Constitutive vs. Corrective: A Causal Taxonomy of Human Runtime Involvement in AI Systems

Recent research from arXiv highlights the ambiguity surrounding terminology for human involvement in AI systems, specifically 'Human-in-the-Loop' (HITL) and 'Human-on-the-Loop' (HOTL). The paper proposes a causal taxonomy to clarify these terms, defining HITL as constitutive (human contribution is necessary for decision output) and HOTL as corrective (human intervention is external to the primary causal chain but can modify or prevent outputs). This clarification aims to improve interdisciplinary collaboration and reduce regulatory uncertainty.

Why it matters

The precise understanding and consistent application of terminology for human involvement in AI are critical for effective governance and responsible development of AI technologies. This clarity can mitigate risks associated with misinterpreting human-AI interaction models, ensuring accountability and ethical considerations are properly embedded in high-stakes decision-making systems.

Key insights

• Existing terminology for human involvement in AI (HITL, HOTL, Human Oversight) is ambiguous.
• This ambiguity complicates interdisciplinary collaboration across fields like computer science, law, and philosophy.
• It also contributes to regulatory uncertainty regarding AI systems.
• A new causal taxonomy distinguishes HITL as 'constitutive,' meaning human input is essential for the AI's decision output.
• HOTL is defined as 'corrective,' indicating human intervention is external to the main AI process but can modify or halt its outputs.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2603.19213