Heni Rothstein Cohen (Tel Aviv University – Buchmann Faculty of Law) has posted Human Roles Across the AI Lifecycle: Shifting Beyond the “Human Supervision” Paradigm, Northwestern Journal of Technology and Intellectual Property, Vol. 24, forthcoming on SSRN. Here is the abstract:
Human supervision is widely regarded as a cornerstone of trustworthy AI in decision-making. To mitigate AI-related risks, such as inaccurate, discriminatory, opaque, or unaccountable decisions, existing legal frameworks and scholarly debate rely on human review of AI outputs. Yet the current model suffers from two fundamental problems. First, legal frameworks lack a clear articulation of what human supervisors are meant to achieve: human presence is treated as a safeguard without asking what objectives it serves or why a human is involved at all. Second, supervision is structured as a post-hoc review of AI outputs, preventing humans from applying their unique capabilities—contextual reasoning, moral judgment, tacit knowledge—at the stages where they could actually shape outcomes. Extensive empirical evidence confirms that after-the-fact human review of AI decisions is largely ineffective. Supervisors tend to over-rely on AI outputs or dismiss them altogether, without any meaningful capacity to critique them. This Article proposes an alternative understanding of human involvement in AI systems. On the conceptual level, it advances a clear articulation of human involvement objectives, distinguishing between those that merely place a human “in charge” to offer reassurance without enabling any meaningful contribution, and “AI betterment” objectives—a term proposed here to capture goals that genuinely seek to improve AI outputs through human capabilities that AI systems lack. On the structural level, the Article proposes a shift from posthoc human supervision to “lifecycle-based human involvement”, reconceptualizing human roles as multiple differentiated functions distributed across the full AI lifecycle, from problem definition, system design, and training, through deployment and post-deployment adaptation. The Article identifies the specific roles humans perform, or should perform, at each stage, linking them to human capabilities and to practices that can more effectively advance AI betterment objectives. Building on this framework, the Article concludes by outlining the legal mechanisms through which lifecycle-based human involvement can be institutionalized, across sectoral regulation, tort liability, and AI governance. This discussion offers a preliminary account of the role-differentiated standards of care that effective and enforceable human involvement in AI systems requires.
Recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum