Leverage zones in Responsible AI: towards a systems thinking conceptualization

The paper addresses whether current Responsible AI (RAI) initiatives adequately engage with root causes of AI-related harms and how to manage complexity to achieve meaningful change. AI systems increasingly influence critical domains (healthcare, education, justice, finance), raising concerns over fairness, accountability, privacy, transparency, and social inequality. While numerous principles, guidelines, and tools exist across government, industry, academia, and professional bodies, responses are often reactive, fragmented, and sometimes accused of ‘ethics washing’. The authors identify two challenges: (1) disciplinary silos that favor technical fixes over structural or normative change; and (2) a lack of practical systems-thinking guidance for a holistic approach to RAI. To fill this gap, they draw on systems dynamics and adapt Meadows’ leverage points into a pragmatic framework—the Five Ps—to help situate, plan, and align interventions toward Responsible AI.

Conceptual framework development grounded in systems thinking and system dynamics literature. The authors adapt Meadows’ (1999) leverage points into ‘leverage zones’ organized within a Five Ps framework comprising two domains (Problem and Response) and four intervention zones with increasing leverage: Parameter, Process, Pathway, and Purpose. The framework: (1) prompts actors to situate the Problem appropriately; (2) categorizes interventions by leverage zone; and (3) visualizes effort versus type of change (incremental to transformational) (Figures 1–2). Each zone is defined: Parameter (tweaks to algorithms/parameters); Process (feedbacks, social-technical processes in design/deployment); Pathway (information flows, rules, governance structures, power); Purpose (norms, values, goals, paradigms). The authors provide a question set for each zone (Table 2) to guide analysis and planning. An illustrative use-case (social media misinformation/extremism) demonstrates how interventions vary across zones and can interact, and how deeper-purpose shifts (e.g., from maximizing engagement to consensus or social cohesion) differ from parameter/process fixes. No empirical data collection or controlled experiments were conducted; rather, the framework is positioned as a scaffold for hypothesis generation, planning, and transdisciplinary dialogue.

• Primary contribution: the Five Ps framework (Problem, Parameter, Process, Pathway, Purpose) to conceptualize, analyze, and plan Responsible AI interventions using a systems-thinking lens.
• Leverage gradient: Lower zones (Parameter, Process) yield incremental changes; higher zones (Pathway, Purpose) enable structural and transformational change. Deeper zones shape and constrain the options available in shallower zones.
• Interdependencies: Interventions in one zone propagate via feedbacks to others; effective change typically requires coordinated, multi-zone strategies rather than siloed actions (e.g., fairness tools or principles alone).
• Practical guidance: Table 2 provides lines of questioning per zone to translate abstract principles into concrete considerations (e.g., quantifying parameters, adjusting processes, modifying rules and information flows, and interrogating goals/assumptions).
• Analytical and planning utility: The framework aids in aligning initiatives with desired change magnitude, anticipating unintended consequences, prioritizing efforts, and fostering transdisciplinary communication among stakeholders.
• Illustrative example: Addressing social media misinformation shows that parameter tweaks and process changes may improve moderation but leave business-model paradigms intact, whereas purpose-level shifts (e.g., prioritizing consensus over engagement) can reorient system goals and outcomes.

The Five Ps framework addresses the need for holistic, practical guidance in Responsible AI by enabling stakeholders to frame problems and interventions across multiple leverage zones. It emphasizes moving beyond isolated technical fixes to consider processes, governance structures, and foundational purposes that drive AI system behavior. By situating interventions within zones and anticipating their interactions, decision-makers can evaluate effectiveness (doing the right things) rather than only efficiency (doing things right). The social media case highlights how business models and paradigms can limit the adoption of deeper changes, explaining why tools and principles alone often fall short. The planning questions cultivate reflexivity, anticipation, and inclusivity, supporting coordinated efforts across disciplines and sectors. The framework also clarifies the role of regulation and governance (Pathway) and value alignment (Purpose) in achieving durable improvements, while recognizing that unified regulation remains nascent and partnerships often produce high-level but non-specific guidance. Overall, the Five Ps operationalizes systems thinking for RAI by linking problem framing to multi-level, mutually reinforcing interventions.

Responsible AI requires engaging with deeper system structures and purposes rather than relying solely on surface-level technical adjustments. The Five Ps framework offers a straightforward systems-thinking tool to analyze and plan interventions, align initiatives across stakeholders, reveal interdependencies, and anticipate long-term effects. Conceptually simple, it encourages structured question-asking and transdisciplinary dialogue, helping avoid fragmented, siloed approaches. Future work should apply the Five Ps in real-world case studies to evaluate short- and long-term outcomes, refine guidance on cross-zone coordination, and further develop practical methods to translate principles into structural and purpose-level change.

• Conceptual nature: No empirical or longitudinal validation is provided; the framework’s effectiveness in practice remains to be demonstrated through real-world applications and case studies.
• Evaluation constraints: Controlled experiments are infeasible in complex socio-technical systems, limiting causal attribution of outcomes to specific leverage-zone interventions.
• Early stage: Application of systems thinking to Responsible AI is in its infancy; additional methodological development and evidence are needed.
• Contextual dependencies: Organizational paradigms (e.g., profit-driven engagement goals) and the current lack of unified regulation may constrain adoption of higher-leverage (Pathway/Purpose) interventions.