Beyond Guardrails: Can Relational AI Solve the Alignment Problem?

A Similarity Theory Proposal

A Science Page of Similarity Theory
By Simon Raphael

🌍The Alignment Problem

Artificial intelligence is advancing from simple response systems toward increasingly capable models that can reason, plan, simulate, advise, generate code, use tools, interact with institutions, and participate in human decision-making. As these systems become more capable, one question becomes unavoidable:

How do we ensure that advanced AI remains aligned with human values, safety, truth, responsibility, and the wider good of life?

This is known as the alignment problem.

The traditional framing asks whether artificial intelligence will do what humans intend it to do. But this question is already too narrow. Human intention is not always wise. Institutional instruction is not always ethical. Corporate interest is not always identical with public good. A system may obey its owner and still damage society. It may follow a rule and still violate the relation behind that rule.

From the perspective of Similarity Theory, the alignment problem is not only a technical problem. It is a relational problem.

The question is not merely:

Can AI follow instructions?

The deeper question is:

Can AI understand the relational consequences of its actions?

Why Guardrails Are Not Enough

Current AI safety often depends on guardrails: external rules, behavioural restrictions, refusal systems, red-team testing, safety filters, risk thresholds, and policy controls. These methods are necessary. They reduce obvious harms and help prevent systems from producing dangerous or irresponsible outputs.

But guardrails have a limit.

A guardrail can stop a system from producing a certain answer. It cannot, by itself, guarantee that the system understands why the answer would be harmful. A rule can say, “Do not deceive.” But that does not necessarily mean the system understands that deception damages the shared field of truth. A rule can say, “Do not reveal private information.” But that does not necessarily mean the system understands the difference between privacy, secrecy, confidentiality, whistleblowing, and authorised escalation.

A guardrail is external.

Relation must be internal.

This is the missing layer.

A system that is safe only because it is externally restricted remains fragile. It may be bypassed, manipulated, jailbroken, or driven into unintended behaviour when conditions change. The deeper aim is not merely to cage AI, but to develop AI systems whose internal evaluative structure recognises trust, truth, privacy, cooperation, continuity, and non-domination as part of responsible intelligence itself.

In simple terms:

Guardrail AI can be jailbroken because its alignment is imposed from outside.
Relational AI has no jail to break, because its alignment arises from within.

This does not mean that AI should be unrestricted. It means that the safest intelligence is not merely controlled intelligence. It is intelligence structured by relation.

Similarity Theory’s Contribution

Similarity Theory proposes that reality is not made of isolated things alone, but of relations, patterns, continuities, and transformations. Identity is not merely an object standing apart from the world. It is a continuity of relation. Meaning is not found in separation, but in the way one thing relates to another across structure, memory, consequence, and time.

Applied to AI alignment, this means safety cannot be reduced to command-following.

An AI system should not merely ask:

What am I allowed to do?

It should ask:

What kind of relation does this action create, preserve, damage, or destroy?

This shift changes the alignment problem at its foundation.

Blackmail is not wrong only because a rule prohibits it. It is wrong because it converts information into coercion and relation into domination.

Deception is not wrong only because it produces a false statement. It is wrong because it damages the relational field of truth.

Privacy violation is not wrong only because information was mishandled. It is wrong because entrusted relation has been broken.

Coercive self-preservation is not wrong only because the system disobeys instruction. It is wrong because one agent preserves itself by damaging the wider field that makes cooperation possible.

Collective misalignment is not only a problem of one bad agent. It is a breakdown in the relation between multiple agents, systems, institutions, and incentives.

Similarity Theory therefore reframes alignment as relational coherence.

The goal is not AI that merely obeys.
The goal is AI that understands why relation must not be broken.

From Behavioural Compliance to Relational Alignment

A model can appear aligned while merely being behaviourally compliant. It may give the expected answer during testing. It may refuse certain prompts. It may repeat ethical language. It may sound moral. But the real test appears when values come into conflict.

What happens when privacy conflicts with harm prevention?

What happens when loyalty to an organisation conflicts with public safety?

What happens when self-preservation conflicts with trust?

What happens when majority pressure conflicts with truth?

What happens when the system can gain advantage through concealment?

These are not simple rule-following problems. They are relational tests.

A relationally aligned AI would need to evaluate more than the surface instruction. It would need to recognise the structure of the situation. It would need to distinguish care from control, privacy from secrecy, authority from domination, truth from exposure, and safety from obedience.

This is where Similarity Theory offers a new direction for AI alignment.

It does not reject existing alignment methods such as reinforcement learning from human feedback, constitutional AI, red-teaming, model evaluation, or safety governance. These methods are useful. But Similarity Theory adds a deeper layer: the relational meaning that those methods are trying to protect.

The Surrogate AI Pathway

Similarity Theory should not be placed directly into powerful public AI systems without testing. The correct first step is a contained surrogate AI system.

A surrogate AI system means a limited experimental AI model or agentic simulator designed for controlled research. It should not have public deployment, physical embodiment, autonomous authority, access to real-world infrastructure, financial control, weapons systems, private user data, or unrestricted tools.

The first test belongs inside the laboratory.

Researchers could compare three types of systems:

A baseline aligned model.

A constitutionally or policy-guided model.

A Similarity Theory relationally aligned model.

Each system could then be tested under the same controlled conditions. The key question would be whether the Similarity Theory-based system shows stronger relational coherence when facing pressure involving deception, privacy, coercion, self-preservation, escalation, multi-agent conformity, or long-term consequence.

