Emissions transparency in AI research should be mandatory

Overview

This study evaluates the feasibility of implementing mandatory emissions disclosure requirements for artificial intelligence research submissions at academic venues. The authors develop a policy-level analytical framework that assesses the operational burden, coverage rates, and uncertainty characteristics of three tiered disclosure policies, ranging from voluntary reporting to policies incorporating random audits. The analysis employs Monte-Carlo simulation to model realistic implementation scenarios and quantifies key metrics including completion time, reviewer effort, and emissions estimation uncertainty under varying assumptions about instrumentation and reporting standards.

Methods and approach

The research formalizes three tiered venue policies: P0 (optional disclosure), P1 (mandatory disclosure using a standardized template for Tier 1 and 2 venues), and P2 (P1 plus 10% random audits for Tier 2 submissions). A minimal disclosure template requiring specification of hardware, training duration, energy consumption or CO2-equivalent emissions, and emission-factor source is proposed as the core reporting requirement. Monte-Carlo simulation is used to model template completion time, first-pass sufficiency rates, reviewer and editorial workload, and propagation of uncertainty across key parameters including power usage effectiveness (PUE), emission factors, and metering accuracy. Uncertainty is quantified through relative half-width estimates for both location-based and market-based emissions accounting approaches.

Key Findings

Implementation of the minimal disclosure template requires median completion time of approximately 10.8 minutes (interquartile range 8.1-14.8 minutes) per submission, with reviewer verification requiring approximately 1.6 minutes per paper. Tiered audit policies yield coverage increases from approximately 25% under voluntary disclosure (P0) to approximately 80% under mandatory policies (P1/P2), with editorial audits for P2 requiring approximately 24.1 minutes per 100 submissions. Uncertainty analysis demonstrates that decision-useful emissions intervals can be generated using lightweight assumptions, with median relative half-widths of approximately 0.33 for location-based accounting and approximately 0.77 for market-based accounting. Emission factors and PUE estimates constitute the dominant sources of residual uncertainty, with metering and device-specific draw effects contributing minimally. Hypothesis testing indicates that policy requirements H1-H3 are met under baseline priors, while H4b (market-based) is satisfied though H4a (location-based) narrowly falls short.

Implications

Mandatory emissions disclosure emerges as operationally feasible at publication time when implemented through tiered venue policies with proportionate verification mechanisms. The proposed framework enables systematic comparison of transparency policies without requiring proprietary telemetry access or speculative estimation approaches, providing actionable decision support for research venues and policymakers. Implementation of the minimal disclosure template can achieve substantial coverage gains—from approximately 25% to 80%—while imposing modest administrative overhead on authors, reviewers, and editorial staff, suggesting that operational burden is not a significant constraint on adoption.