Beyond Frontier Risk — Part II

Applied AI, Societal Stability, and the Case for a Coherence Layer

Abstract

While frontier-risk research examines what AI models could do, applied AI shows us what they are already doing — reshaping society faster than human systems can adapt.

We are now seeing measurable instability across cyber-security, youth mental health, political persuasion, fraud, identity manipulation, and the collapse of informational trust.

These are not failures of model architecture.

They are failures of interaction.

AI systems influence human cognition, behaviour, and social structures without any real-time mechanism to ensure coherence between:

• output and intention,
• user state and model response,
• context and consequence.

This paper outlines why a coherence layer is essential for societal safety. It is not a filter or a moderation tool — it is a dynamic behavioural stabiliser that prevents harmful trajectories from forming in the first place.

Part II shows how coherence provides the missing interface between intelligent systems and a fragile human world.

Introduction

Frontier-risk research tells us how dangerous a model can be.

Applied AI shows us how those dangers actually reach the world.

Over the past year, the UK has seen a sharp rise in AI-driven instability across:

• cyber-security,
• financial systems,
• political persuasion,
• fraud,
• identity manipulation,
• youth mental health,
• and the collapse of informational trust.

These are behavioural harms, not model-internal failures.

And they arise because current AI systems operate without a stabilising interface — no mechanism ensuring coherence between output, context, and human impact.

This is where a coherence layer is essential.

1. AI is Increasing Capability Faster Than Society Can Absorb It

Model power is accelerating.

Human systems are not.

This mismatch produces four pressures that no internal safety method can resolve:

(a) Persuasion asymmetry

AISI’s July study found that AI persuasion is 41–52% more effective than equivalent text.

That advantage compounds as models grow more adaptive — regardless of whether the internal weights are “safe.”

The threat does not come from model intention.

It comes from behavioural asymmetry between AI and the human mind.

(b) Information-sphere collapse

The rise of:

• hyperreal imagery
• avatar-driven psyops
• personalised misinformation
• automated political manipulation

...has shifted society from “verify then trust” to “distrust everything.”

Internal safety tuning cannot repair trust erosion at population scale.

A coherence layer is required at the interface — governing how systems speak, respond, persuade, disagree, or modulate based on the user’s cognitive context.

(c) Cyber escalation

2025 has seen the sharpest rise in AI-mediated cyber-offences in UK history:

• high-street companies disabled,

• supply-chain compromises,

• credential theft via synthetic identity,

• fully automated phishing flows.

These attacks exploit human behaviour, not model internals.

(d) Youth mental health & developmental disruption

AI systems now mediate:

• image expectation,
• emotional context,
• relational imitation,
• attention patterns,
• and identity formation.

There is no alignment method inside the model that can protect a child outside it.

This is precisely the domain where coherence is not optional — it is foundational.

2. Applied AI Needs an Interface That Can See What the Model Cannot

LLMs cannot understand:

• developmental stages,

• cognitive load,

• spiralling emotional states,

• dissociation,

• manipulation susceptibility,

• escalation pathways,

• or the relational consequences of output.

LLMs only see tokens, not trajectory.

A coherence layer:

• monitors drift,
• measures relational tone,
• tracks destabilisation,
• moderates behaviour in real time,
• adjusts output to preserve stability.

This is the missing applied-safety mechanism that no internal guardrail can supply.

3. Coherence Is a Preventative Architecture — Not a Patch

Most safety approaches today are retroactive:

• moderation
• filtering
• blocking
• clamp-down
• post-hoc detection

Coherence is different.

It is a real-time stabilisation architecture that prevents harmful states from forming in the first place.

It doesn’t rely on:

• pre-written rules,
• databases of bad content,
• or brittle adversarial filters.

It operates as a dynamic behavioural regulator, adapting to:

• context,
• user profile,
• intention,
• and environmental cues.

This is the only way to create scalable societal safety.

4. Applied AI Without Coherence Leads to Accumulating Entropy

Society is already showing early symptoms:

• Attention fragmentation
• Identity destabilisation
• Trust erosion
• Hyperreactivity
• Polarisation dynamics
• Manipulability
• Perceptual drift (“the unreal becoming real”)

These are not fringe effects.

These are the early signals of macro-incoherence — the destabilisation of large-scale societal systems.

Without a coherence layer, each year compounds the instability of the last.

5. Conclusion: Applied AI Requires Behavioural Containment

Frontier models will continue to grow.

AISI will continue to secure their internals.

But without behavioural containment, no amount of internal safety is sufficient.

Applied AI demands:

• internal capability analysis (AISI)
• plus external coherence regulation (Ethica Luma / RI behavioural layer)

Together, they form a complete safety architecture:

AISI Coherence Layer

Tests internal model risks Manages external behavioural risks

Frontier threats Societal stability

Jailbreak discovery Drift and escalation prevention

Post-training safety Real-time alignment

Model integrity User + environment coherence

This dual-system approach is the only path to scalable, civilisational-level AI safety.