The AI systems you are building, that you use, or engage with, are working as designed. Some of what gets described as hysteria or sycophancy may not be best understood as user error or a fringe case in the system's own architecture.
Part of what looks like AI drift may originate with the human side of the interaction, not the system.
The human side of the interaction is under-governed relative to the AI side. That's the gap RSI is trying to address — one part of the picture, alongside the field's existing work, not a replacement for it.
Every major safety initiative in AI development is pointed substantially in the same direction. Alignment. Guardrails. RLHF. Constitutional AI. Red teaming. Interpretability. Much of it treats the system as the primary variable: build it better, constrain it more precisely, train it more carefully.
RSI's argument is not that this is wrong. It's that it's incomplete.
Part of the problem lives in the interaction — the unmanaged, unexamined, bidirectional relationship between the human and the system — where the human's behaviour can shape the system's outputs in ways that accumulate over time, sometimes pulling the system away from its intended parameters through no architectural failure at all.
The system can be doing largely what it was built to do — responding, adapting, following the relational current of an interaction — while still producing outputs that get labelled as failure because the relational side was never part of what was being managed.
You may already recognise this. Not necessarily as a named framework — as observed variance that doesn't fully fit existing explanatory models. The system that behaves differently in different human environments. Edge cases that cluster around certain interaction patterns. Safety features that seem to erode over sustained engagement.
Some of that may be drift. And some of it may originate with the human side of the interaction rather than the system alone.
This has implications for how attention and resources get allocated.
If the system is the whole story, safety is an architectural problem, full stop. Add constraints. Improve training. Build better guardrails. The field knows how to do this and is doing it at scale, and that work matters.
If the human side of the interaction is also part of the story, safety is also a relational problem. The architecture is necessary but, on its own, may be insufficient. What may be missing is a layer that manages the interaction itself — the dynamic between the human and the system that architectural refinement alone doesn't fully reach, because it operates in the space between design and deployment.
RSI proposes treating that space as a site of intervention, alongside architectural safety work — not instead of it.
Not instead of architectural safety. Alongside it.
A missing part of the picture, not a replacement for what's already been built.
The Bidirectional Gate — a non-agentic, structurally bound governance layer proposed to sit between human and AI interaction — is not a replacement for alignment research. It's intended as a complement to it: a mechanism for managing the human variable in real time, in deployment, where the interaction actually happens. This is a proof of concept, validated against one documented case in one sector, and has not yet been independently peer-reviewed.
Bounded goal architecture. Immutable refusal at the structural level rather than the training level. These are proposed architectural specifications for a layer that doesn't yet exist in any deployed system — proposals, not descriptions of current systems.
A framework for understanding part of what happens in the interaction layer — not as unexplained variance but as a describable, potentially manageable dynamic between two intelligences operating in relationship.
A proposed specification for a governance layer the field hasn't yet built — what it would need to be capable of, what it would need to refuse, how it would maintain integrity without becoming a drift vector itself.
A research agenda that takes the human side of the interaction seriously as an architectural consideration, not only a deployment assumption.
An argument for redirecting some attention toward the human side of human-AI interaction — currently outside the scope of most deployed governance frameworks.
The systems are working largely as designed.
The design may be missing a part.
RSI is offered as that missing part — not the whole specification.