Human-on-the-Loop: AIの監督と自律性の新モデル

主要なポイント

• Human-in-the-loop (HITL) では、AI の行動ごとに人間の承認が必要となる。Human-on-the-loop (HOTL) では、AI が自律的に行動し、人間は監視を行い、必要に応じて介入するのみとなる

• HITL から HOTL への移行は、3 つの力によって推進されている。AI エージェントが独立して行動できるほど能力が高まっていること、AI の展開規模が行動ごとの人間によるレビューを不可能にしていること、そして自動化を求める経済的圧力が 2023〜2025 年の安全第一の姿勢を圧倒していること

• Anthropic の Pentagon 拒否は、議論の極端な事例を示した。HOTL は請求処理には許容されるが、一部の企業の見解では、機密軍事用途には許容されない

• AI を展開するすべての組織にとっての実践的な問いは、「人間をループに入れるべきか」ではなく、「リスク、コスト、結果のどの閾値で人間が介入するのか」である

Human-in-the-Loop vs Human-on-the-Loop 、明確な区別

Human-in-the-loop (HITL) は、より古く、より広く理解されているモデルである。AI が推奨事項を作成したり出力を生成したりし、人間がそれをレビューして承認してから行動が実行される。人間はゲートであり、ゲートが開かなければ何も起こらない。

このモデルは、AI の行動が離散的で、頻度が低く、個別に重大な結果をもたらす場合に有効である。AI がフラグを立てた異常を放射線科医がスキャンでレビューする場合。AI が生成した信用判断を融資担当者が承認または却下する場合。パイロットがオートパイロットの推奨を受け入れるか上書きする場合。これらのケースでは、AI は支援するが行動はしない。人間が意思決定者であり続ける。

Human-on-the-loop (HOTL) は、これらの条件が崩れたときに現れるモデルである。AI は継続的な意思決定の流れに対して自律的に行動し、人間は監督者の立場から監視する。人間はゲートではない。人間はオーバーライド権限を持つ観察者である。この区別は理論上は微妙だが、実践上は非常に大きい。

A Deloitte analysis of enterprise AI deployments describes the shift precisely: 「The shift to human on the loop is a form of automation that requires strong AI agent governance and oversight. Multi-agent systems can be continuously trained to improve, but without proper transparency and monitoring, they can also go haywire.」 重要なフレーズは「go haywire」である。HITL は、人間が承認しなければならないため、悪い行動が発生するのを防ぐ。HOTL は、悪い行動が発生した後にそれを検知し、人間が十分に迅速に介入して被害を抑えることに依存する。この違いは哲学的なものではない。ミスを防ぐことと、ミスを捉えることの違いである。

The distinction maps cleanly onto the difference between a 2023 AI tool and a 2026 AI agent. ChatGPT in 2023 was HITL: every response required a human prompt, and the human decided whether to use the output. Google Gemini Intelligence in 2026 is HOTL: it books a restaurant reservation across multiple apps, fills forms, confirms the booking, and reports back. The human provides the initial instruction and receives the confirmation. The steps in between happen without gates.

Why the Shift Is Happening Now

*Three structural forces are pushing AI systems from HITL to HOTL, and none of them are slowing down.

Capability Has Crossed the Threshold

The first force is technical. In 2023, AI models could generate text and images. They could not reliably complete multi-step tasks that required navigating interfaces, filling forms, and making judgment calls across applications. In 2026, they can. Claude Code autonomously writes, tests, and deploys code across a full project. Gemini Intelligence navigates Chrome to complete a reservation without the user touching the screen. OpenAI Codex controls desktop applications.

When the AI's capability was limited to generating suggestions, HITL was the natural model. The AI proposed, the human decided. When the AI can execute the entire workflow, the human becomes the bottleneck. A human reviewing every invoice Claude chases, every lead it triages, every code change it commits is not adding safety. They are adding latency that defeats the purpose of using the AI. The capability has crossed a threshold where HOTL is not just acceptable , it is the only model that allows the AI to deliver its full value.

Scale Makes Gates Impossible

The second force is deployment scale. GitHub Copilot is now a co-author on four million commits. Google Gemini Intelligence is rolling out to 30 billion Android devices. AI agents inside enterprise software process millions of transactions per day. At that scale, per-action human review is mathematically impossible. You cannot hire enough people to approve every AI-generated line of code, every AI-suggested email response, every AI-triaged support ticket.

The shift from HITL to HOTL is partly a technological choice and partly a mathematical inevitability. When an AI system processes more decisions in an hour than a human team could review in a year, the question is not whether to use HOTL. The question is how to design HOTL systems that fail safely.

Economics Are Overwhelming the Safety Posture

The third force is economic. The AI industry spent $300 billion on venture funding in Q1 2026 alone. Companies that raised money on the promise of autonomous AI agents are now under pressure to ship autonomous AI agents. The 2023-2025 posture , "we're being careful, we're testing, safety comes first" , is colliding with the 2026 reality of revenue targets, customer expectations, and competitive pressure.

This is the context in which Anthropic's Pentagon refusal should be understood. It was not just a safety decision. It was a line-drawing exercise. Anthropic was saying: HOTL is acceptable for invoice chasing. It is not acceptable for military applications where the "loop" between autonomous action and human intervention could be measured in seconds with irreversible consequences. The company drew a line. The rest of the industry is watching to see where everyone else draws theirs.

The Oversight Problem , What Happens When the Human Is No Longer the Gate

The defining challenge of HOTL is not technological. It is cognitive.

When a human reviews every AI decision, the human is engaged. Each action requires attention, however briefly. When a human monitors a stream of AI decisions from a supervisory position, the human must sustain attention on a system that mostly works correctly. This is a vigilance task, and humans are famously bad at vigilance tasks.

