Human-in-the-Loop in AI and Ethics Washing

The key message of the paper is that as AI systems become more autonomous, interconnected, and complex, the traditional oversight mechanism of "Human-in-the-Loop" (HITL) is becoming dangerously obsolete. Instead of providing genuine oversight, HITL is frequently exploited for "AI Ethics Washing"—a practice where organizations create a false illusion of human control to protect corporate stakeholders and unfairly shift accountability onto users when AI systems fail.

Key Facts and Challenges Several critical factors are currently undermining human agency and making HITL difficult to implement:

• Speed and Scale: AI computing and data processing far outpace human cognitive and reaction times, making real-time oversight practically impossible in high-stakes scenarios, as tragically demonstrated by the Boeing 737 MCAS failures.

• Policy Fragmentation: There is no universally accepted definition for HITL. Terms like Human-on-the-Loop (HOTL) or Human-in-Command (HIC) are applied inconsistently across global jurisdictions, leading to ambiguity regarding who is accountable.

• Loss of Agency: Humans are increasingly subjected to algorithmic bias, manipulation, and "black-box" systems where the AI's decision-making process is hidden, stripping users of the ability to challenge or correct automated choices.

Recommended Solution To combat AI Ethics Washing, the authors propose a new governance model that evaluates the credibility of HITL using three diagnostic pillars: Power, Trust, and Complexity.

• Power: Governance must establish genuine human power over the AI system across its entire lifecycle (design, deployment, and decommissioning), rather than just providing an illusion of control.

• Trust: Policymakers must scrutinize whether systems manipulate users into over-trusting machines (e.g., through human-like chatbots), which artificially bypasses human oversight.

• Complexity: As human and AI actions become conjoined, regulations must transparently untangle and assign collective accountability.

Ultimately, the authors call for globally accepted norms for HITL, transparent data provenance, and a stronger focus on AI literacy to ensure ethical human responsibility remains a reality.

Deconstructing using the Diagnostic Pillars

Power This pillar examines the balance of control between the human and the AI. Power is defined as the ability of one agent to get another to do something they normally would not, whether through influence, manipulation, or coercion. The sources distinguish between the "power to" initiate actions and the "power over" social structures and independent actions.

• Genuine Oversight: For HITL to be credible, humans must maintain power over the AI across its entire lifecycle—from design and development to deployment and decommissioning.

• The Risk: As AI systems become more precise and predictable, they gain "personal power" and expertise that can eclipse human capabilities. If human power is only integrated at selective, superficial stages, the oversight is an illusion, shifting control to the AI and enabling AI Ethics Washing.

Trust This pillar evaluates how vulnerability and expectations are managed between humans and machines. Trust is defined as the willingness to be vulnerable based on positive expectations.

• Manipulation of Trust: AI systems can artificially engineer trust through "anthropomorphization" (giving the AI human-like conversational traits) or by providing a systemic illusion of control (such as a user-friendly interface that hides a lack of actual agency).

• The Risk: When an AI performs its tasks with high precision, humans naturally develop trust and are subsequently manipulated into staying out of the loop. As humans cede more trust to the machine, true accountability becomes impossible to attribute, reducing HITL to a façade meant to wash away ethical concerns.

Complexity This pillar assesses the tangled nature of "inter-agency," where decision-making is co-produced by a collective network of people, machines, and programs.

• Levels of Conjoined Agency: The complexity of human-AI collaboration spans four levels: assisting technologies (where AI cannot develop protocols or select actions), arresting, augmenting, and automating technologies (where AI can both develop protocols and select actions).

• The Risk: As systems become highly complex and interconnected, it becomes extremely difficult to trace whether an AI is intentionally or accidentally substituting human routines. When AI gains greater autonomous control over processes and outcomes within this web of complexity, transparent accountability is lost, making it easier for organizations to obscure liabilities and engage in AI Ethics Washing

Want to know more?

✏️ Authors: Melodena Stephens, Dr. Sadaf Khurshid and Mark Esposito, PhD 

Citation: Stephens, M., Khurshid, S., & Esposito, M. (2027). Human-in-the-Loop for Artificial Intelligence: Avoiding AI Ethics Washing. In Human-AI Cognitive Collaboration for Industrial Evolution (pp. 421-454). IGI Global Scientific Publishing.

🛒 Available here: https://lnkd.in/eBxGDpST