Position: AI Must Become Planet-Centered, Not Just Human-Centered

Technical Summary: Position: AI Must Become Planet-Centered, Not Just Human-Centered

Problem Statement
The paper argues that contemporary AI paradigms, including Human-Centered AI (HCAI), Responsible AI, and AI for Social Good, are insufficient for addressing the complex, systemic challenges of the Anthropocene. Current frameworks primarily focus on protecting individual humans from harm, optimizing for narrow performance metrics, and operating within stable, well-defined problem spaces ("tame problems"). However, planetary challenges—such as climate change, biodiversity loss, and resource depletion—are "wicked problems" characterized by non-stationarity, deep uncertainty, feedback loops, and contested objectives.

The core technical mismatch identified is that standard AI assumptions (stationarity, sample independence, closed-world formulations, and fixed objective functions) structurally violate the dynamics of socio-ecological systems. When AI is deployed in these "wicked" contexts without a systems perspective, it risks amplifying systemic instability through mechanisms such as:

• Technical Misalignment: Models optimized for local efficiency fail to account for endogenous system changes and regime shifts.
• Narrow Value Framing: Progress is equated with optimization and expansion, often ignoring distributional fairness, resilience, and ecological limits.
• Impact Blind Spots: Evaluation is confined to task-level metrics, missing indirect, cumulative, and long-horizon effects (e.g., rebound effects where efficiency gains lead to increased total consumption).
• Risk Amplification: AI can accelerate feedback loops, lock societies into unsustainable trajectories, and propagate failures across coupled infrastructures.

Methodology and Framework
This is a position paper that proposes a new design philosophy and research agenda called Planet-Centered AI (PCAI). Rather than presenting new experimental results, the paper employs a diagnostic and conceptual methodology:

1. Diagnostic of Wickedness: The authors adapt frameworks from systems science (e.g., Rittel & Webber, 1973; Hisschemöller & Hoppe, 2018) to classify problems based on "knowledge certainty" (understanding of causal dynamics) and "value consensus" (agreement on objectives). This diagnostic maps specific AI failure mechanisms to the properties of wicked systems.
2. Lifecycle Reshaping: PCAI is defined not as a single algorithm but as a reorientation of the entire AI lifecycle. The paper outlines how this philosophy transforms four key stages:
   • Problem Setting: Moving from isolated task formulation to "system mapping" and "theories of change" that explicitly define system boundaries, feedbacks, and intervention pathways before modeling begins.
   • Model Design: Shifting from prediction and control to "foresight" and "epistemic infrastructure." This involves prioritizing robustness under deep uncertainty, non-stationarity, and regime shifts. It advocates for AI as a tool for collective sensemaking (e.g., scenario analysis, exploratory modeling) rather than automated decision-making.
   • Evaluation: Replacing single-score benchmarking with "trajectory-oriented evaluation." This includes Pareto-frontier analysis to surface trade-offs, systemic risk probes (testing for rebound effects), and counterfactual baselines (comparing against no-AI or simple heuristic interventions).
   • Deployment: Treating deployment as a provisional, revisable intervention. This requires system-level monitoring (beyond data drift) to detect behavioral adaptation and early warning signals, with predefined pathways for rollback or modification.

Key Contributions

• Conceptual Framework (PCAI): The paper introduces Planet-Centered AI as a distinct paradigm that treats the integrity, resilience, and limits of ecological systems as first-class design constraints, rather than externalities inferred through human outcomes.
• Diagnostic Mapping: It provides a structured mapping of how specific properties of wicked systems (e.g., non-stationarity, path dependence, emergent behavior) conflict with standard AI assumptions, offering a taxonomy for identifying when standard methods are likely to fail.
• Operational Guidelines: The paper offers concrete actions for researchers, funders, and institutions (summarized in Table 1), such as mandating system mapping, requiring comparative baselines, and redefining scientific contribution to include system-level stress-testing.
• Falsifiable Claim: The paper advances a specific, testable hypothesis: In domains governed by wicked dynamics, AI systems optimized for efficiency or narrow objectives without explicit consideration of systemic feedbacks are more likely to exacerbate systemic instability than to mitigate it.

Results and Claims
As a position paper, the work does not report empirical results from new experiments. Instead, its "results" are the logical synthesis of existing literature and case studies (e.g., autonomous vehicles, precision agriculture, conservation monitoring) that demonstrate the limitations of current approaches.

• Evidence of Failure: The paper synthesizes evidence showing that AI-driven efficiency often triggers rebound effects (e.g., autonomous vehicles increasing total travel demand) and that narrow evaluation metrics fail to capture systemic risks like biodiversity loss or social legitimacy erosion.
• Theoretical Necessity: The paper argues that current AI safety research (focused on internal alignment and control) is necessary but insufficient because it does not address how well-aligned systems reshape external socio-ecological dynamics.

Significance
The paper claims that the significance of PCAI lies in its potential to align AI development with the defining challenges of the Anthropocene. By shifting the unit of analysis from the individual user to the coupled social-ecological system, PCAI aims to:

1. Prevent AI from inadvertently reinforcing "Anthropocene traps" (self-reinforcing trajectories of instability).
2. Transform AI from a tool of optimization into a tool for "collective intelligence" and "foresight," enabling societies to navigate deep uncertainty and contested values.
3. Provide a rigorous, falsifiable framework for evaluating whether AI contributes to planetary stability or exacerbates systemic risk.

The authors position PCAI not as a replacement for Human-Centered AI, but as a necessary extension that addresses the temporal horizons and systemic boundaries that HCAI often overlooks. The ultimate goal is to equip societies with AI tools capable of supporting responsible intervention in complex, high-stakes planetary challenges.