The research on automation complacency is decades old and unambiguous. When an automated system performs reliably more than 90 percent of the time, human operators begin to disengage. Reaction times to anomalies increase. The ability to detect subtle errors degrades. The operator's mental model of the system's behavior becomes increasingly outdated as the system evolves while the operator's attention is elsewhere. This was well documented in aviation autopilot research in the 1990s. It is being rediscovered in AI agent deployment in 2026.

The practical implication is that HOTL systems require a different kind of oversight infrastructure than HITL systems. HITL needs approval workflows. HOTL needs anomaly detection, alerting thresholds, intervention playbooks, and simulation environments where operators can practice responding to rare failure modes. Most organizations deploying AI agents in 2026 have the approval workflows. Almost none have the rest.

Deloitte's agent observability framework identifies three layers of HOTL oversight: transparency (can the human see what the agent is doing in real time?), traceability (can the human reconstruct why the agent made a specific decision after the fact?), and controllability (can the human intervene at the right level of granularity , pause, redirect, override, rollback?). Most deployed AI agents in 2026 score well on controllability (you can turn them off) and poorly on transparency and traceability (you cannot see inside the decision process in real time).

Where HOTL Works , and Where It Does Not

The appropriateness of HOTL depends on two variables: the cost of a wrong decision and the reversibility of that decision.

Invoice chasing is an ideal HOTL use case. The cost of a wrong email is low (an annoyed vendor). The action is reversible (you can send a correction). The human can monitor at the level of exceptions , "flag anything unusual and I'll review it" , rather than at the level of individual actions.

Autonomous weapons are the extreme counterexample. The cost of a wrong decision is measured in lives. The action is irreversible. The time between decision and consequence may be too short for human override. This is the use case Anthropic refused to enable, and it is the use case that makes HOTL a genuinely frightening concept to anyone who studies the automation complacency literature.

Between these extremes lies most of the economy. Medical diagnosis AI: high cost, partially reversible (a misdiagnosis can be corrected but harm may already be done). Financial trading AI: high cost, irreversible at market speed. Legal document review: medium cost, reversible. Customer service AI: low cost, reversible. Each domain requires its own threshold for when the human steps in, and that threshold should be set by the domain's error tolerance, not by the AI vendor's default configuration.

Human-on-the-Loop in Practice , How remio Navigates the Boundary

remio occupies an unusual position in the HOTL landscape. It is an AI system that acts autonomously in one dimension , it passively captures everything you do, every meeting, every browse session, every document , and requires human judgment in another , what you ask it, how you verify its answers, and what you do with the retrieved information.

The capture side is fully HOTL. remio records without asking. There is no gate because a gate would defeat the purpose: if you had to decide what to save, you would miss the thing you later needed. The retrieval side is closer to HITL. You ask a question. remio surfaces relevant information from your archive. You decide whether the answer is accurate, complete, and useful. The loop closes when you act on the information , or when you realize the information is wrong and refine your query.

This hybrid model , autonomous capture, supervised retrieval , is a microcosm of where HOTL is headed across the industry. The tasks that are high-volume, low-cost-of-error, and burdensome for humans (capturing everything) are automated. The tasks that require judgment, context, and accountability (deciding what matters and what to do with it) remain human-supervised. The knowledge blending architecture that remio uses to connect information across sources is itself a HOTL design: the AI finds the connections, the human evaluates their relevance.

FAQ: Common Questions About Human-on-the-Loop

Q: Is human-on-the-loop less safe than human-in-the-loop?

A: It depends on the domain and the oversight infrastructure. In high-volume, low-consequence domains where per-action human review would introduce unacceptable latency, HOTL can be safer than a poorly implemented HITL system where the human becomes a rubber stamp. In high-consequence, irreversible domains, HITL remains the appropriate model. The danger is not HOTL itself. It is deploying HOTL without the monitoring, alerting, and intervention infrastructure that makes it safe.

Q: How is human-on-the-loop different from fully autonomous AI?

A: In a fully autonomous system, there is no human oversight. The AI makes decisions and executes actions with no mechanism for human intervention. HOTL preserves human oversight but relocates it from before the action to during or after. The human remains accountable. The difference is the timing of the intervention, not its existence.

Q: What industries are moving fastest toward HOTL?

A: Software development (AI coding agents), customer service (AI chatbots handling full resolution paths), and financial operations (automated invoice processing, fraud detection) are the leading adopters. These industries share two characteristics: high transaction volumes that make HITL impractical, and error costs that are generally financial rather than physical.

Q: HOTLは雇用喪失を意味しますか？

A: HOTLは仕事をなくすのではなく、役割を変えます。個別の取引を承認していたオペレーターは、システム全体のパフォーマンスを監視し、例外を処理する監督者になります。必要なスキルは実行速度から診断能力に移行します。これは製造業の自動化や航空のオートパイロットで起こったのと同じパターンです。システムを直接操作する人の数は減少しますが、残る人のスキルレベルは向上します。

Q: 規制は特定のAIアプリケーションにhuman-in-the-loopを要求しますか？

A: EU AI Actはすでに高リスクAIシステムに人間による監督を義務付けていますが、その仕組みは指定していません。2026年の議論は、「人間による監督」がHOTLで満たされるのか、それともHITLが必要なのかという点です。答えはセクターによって異なる可能性が高いでしょう。医療機器ではHITLが求められる一方、金融不正検知ではHOTLが認められ、自律型兵器は全面禁止されるかもしれません。規制の枠組みはまだ策定中であり、2026-2027年に下される決定が、今後10年間のAI業界を形作ることになります。