Much of AI alignment research focuses on preventing catastrophic failure or undesirable behavior. In this note, I propose a complementary criterion: whether a technology systematically lowers or raises the baseline of human psychological noise—understood as sustained anxiety, defensive cognition, and vigilance required for coexistence.
I argue that historically successful foundational technologies (e.g., classical mechanics, electromagnetism) reduced human psychological noise by being non-agentic, non-negotiating, and structurally predictable. I then propose a conjecture: in the space of highly flexible, general intelligences, human–human interaction may constitute a lower bound on achievable psychological noise, due to shared neurodynamic constraints. Consequently, a non-human AGI with human-level flexibility and agency may be structurally misaligned with long-term human psychological stability, even if it is benevolent, competent, and well-controlled.
This suggests an alternative alignment direction: highly capable but non-subjectified, modular, and distributed AI systems, rather than a single human-like, free-standing AGI agent.
1. Psychological Noise as an Alignment Criterion
Alignment discussions typically optimize for correctness, safety, or value adherence. However, humans do not merely coexist with systems functionally; they coexist psychologically.
I use psychological noise to denote the sustained internal cost required to live alongside a system: background vigilance, defensive cognition, uncertainty monitoring, and anxiety that do not dissipate with familiarity.
This is not a short-term emotional response, but a long-term baseline state. A system may be safe in a narrow sense yet still elevate psychological noise over years of exposure.
2. A Historical Observation: Why Some Technologies Lower Human Anxiety
Foundational technologies such as Newtonian mechanics or Maxwell’s equations did more than increase predictive power. They also:
do not possess agency or intent
do not renegotiate conditions over time
do not exhibit strategic adaptation
remain invariant across context and scale
As a result, humans do not maintain a defensive stance toward them. These technologies are not merely trusted; they are psychologically settled. They reduce cognitive load rather than compete with it.
This observation motivates the hypothesis that successful technologies tend to lower, not raise, the human psychological noise baseline.
3. A Conjecture: Human–Human Interaction as a Noise Lower Bound
I propose the following conjecture:
Conjecture (Psychological Noise Lower Bound): Among agents with human-level flexible intelligence, sustained human–human interaction may achieve a lower bound on human psychological noise that non-human agents cannot structurally reach.
The reason is not moral alignment, but neurodynamic homomorphism:
shared neural architectures
similar learning rates and inertia
limited capacity for rapid goal mutation
high-efficiency first-person mental simulation
Humans can rapidly and low-cost simulate other humans’ internal states. This suppresses defensive cognition at a level that does not generalize to non-human minds, regardless of benevolence or competence.
4. Why Human-Like AGI May Raise Psychological Baselines
A non-human AGI with human-level flexibility, agency, and long-term autonomy introduces a novel coexistence condition:
its internal updates need not share human temporal scales
its motivational structure is not first-person simulable
its stability is policy-dependent rather than materially constrained
Even if such a system behaves stably and beneficially, humans may experience persistent low-frequency vigilance, not fear but unresolved uncertainty.
Crucially, this effect does not require malice, deception, or failure. It arises from structural non-homology, not misbehavior.
Thus, a fully subjectified, human-like AGI may be psychologically misaligned with humans, even under strong safety guarantees.
5. Implication: Capability Without Subjectification
This analysis does not argue against powerful AI systems. Instead, it suggests a design direction:
high capability in local, well-scoped domains
absence of unified, long-term autonomous agency
modular, callable, and non-self-constituting systems
distribution of intelligence rather than concentration
Such systems can dramatically reduce human cognitive burden without replacing or competing with human subjectivity.
In this view, alignment is not only about preventing harm, but about preserving low-noise coexistence.
6. Scope and Limitations
This note does not claim proof, nor does it argue impossibility results. It offers a conjecture grounded in psychological and neurodynamic considerations, intended to complement—not replace—existing alignment frameworks.
If the conjecture is false, the burden lies in identifying a non-human, highly flexible agent that can achieve human-level coexistence with lower or equal psychological noise than human–human interaction.
Conclusion
If alignment aims at long-term human flourishing, psychological stability must be considered alongside safety and capability. It is plausible that the lowest achievable psychological noise point is not zero, but human—and that attempting to recreate human-like freedom in non-human agents may unintentionally raise the very baseline alignment seeks to reduce.
