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

Based on the paper "Constitutive vs. Corrective: A Causal Taxonomy of Human Runtime Involvement in AI Systems" by Kevin Baum and Johann Laux, here is a detailed technical summary.

1. Problem Statement

The terminology surrounding human involvement in AI systems—specifically "Human-in-the-Loop" (HITL), "Human-on-the-Loop" (HOTL), and "Human Oversight"—is currently plagued by semantic ambiguity. This "Tower of Babel" creates significant friction in interdisciplinary collaboration (between computer science, law, philosophy, psychology, and sociology) and leads to regulatory uncertainty.

• The Core Issue: Current definitions often rely on vague spatial metaphors (e.g., being "in" or "on" a loop) rather than precise causal mechanisms.
• Consequence: Legal frameworks, such as the EU AI Act and GDPR, struggle to operationalize requirements like "meaningful human oversight" because it is unclear whether a human is merely a rubber-stamp (constitutive but passive) or an active supervisor (corrective).
• Gap: Existing literature often conflates the presence of a human with effective oversight, failing to distinguish between a human who is causally necessary for a decision and one who merely has the capacity to intervene.

2. Methodology

The authors propose a conceptual taxonomy grounded in causal structure rather than spatial metaphors.

• Scope: The analysis focuses on runtime involvement (operational deployment) rather than pre-deployment design or post-hoc auditing, though it acknowledges the permeability of these boundaries.
• Framework: The authors adopt the EU AI Act's distinction between Providers (developers) and Deployers (users) to create a nested temporal model of system lifecycles.
• Approach: Instead of a broad literature review, the authors synthesize existing concepts to propose a new, operationally useful taxonomy. They utilize counterfactual reasoning (asking "what would happen if the human were absent?") to define causal roles.
• Dimensions: The taxonomy is built on two orthogonal dimensions:
  1. Causal Structure: Constitutive vs. Corrective agency.
  2. Cognitive Integration: Complementary vs. Hybrid intelligence.

3. Key Contributions and Taxonomy

A. The Central Causal Distinction: HITL vs. HOTL

The authors redefine HITL and HOTL based on the human's role in the causal chain of decision-making:

• Human-in-the-Loop (HITL) – Constitutive Agency:
  • Definition: A human is a necessary link in the causal chain. Without their contribution, the system cannot generate an output.
  • Metaphor: The human is a "Gate." The process pauses and waits for human input to proceed.
  • Implication: The human is causally required for every decision output.
• Human-on-the-Loop (HOTL) – Corrective Agency:
  • Definition: The system generates outputs autonomously by default. The human is external to the primary causal chain but possesses the capacity to prevent, modify, or override outputs.
  • Metaphor: The human is a "Switch." They are not needed for the process to run but can redirect it.
  • Implication: The human is causally necessary only if they intervene; otherwise, the system operates independently.

B. Temporal Gradations within HOTL

The authors distinguish three temporal modes of HOTL, which are critical for understanding intervention capacity:

1. Synchronous HOTL: Real-time monitoring with immediate intervention capability. Upon intervention, the human becomes causally necessary for that specific outcome (transitioning to an outcome-HITL for that token).
2. Asynchronous HOTL: Periodic review (e.g., logs, audits). Interventions are delayed and target future system behavior (e.g., retraining, parameter adjustment) rather than the current decision.
3. Anticipatory HOTL: Prospective influence where humans set normative constraints or alignment parameters before operation. This shapes the "space of permissible actions" rather than individual decisions. It operates within the provider's runtime but targets the deployer's decision space.

C. Cognitive Integration: Complementary vs. Hybrid

A second dimension describes the cognitive relationship between human and machine:

• Complementary Intelligence: Human and AI work side-by-side with separate cognitive architectures. They coordinate but do not share a unified cognitive state (e.g., a radiologist reviewing an AI-flagged mammogram).
• Hybrid Intelligence: Human and machine form an entangled cognitive unit. The human's reasoning is partially constituted by the machine's representations (e.g., a surgeon using AR overlays where the visual perception is inseparable from the AI data).
• Interaction: The authors map these against the causal roles, noting that while Hybrid Intelligence is often associated with HITL, it can exist in HOTL configurations (e.g., a clinician whose understanding of risk is reshaped by an AI system they oversee).

D. Redefining "Human Oversight" (HO)

The paper argues that Human Oversight is a specific normative mode of HOTL, not a synonym for HITL.

• Definition: A human overseer must be systematically prepared, positioned, and capable of conscious monitoring and intervention to substantially reduce risks.
• Key Insight: A formal HITL configuration (where a human must click "approve") can fail to provide oversight if the human acts as a "rubber stamp." Conversely, a well-designed HOTL system can provide stronger oversight than a perfunctory HITL system.
• Role Duality: The authors address the tension where a single human may occupy both HITL (constitutive) and HOTL (oversight) roles simultaneously. They argue this creates "automation bias" and role conflict, requiring architectural mitigation (e.g., layered oversight structures) rather than mere acknowledgment.

4. Results and Implications

• Clarification of Regulatory Requirements: The taxonomy clarifies that Article 22 GDPR (no "solely automated processing") often implies a HITL structure, while Article 14 of the AI Act (requiring "effective oversight") demands a robust HOTL structure with genuine capacity for intervention.
• Operationalization: The framework provides a checklist for system architects and regulators to assess whether a system meets normative goals. It moves the conversation from "Is a human present?" to "What is the human's causal role and cognitive state?"
• Design Guidance: The paper highlights that moving from HITL to HOTL does not reduce cognitive load; it changes the nature of the task, often making it more demanding (the "ironies of automation"). It suggests that role duality (one person doing both HITL and HOTL) is a design problem that requires specific mitigation strategies.

5. Significance

This paper is significant because it moves the discourse on Human-AI interaction from metaphorical descriptions to causal and architectural specifications.

• Interdisciplinary Bridge: It provides a common language for computer scientists (who design causal chains), legal scholars (who define oversight requirements), and ethicists (who assess agency and responsibility).
• Regulatory Precision: It offers a tool to interpret and enforce regulations like the EU AI Act, distinguishing between systems that merely have a human in the loop and those that provide effective human oversight.
• Future-Proofing: By introducing "Anticipatory HOTL," the framework addresses the challenges of agentic AI systems where reactive oversight may be too slow, emphasizing the need for proactive constraint-setting.

In summary, Baum and Laux argue that to ensure safe and compliant AI, we must stop asking if humans are "in the loop" and start analyzing how they are causally and cognitively integrated into the system's decision-making architecture.