Emissions transparency in AI research should be mandatory

The rapid scaling of computational workloads in artificial intelligence (AI) has raised concerns about the energy use and carbon emissions associated with model training. While several prior studies have called for improved reporting of environmental impacts, emissions disclosure in AI research remains largely voluntary, inconsistent, and difficult to verify. This paper evaluates the feasibility, burden, and uncertainty properties of mandatory emissions disclosure at AI research venues. Rather than estimating the emissions of specific models, we develop a policy-level analytical framework that models disclosure requirements, reviewer and editorial workload, and uncertainty propagation under realistic instrumentation assumptions. We formalize tiered venue policies – P0 (optional disclosure), P1 (template for Tier 1/2), and P2 (P1 plus 10% random audits for Tier 2) – and test explicit hypotheses using Monte-Carlo simulation. We quantify template completion time, first-pass sufficiency, reviewer/editor effort, and uncertainty interval width for reported emissions. Our results indicate that a minimal disclosure template – requiring hardware, duration, energy or CO 2 e, and emission-factor source – can achieve high coverage with modest additional burden (median completion time ≈ 10 . 8 min; IQR 8.1–14.8; reviewer checklist ≈ 1 . 6 min/paper; P2 editorial audits ≈ 24 . 1 min per 100 submissions). Coverage rises from ∼ 25% of submissions (P0) to ∼ 80% (P1/P2). Uncertainty analysis shows that decision-useful emissions intervals can be reported using lightweight assumptions, with median relative half-widths R ≈ 0 . 33 (location-based) and R ≈ 0 . 77 (market-based); the emission-factor band dominates residual uncertainty, followed by power usage effectiveness (PUE), with metering/device-draw effects smaller. Under baseline priors, H1–H3 are met; H4b (market-based) is met while H4a (location-based) is narrowly missed. We conclude that mandatory, uncertainty-aware emissions disclosure is operationally feasible at publication time when implemented through tiered requirements and light-touch verification. The framework offers venues and policymakers decision support for comparing transparency policies without relying on proprietary telemetry or speculative “big-number” estimates. • Mandate publication-time emissions transparency for AI training runs. • Tiered disclosure+verification (P0/P1/P2): self-report to light audit. • Compact Model Emissions Datasheet with uncertainty-aware C O 2 e intervals. • Monte Carlo case study: high compliance gains with low overhead. • Near-universal disclosure enables comparable, reproducible emissions reports.