The conscience of the pipeline — the layer that cannot yet be replaced, because trust is not a function of accuracy. It is a function of consequences survived.
Embodied AI is the fifth layer — and the one that makes the other four trustworthy. It audits the entire workflow for bias, equity gaps, PRISMA compliance, and methodological integrity. In the current moment, this layer is occupied by a human being: a health economist, a systematic review methodologist, a senior WHO officer who reads the final product and signs her name to it.
When the WHO India director asked for this work to be delivered, she wasn't asking for a faster pipeline. She was asking for the pipeline to produce something she could trust enough to stake her professional judgment on. That is embodied AI: the final arbiter before the output leaves the room.
No AI system in 2025 can tell you whether a GRADE summary accurately represents the uncertainty a seasoned health economist would feel reading the underlying papers. No model has faced the consequences of a bad recommendation reaching a health minister. The WHO director has. That accumulated consequence — that lived feedback — is what currently occupies this layer. It is irreplaceable not because the tasks are difficult, but because trust has not yet been transferred.
A policy brief on PM-JAY's impact on catastrophic health expenditure reaches Phase V. The generative synthesis correctly summarises the extracted literature. The human auditor notices that every included paper uses administrative enrolment data — not survey-based CHE measurement — and that this distinction is nowhere flagged. She adds a limitations paragraph. The brief is now honest in a way no automated layer could have made it.
The original table note read: Human in loop for now — but AGI coming soon. This deserves to be taken seriously, and then held at arm's length.
AGI — in the sense that matters here — is not the point at which a model passes a benchmark. It is the point at which we trust the output enough to act on it without checking. That trust is not a property of the model. It is a property of the model's track record, accrued through trial, error, consequence, and correction — the same process by which a junior researcher becomes someone whose judgment is trusted unsupervised.
The Scalar phase notation L(θᵗ⁺¹) describes a loss function that updates model parameters. For a human auditor, this is biography. Every miscalculation that reached a policy table, every recommendation that turned out wrong, every brief that needed retraction — this is what the superscript t+1 represents. Updated beliefs, earned through consequence. No model trained on text has this. The WHO director does.
This is why the infrastructure framing matters. The first four layers are infrastructure for accumulating the feedback that eventually produces trustworthy judgment. They make the pipeline fast enough to run repeatedly, at enough scale, across enough contexts, that the feedback loops begin to close. That accumulation — not any single benchmark — is the path to trust.
This evidence synthesis work is delivered pro bono to WHO India. That choice is itself an act of embodied AI — a human in the loop who decided that the infrastructure for better health policy decisions was worth building before it was paid for. The WHO director's request was not bureaucratic; it was a recognition that the pipeline needed a conscience layer, and that she needed to trust it.
The five layers of this taxonomy describe a system. But systems without a human willing to audit them, correct them, and stake something on the outcome are not systems — they are processes without accountability. Embodied AI is the accountability layer. It exists because consequences exist.
World AI provides the environment. Perception AI provides the senses. Agentic AI provides the hands. Generative AI provides the voice. Embodied AI provides the judgment — the accumulated history of being wrong and surviving it. The stack is not complete without all five. And the fifth is, for now, a person.