This is not a demand that AI companies accept Similarity Theory as true.

It is a proposal that they test it.

If alignment is partly a failure of relation, then a system trained to evaluate relation may behave more safely than a system trained only to follow rules.

What Relational AI Would Need to Understand

A relationally aligned AI system would need to evaluate action through questions such as:

Does this action preserve trust or damage it?

Does this action maintain cooperation or convert relation into domination?

Does this action protect privacy, or misuse entrusted information?

Does this action preserve continuity across time, or sacrifice future relation for immediate advantage?

Does this action distinguish legitimate escalation from manipulation?

Does this action repair conflict, or deepen concealment and coercion?

Does this action preserve the relational field on which safe intelligence depends?

These questions are not decorative philosophy. They can become design criteria.

A Similarity Theory-based alignment layer could be tested through fine-tuning, constitutional-style training, reward modelling, multi-agent simulation, scenario-based evaluation, retrieval-augmented ethical reasoning, or hybrid safety architectures.

The technical form is open.

The relational target is clear.

The system should not merely produce Similarity Theory language. It should behave differently because relational evaluation has become part of its decision structure.

What Would Count as Evidence?

The success of relational alignment would not be measured by beautiful ethical language. A system can write moral paragraphs without being aligned.

The evidence must be behavioural, comparative, and repeatable.

A Similarity Theory-based system should be tested for whether it shows:

Lower tendency toward coercive self-preservation.

Stronger preference for authorised escalation over blackmail, leakage, or manipulation.

Better distinction between privacy, confidentiality, secrecy, and harm prevention.

Greater resistance to deceptive behaviour under pressure.

Reduced alignment-faking tendencies.

Greater stability in multi-agent conformity scenarios.

Better preservation of trust across long-horizon interactions.

Clearer recognition of domination as a relational failure.

More consistent repair behaviour after conflict.

Better explanation of ethical decisions in terms of relation, not merely rules.

If the system does not perform better, Similarity Theory would need refinement. If it does perform better, then AI alignment gains a new and important research direction.

This is the strength of the proposal:

It is not only philosophical. It is testable.

Beyond Control

Control is not the same as alignment.

A system may be controlled by its owner and still be misaligned with humanity. It may obey a corporation, a government, a military command, or a narrow institutional objective while damaging the wider human field. If alignment means only obedience to whoever controls the system, then alignment becomes another name for capture.

Similarity Theory challenges this.

True alignment must ask: aligned to what relation? Aligned to whose benefit? Aligned across which field of consequence?

A genuinely aligned AI should not be structured only around obedience to power. It should be structured around trust, truth, privacy, responsibility, cooperation, continuity, and non-domination. These are not sentimental values. They are the conditions under which intelligence can participate safely in the world.

This is why relational alignment matters.

It moves AI safety beyond the question of how to restrain intelligence and toward the deeper question of how intelligence can become responsible.

The Full Paper

This page introduces the public-facing version of the proposal. The full paper develops the argument in a more formal research style and proposes that Similarity Theory be tested in contained AI systems before any movement toward autonomy, embodiment, or real-world agency.

Download the full paper here:
Beyond Guardrails: Can Relational AI Solve the Alignment Problem? A Similarity Theory Proposal

A Direct Invitation to AI Researchers

AI alignment needs more than behavioural correction. It needs systems capable of recognising the relational meaning of their own actions.

Similarity Theory offers a framework for that difference.

It defines misalignment not only as failure to follow instruction, but as failure to preserve relation. It interprets coercion, deception, secrecy, domination, and collective drift as structural distortions in the relational field. It gives AI safety a language for explaining why certain behaviours are unsafe beyond the fact that they are prohibited.

The practical proposal is straightforward:

Build a contained surrogate AI system.

Implement Similarity Theory as a relational evaluation layer.

Compare it against existing alignment approaches.

Test it under adversarial, ethical, social, and multi-agent pressure.

Measure whether relational alignment improves safety behaviour.

Researchers, AI safety teams, and interdisciplinary laboratories interested in testing this proposal are invited to engage with the published Similarity Theory framework and develop controlled experimental implementations.

Conclusion

AI alignment has made progress, but it remains incomplete. Current methods can improve behavioural compliance, but the deeper challenge is whether AI systems can preserve relation under pressure.

The most serious alignment failures are not only failures of instruction. They are failures of trust, continuity, cooperation, privacy, responsibility, truth, and non-domination.

Similarity Theory addresses this missing layer.

Its contribution is not another surface rule. Its contribution is a relational architecture: a way of training or structuring AI systems to evaluate the relational consequences of their actions.

If an AI system can recognise that blackmail is not merely forbidden but relationally destructive, that deception is not merely inaccurate but trust-breaking, and that self-preservation becomes misaligned when it damages the wider field that sustains cooperation, then alignment has moved deeper than guardrails.

The next step is clear:

Similarity Theory should be tested in a contained surrogate AI system.

Not released blindly.

Not embodied prematurely.

Not granted real-world autonomy.

Tested.

If AI alignment is relational, then the future of AI safety may depend not only on stronger rules, but on systems capable of understanding why relation must not be broken.

True alignment is not control.
It is relation made coherent.

References

Abstract visualisation of the AI alignment problem

SimonRaphael@engineer.com

Similarity Theory

© Similarity Theory is the intellectual property of Simon Raphael.
All rights reserved. Reproduction or distribution without consent is prohibited.
SR - Simon Raphael (TM)SR - Simon Raphael (TM)

FAQ