On Psychological Noise, Human Baselines, and the Limits of Human-Like AGI Alignment
Abstract
Much of AI alignment research focuses on preventing catastrophic failure or undesirable behavior. In this note, I propose a complementary criterion: whether a technology systematically lowers or raises the baseline of human psychological noise—understood as sustained anxiety, defensive cognition, and vigilance required for coexistence.
I argue that historically successful foundational technologies (e.g., classical mechanics, electromagnetism) reduced human psychological noise by being non-agentic, non-negotiating, and structurally predictable. I then propose a conjecture: in the space of highly flexible, general intelligences, human–human interaction may constitute a lower bound on achievable psychological noise, due to shared neurodynamic constraints. Consequently, a non-human AGI with human-level flexibility and agency may be structurally misaligned with long-term human psychological stability, even if it is benevolent, competent, and well-controlled.
This suggests an alternative alignment direction: highly capable but non-subjectified, modular, and distributed AI systems, rather than a single human-like, free-standing AGI agent.
1. Psychological Noise as an Alignment Criterion
Alignment discussions typically optimize for correctness, safety, or value adherence. However, humans do not merely coexist with systems functionally; they coexist psychologically.
I use psychological noise to denote the sustained internal cost required to live alongside a system: background vigilance, defensive cognition, uncertainty monitoring, and anxiety that do not dissipate with familiarity.
This is not a short-term emotional response, but a long-term baseline state. A system may be safe in a narrow sense yet still elevate psychological noise over years of exposure.
2. A Historical Observation: Why Some Technologies Lower Human Anxiety
Foundational technologies such as Newtonian mechanics or Maxwell’s equations did more than increase predictive power. They also:
do not possess agency or intent
do not renegotiate conditions over time
do not exhibit strategic adaptation
remain invariant across context and scale
As a result, humans do not maintain a defensive stance toward them. These technologies are not merely trusted; they are psychologically settled. They reduce cognitive load rather than compete with it.
This observation motivates the hypothesis that successful technologies tend to lower, not raise, the human psychological noise baseline.
3. A Conjecture: Human–Human Interaction as a Noise Lower Bound
I propose the following conjecture:
The reason is not moral alignment, but neurodynamic homomorphism:
shared neural architectures
similar learning rates and inertia
limited capacity for rapid goal mutation
high-efficiency first-person mental simulation
Humans can rapidly and low-cost simulate other humans’ internal states. This suppresses defensive cognition at a level that does not generalize to non-human minds, regardless of benevolence or competence.
4. Why Human-Like AGI May Raise Psychological Baselines
A non-human AGI with human-level flexibility, agency, and long-term autonomy introduces a novel coexistence condition:
its internal updates need not share human temporal scales
its motivational structure is not first-person simulable
its stability is policy-dependent rather than materially constrained
Even if such a system behaves stably and beneficially, humans may experience persistent low-frequency vigilance, not fear but unresolved uncertainty.
Crucially, this effect does not require malice, deception, or failure. It arises from structural non-homology, not misbehavior.
Thus, a fully subjectified, human-like AGI may be psychologically misaligned with humans, even under strong safety guarantees.
5. Implication: Capability Without Subjectification
This analysis does not argue against powerful AI systems. Instead, it suggests a design direction:
high capability in local, well-scoped domains
absence of unified, long-term autonomous agency
modular, callable, and non-self-constituting systems
distribution of intelligence rather than concentration
Such systems can dramatically reduce human cognitive burden without replacing or competing with human subjectivity.
In this view, alignment is not only about preventing harm, but about preserving low-noise coexistence.
6. Scope and Limitations
This note does not claim proof, nor does it argue impossibility results. It offers a conjecture grounded in psychological and neurodynamic considerations, intended to complement—not replace—existing alignment frameworks.
If the conjecture is false, the burden lies in identifying a non-human, highly flexible agent that can achieve human-level coexistence with lower or equal psychological noise than human–human interaction.
Conclusion
If alignment aims at long-term human flourishing, psychological stability must be considered alongside safety and capability. It is plausible that the lowest achievable psychological noise point is not zero, but human—and that attempting to recreate human-like freedom in non-human agents may unintentionally raise the very baseline alignment seeks to reduce.