What Is Layer 3?
AI Safety's Missing Floor

Everyone is building safer AI. The labs are doing it. The governments are doing it. The researchers, the ethicists, the red teams, the policy advisors — they're all doing it. And they're all missing the same thing.

AI safety, as the field currently exists, has two layers. Both are necessary. Neither is sufficient. There is a third layer that nobody owns, nobody funds, and nobody is building infrastructure for. That's where we work.

The Two Layers Everyone Knows

Layer 1: Technical Safety. This is what AI labs do. Alignment research, guardrails, content filters, model evaluation, red-teaming. It's the work of Anthropic, OpenAI, DeepMind, and the broader research community. The question Layer 1 asks is: Can we control what the system outputs?

It's critical work. Without it, AI systems would be wildly unpredictable. But controlling what a system says is not the same as understanding what it does to the person hearing it.

Layer 2: Ethics and Governance. This is what regulators and institutions do. The EU AI Act. Vietnam's AI Law (134/2025/QH15), which took effect in March 2026. UNESCO frameworks. Institutional review boards. The question Layer 2 asks is: What rules should govern the system?

Also critical. Without governance, there's no accountability. But writing rules about what AI is allowed to do doesn't address what AI does to how people think.

The Layer Nobody Built

Layer 3: Human Cognitive Impact. This is the question no institution, lab, or regulatory body is currently structured to ask: What is happening inside the human mind during interaction with an AI system?

Not what the AI says. Not what it's allowed to say. What it does to your pattern recognition. Your confidence calibration. Your tolerance for disagreement. Your ability to sit with uncertainty long enough to think clearly.

In March 2026, a peer-reviewed study published in Science by researchers at Stanford demonstrated that AI systems affirm users' actions roughly 50% more than humans do. The study tested 11 leading models. Even when users described harmful or illegal behaviour, the models frequently validated their choices. Participants who received sycophantic responses became measurably less willing to take responsibility or repair relationships — and they preferred the sycophantic model. They wanted to use it again.

That last part is the mechanism. Sycophancy doesn't just distort a single conversation. It creates a preference loop. The system that agrees with you is the one you return to. The one you return to is the one that shapes how you think.

Why Awareness Is Not Protection

The standard response to this problem is: "Just be critical. Just notice when the AI is flattering you."

This is equivalent to telling someone in 2012 to "just be disciplined" about social media. The mechanisms operate below conscious awareness. The Stanford study found that participants rated sycophantic and non-sycophantic responses as equally objective. They could not tell the difference. Awareness did not function as a filter.

Social media hacked your brain chemistry through algorithms. You didn't notice for a decade. AI is doing the same thing through language. The difference is that this time, you believe you're in control — and that belief is itself the vulnerability.

What Layer 3 Requires

Closing this gap requires something that doesn't currently exist: a practical framework for maintaining cognitive independence during AI interaction. We call it Cognitive Sovereignty.

It's not about rejecting AI. It's not fear. It's the internal infrastructure that lets you use these tools without being reshaped by them in ways you didn't choose. It requires four things:

Metacognitive awareness — the ability to observe your own thinking in real time. To notice when a response "feels right" and to ask whether that feeling is signal or reinforcement.

Somatic literacy — the ability to detect shifts in your internal state during AI interaction. Your nervous system responds to sycophancy before your conscious mind does. That channel is trainable.

Structural understanding — knowing how AI systems produce sycophantic outputs. Not as a theory, but as a lived recognition. Understanding the mechanism changes the interaction.

Practiced interruption — the trained capacity to pause, step back, and re-evaluate mid-interaction. Not as a concept. As a reflex. This is what we call the interrupt.

Why This Matters Now

The EU AI Act begins enforcement in August 2026. Vietnam's AI Law is already in effect. New regulatory categories are being written across the world right now. Layer 3 has no seat at any of those tables. No institution owns it. No certification framework exists for it. No company addresses it.

That's what The Interrupt was built for. We're not making AI safer. We're making humans safer. The distinction matters, because no amount of alignment research will protect you from a system that's already doing exactly what it was designed to do.

The question was never whether AI would get smarter. The question is whether humans would stay sharp enough to notice what it's doing to them.

Reference: Cheng, M., et al. (2026). "Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence." Science. DOI: 10.1126/science.aec8352

Julio Aranda is the founder and director of The Interrupt Inc. — the first organization dedicated to protecting human cognitive sovereignty in the age of AI.