Telling AI exactly how to act — not vaguely appealing to 'be green' — cuts the energy footprint of a GenAI research workflow substantially without changing results. The study shows operational constraints and decision-rule prompts are a practical human-in-the-loop lever to align GenAI productivity with environmental efficiency.

Main Finding

The paper advances a workflow-level view of Green AI and shows empirically that in GenAI-assisted research workflows (here, an LDA-based literature mapping implemented with GenAI-assisted coding) prompt design matters for environmental footprint. Generic “green” language in prompts does not reliably reduce runtime or CO2e, whereas prompts that impose operational constraints and explicit decision rules (i.e., constrain exploration, stopping, and outputs) produce large, stable reductions in runtime and estimated emissions while preserving decision-equivalent topic-model outputs. Conceptually, prompts should be treated as decision policies that allocate discretion between researcher and system and thus are a practical lever for computational governance.

Key Points

• Literature mapping
  • Built a curated arXiv corpus (n = 310) of Green AI papers (filtering with an LLM classifier) and used topic models to map the field into seven themes.
  • The largest cluster in the Green AI literature is training footprint; inference efficiency and system-level optimization are fast-growing themes. Other identified themes include measurement protocols, green algorithms, governance, and security/efficiency trade-offs.
• Conceptual contribution
  • Shift unit of analysis from models to workflows: researchers run pipelines, not isolated models.
  • Treat prompts as decision policies that allocate which computations the system performs and when iteration stops—this allocation governs runtime and CO2e as much as model selection.
• Empirical benchmark
  • Benchmark task: a representative economic-survey pipeline (LDA-based literature mapping) implemented in a fixed cloud notebook with GenAI-assisted coding.
  • Measurements: runtime and estimated CO2e recorded with CodeCarbon; statistical comparisons (paired tests and robustness checks) across prompt strategies.
  • Main experimental result: adding generic green language to prompts yielded no reliable reductions in runtime or emissions. In contrast, prompts embedding operational constraints and explicit decision rules delivered substantial and consistent reductions in both runtime and CO2e, without materially altering the practical topic outputs (decision-equivalent results).
• Scope and limits of measurement
  • Footprint measurements cover execution of the AI-assisted notebook workflow (local/cloud execution and the resulting compute) and do not aim to measure the data-center footprint of the LLMs used for code generation or filtering; that component is discussed as a separate literature and would require different instrumentation.

Data & Methods

• Green AI literature mapping
  • Data source: arXiv search using targeted keywords; initial retrieval n = 410.
  • Filtering: LLM classifier (ChatGPT 5.2) applied to titles/abstracts under explicit inclusion/exclusion rules; manual validation on a random subset to check misclassification.
  • Final corpus: 310 Green AI-focused preprints (methodological, algorithmic, system-level contributions).
  • Analysis: deterministic preprocessing and LDA topic modeling (K tuned; final map reported for K = 7) to extract seven interpretable themes.
• Controlled prompting experiment (benchmark)
  • Task: LDA-based literature mapping (representative survey workflow) implemented in a fixed cloud notebook; code generation and refactoring assisted by GenAI.
  • Treatments (prompt strategies): baseline (no green language), generic green language prompts, and prompts embedding operational constraints/decision rules (explicit stopping rules, output constraints, exploration limits).
  • Instrumentation: runtime tracking and CO2e estimation via CodeCarbon for the executed notebook. Statistical testing included paired t-tests and small-sample robustness checks (see appendices).
  • Output equivalence: because topic models are not identified across runs, authors evaluated “practical” or decision-equivalent equivalence of topic outputs to ensure footprint reductions did not come at the cost of substantively different analytical objects.

Implications for AI Economics

• For researchers and labs
  • Treat prompts as part of computational governance: implement explicit operational constraints and stopping rules in prompts when generating or iterating code to materially reduce compute and emissions.
  • Report workflow-level compute metrics (runtime, energy/CO2e estimates) alongside methodological appendices for reproducibility and sustainability transparency.
  • Use the paper’s benchmark as a replicable template to measure the footprint of other GenAI-assisted workflows (different methods, datasets, or cloud environments).
• For methodology and tool design
  • Green-aligned tooling should support easy expression of decision rules (e.g., iteration caps, output selection, selective caching) and expose workflow-level cost metrics to users in real time.
  • Integrate lightweight footprint monitoring libraries (e.g., CodeCarbon) into research notebooks and pipeline templates as a best practice.
• For policy and evaluation
  • Evaluations of GenAI’s impact on research productivity should incorporate environmental costs at the workflow level, not only model training/inference footprints.
  • Funding agencies and journals could require minimal reporting of computational costs (runtime and CO2e estimates) for large-scale computational studies, especially when GenAI significantly alters production processes.
• For future research
  • Expand measurement to end-to-end footprints that include data-center inference costs of LLMs used during code generation/filtering.
  • Apply the workflow-level prompt-as-policy framework to other common economics tasks (simulation, counterfactual inference, large-scale econometrics) to generalize findings.
  • Investigate automated or UI-level interventions that translate high-level researcher objectives into explicit, low-cost decision rules without harming productivity.

Limitations to keep in mind: the benchmark focuses on one representative survey pipeline (LDA) executed in a specific cloud environment and measures notebook-execution emissions (via CodeCarbon); it does not provide a global estimate of the total footprint of GenAI-enabled research because it excludes model-hosting/data-center inference emissions.