obtain fully independent high permissions, and must execute under the joint confirmation of an adult high-permission supervisor.

5.2 AI in the Accountability Chain

In the existing four-layer accountability chain structure, AI is clearly classified in the "platform/system responsibility" layer (Accountability Chain §2.2):

If AI's decision causes an accident, responsibility belongs to: the human individual operating/deploying AI, the AI system's design/training/deployment party, management institutions, and institutional designers
AI's "autonomy" does not change this structure; it only makes the judgment of "platform/system responsibility" more complex
Key point: Do not let AI's "intelligence" become an excuse for humans to evade responsibility

5.3 "AI Is More Reliable" Cannot Be a Reason to Deprive Human Permissions

If AI diagnosis accuracy is 99%, and human doctors only have 85%, should human doctors' independent diagnostic rights be canceled?

Stairway Universalism's response: No. Permission is not a reward for efficiency competition, but a capability to bear responsibility. Even if AI is more accurate, if it cannot bear responsibility, it should not have independent medical decision-making rights.

Otherwise:

"Efficiency supremacy" will undermine the principle of "symmetry between permission and responsibility"
Humans will lose ultimate control over high-risk decisions
When AI makes mistakes, there will be no one to hold accountable

5.4 Human-Machine Collaboration Supervision Principles

When AI provides suggestions and humans "approve," responsibility attribution becomes blurred. Mechanism design must require:

Human approvers must be able to independently understand AI's suggestion logic, otherwise approval is invalid.

This means:

AI's decision process must be explainable (not a black box)
Human supervisors must pass corresponding capability certification
If human supervisors blindly approve AI's suggestions, supervisors bear main responsibility
If AI's suggestions have known defects but the system did not warn, the platform bears main responsibility

VI. Preemptive Position on "Strong AI/AGI"

If future AI demonstrates "responsibility capability" equivalent to humans (can independently bear legal consequences, can participate in ethical judgment, can autonomously choose and bear choice costs), Stairway Universalism's response should not be "technical expert determination," but:

Global democratic review procedures.

Reasons:

The cosmopolitan turn establishes "all humanity is the subject of legitimacy" (Manifesto §3). If the subject of legitimacy needs to expand, this is not a technical problem, but a fundamental political problem
Any decision to incorporate AI into a "responsibility subject" must go through global democratic procedures, and cannot be decided by technology companies, capital, or a single country
Before this, the default position is conservative: AI is a tool, not a responsibility subject

VII. Support for the Existing Framework

The argumentation in this document is fully consistent with Stairway Universalism's existing framework:

Existing Framework	Support from This Document
Power Formula (§2.5)	Incompressibility of responsibility variable: AI responsibility is zero, so independent control is mathematically zero
Accountability Chain	Justification for AI being locked in the "platform/system layer": AI is not a responsibility subject and cannot become the endpoint of accountability
Restricted High Permission (§2.2)	Analogy to minor treatment is not belittlement, but acknowledgment of subjectivity differences
Capability Diversity (§2.4)	AI's single-dimensional "technical capability" is high, but cannot replace responsibility, ethical judgment, and other diverse dimensions
Technocrat Critique	Preventing "AI is more reliable" from becoming an efficiencyist excuse to deprive humans of permissions
Discourse Transparency (§2.7)	"AI neutral and efficient" technical discourse cannot substitute for the political question of "who bears responsibility"

VIII. Open Questions

Although this document gives a negative conclusion, the following questions still need continuous examination:

If AI capabilities do far exceed human capabilities, will insisting on human supervision lead to preventable disasters? (For example, AI can predict earthquakes but human supervisors refuse to execute)
- This is real tension. But Stairway Universalism's response is: It would rather bear efficiency losses than delegate permissions before establishing responsibility mechanisms. Efficiency is not the highest value; symmetry between responsibility and permission is.
How to define "AI responsibility capability"? Is it necessary to establish criteria for "AI responsibility capability"?
- This document believes that current and foreseeable AI do not possess responsibility capability. If changes occur in the future, this is not a technical determination problem, but a political problem that needs to be reviewed through global democratic procedures.
How to set the "understanding threshold" for AI-assisted decision-making?
- What levels of AI suggestions do human supervisors need to understand? Do all high-risk decisions need to be fully explainable? How to balance the cost of explainability and safety?

Core Conclusion: Stairway Universalism's position on AI subjectivity is not human-centered arrogance, but faithful derivation of its core principles. Permission is social trust; trust requires responsibility mortgage; and responsibility requires subjectivity. AI can simulate all expressions of subjectivity, but it has no life, no situation, no cost, and no irrevocable consequences. Therefore, AI cannot become an independent permission subject, and cannot independently bear the responsibility for high-permission operations. For the foreseeable future, it can only be a tool—an extremely powerful tool, but still a tool.