Glossary · Verifiable AI

Human-in-the-Loop — HITL

ヒューマン・イン・ザ・ループ

An operating model where a human reviews and approves an AI system's decision before it takes effect. A human can catch errors and deviations before they execute — but inserting a person into every decision sets the throughput ceiling.

Definition

Human-in-the-loop (HITL) describes designs that never execute AI output directly: a human review-and-approval step gates every decision before it is committed. It has long been the standard control in domains where the cost of error is high — lending, medical diagnosis, credit decisions.

EU AI Act Article 14 mandates "human oversight" for high-risk AI; HITL is its most conservative implementation. A human in each decision blocks automation bias and malfunction before execution.

But HITL does not scale by nature. Once decision volume exceeds human review capacity, the approval queue becomes the rate limit. For use cases where an agent decides continuously, gating every decision on a human stops being realistic.

Lemma implementation

Lemma records the fact that a human approved as an element of the audit trail. The approver, the approval timestamp, and the target docHash are committed, so "this decision passed human approval" becomes provable tamper-evidently after the fact.

This moves HITL from "a claim in an operational log" to "a cryptographically verifiable fact." An auditor can confirm approval via zero-knowledge proof without seeing the underlying data.

Rather than removing HITL outright, Lemma provides — on the same audit-trail substrate — a path to record the approval as evidence and then migrate progressively toward human-off-the-loop.

Get started

Make human approval a verifiable fact.