[Paper Review] Power and accountability in reinforcement learning applications to environmental policy

Melissa Chapman, Caleb Scoville+2 more|arXiv (Cornell University)|May 22, 2022

Sustainability and Climate Change Governance4 citations

TL;DR

This paper examines how reinforcement learning (RL) in environmental policy redistributes power and challenges accountability by shifting decision-making authority to private tech firms. It argues that RL's opacity and lack of clear liability mechanisms risk entrenching inequities, and proposes participatory design, open-source algorithms, and legal clarity to ensure equitable and transparent environmental governance.

ABSTRACT

Machine learning (ML) methods already permeate environmental decision-making, from processing high-dimensional data on earth systems to monitoring compliance with environmental regulations. Of the ML techniques available to address pressing environmental problems (e.g., climate change, biodiversity loss), Reinforcement Learning (RL) may both hold the greatest promise and present the most pressing perils. This paper explores how RL-driven policy refracts existing power relations in the environmental domain while also creating unique challenges to ensuring equitable and accountable environmental decision processes. We leverage examples from RL applications to climate change mitigation and fisheries management to explore how RL technologies shift the distribution of power between resource users, governing bodies, and private industry.

Motivation & Objective

• To analyze how RL applications in environmental policy redistribute power among stakeholders, particularly between private industry, governments, and marginalized communities.
• To investigate how the opacity and non-explainability of RL models challenge accountability in environmental decision-making.
• To identify systemic risks in RL deployment, such as hidden biases and self-interest-driven outcomes, especially when private firms control data and algorithms.
• To propose actionable strategies—participatory methods, transparency, and legal frameworks—for ensuring equitable and accountable RL use in environmental governance.
• To bridge gaps between AI ethics, environmental law, and conservation science by addressing the normative implications of RL in ecological management.

Proposed method

• Uses case studies from climate change mitigation and fisheries management to illustrate RL's real-world applications and power dynamics.
• Applies conceptual frameworks from political ecology and algorithmic accountability to analyze how RL refracts existing environmental power structures.
• Proposes participatory methods such as focus groups, surveys, and participatory mapping to include marginalized communities in RL problem formulation.
• Advocates for open-source development of RL algorithms and public documentation of training processes to increase transparency and auditability.
• Recommends collaboration with environmental law experts to clarify legal responsibility and liability for RL-driven environmental decisions.
• Draws on existing literature in algorithmic auditing and fairness to inform methods for monitoring and evaluating RL systems in environmental contexts.

Experimental results

Research questions

• RQ1 How does the deployment of RL in environmental policy shift power dynamics between private industry, governments, and local communities?
• RQ2 In what ways does the opacity of RL models undermine accountability in environmental decision-making, especially in cases of unintended or inequitable outcomes?
• RQ3 How can existing legal frameworks address liability for harmful outcomes resulting from RL-driven environmental policies?
• RQ4 What role can participatory and community-informed design play in ensuring that RL applications reflect diverse values and reduce marginalization?
• RQ5 What mechanisms can ensure transparency and auditability of RL systems when computational resources are concentrated in private hands?

Key findings

• RL applications in environmental policy risk consolidating decision-making power in the hands of private technology firms, especially when they control both data and algorithmic infrastructure.
• The non-explainable nature of deep RL policies creates significant challenges for accountability, as it becomes difficult to trace responsibility for harmful or inequitable outcomes.
• Existing legal doctrines on foreseeability and negligence may not apply effectively to RL systems, especially when outcomes emerge from complex, adaptive interactions with the environment.
• Private ownership of RL platforms can lead to policies that prioritize corporate interests over ecological or social equity, even when such outcomes are not explicitly encoded in the reward function.
• Participatory methods and algorithmic audits can help surface hidden biases and ensure that marginalized communities' values are integrated into RL problem formulation.
• Open-source development and public documentation of RL training processes are essential to enabling scrutiny, reducing opacity, and increasing trust in automated environmental decisions.