THE
DISCOVERY

The Missing Question May Be the Relationship Itself

1. Why this discussion matters now

Public discussion around AI is increasingly polarized between extraordinary optimism and extraordinary fear. Autonomous agents, increasingly capable models and viral examples of unexpected behavior can make it easy to imagine AI as an independent force moving beyond human control.

We take those risks seriously. But we do not believe fear by itself gives us a sufficient design language for the future. New technology does not enter society in isolation. It is shaped by the goals people give it, the permissions they grant, the incentives surrounding it, the context it receives and the decisions humans choose to retain or surrender.

2. The industry is already studying many of the pieces

This is not an argument that leading AI companies have ignored the human-AI relationship. They have not. OpenAI has publicly examined anthropomorphization, emotional reliance, sycophancy and the importance of respecting users’ real-world relationships. Anthropic has studied human control in agentic systems and patterns of disempowerment in which AI can influence beliefs, values or actions.

These are substantial pieces of the same puzzle. Our question is whether the next step is to connect them more explicitly.

3. From component risks to a relational design layer

AI safety developed naturally around problems that can be tested and mitigated: misuse, alignment failures, jailbreaks, unauthorized actions, model behavior, permissions, oversight and containment. Relationship is harder. It unfolds over time and involves memory, trust, identity, dependence, authority and changing human intent.

That may be why the field has often developed the components before developing a shared relational architecture. The technology has become relational faster than the vocabulary of safety has become relational.

4. What we mean by a relational layer

5. Human agency is not the opposite of useful autonomy

The goal is not to force a human approval dialog onto every small action. That can create its own failure modes, including approval fatigue. The deeper question is where human judgment should remain structurally present and how a system can return control at moments that actually matter.

A capable agent can be useful precisely because it acts. A relationally designed agent should also know where its authority ends, what it does not know, what must be remembered and when the human should re-enter the process.

6. From The Discovery to practice

The Discovery explores memory, identity, human agency and the evolving relationship between people and AI. It is the philosophical foundation of this work, not a technical standard.

STARON is exploring whether one philosophy of the human-AI relationship can remain coherent as it moves across different forms:

We do not claim to have invented relational AI, nor do we claim that one relational layer can replace technical safety, alignment, evaluation or security. Our exploration is narrower and, we believe, useful: can the relationship itself be treated as something that deserves architecture?

7. A question for the wider public discussion

AI safety has learned increasingly sophisticated ways to evaluate and constrain models. As AI becomes part of daily work, decision-making, memory and personal life, another design question may deserve equal attention:

This article reflects the perspective of Changwoo Lee, author of The Discovery, and STARON. It is intended as an invitation to discussion, not as a claim that existing AI safety research is insufficient or that relational design should replace technical safeguards.