cs.SEMar 17, 2026
SaveOn the Carbon Footprint of Economic Research in the Age of Generative AI
Organizations: Banco de Espa˜na · Universidad de Castilla-La Mancha
Abstract
Generative artificial intelligence (AI) is increasingly used to write and refactor research code, expanding computational workflows. At the same time, Green AI research has largely measured the footprint of models rather than the downstream workflows in which GenAI is a tool. We shift the unit of analysis from models to workflows and treat prompts as decision policies that allocate discretion between researcher and system, governing what is executed and when iteration stops. We contribute in two ways. First, we map the recent Green AI literature into seven themes: training footprint is the largest cluster, while inference efficiency and system level optimisation are growing rapidly, alongside measurement protocols, green algorithms, governance, and security and efficiency trade-offs. Second, we benchmark a modern economic survey workflow, an LDA-based literature mapping implemented with GenAI assisted coding and executed in a fixed cloud notebook, measuring runtime and estimated CO2e with CodeCarbon. Injecting generic green language into prompts has no reliable effect, whereas operational constraints and decision rule prompts deliver large and stable footprint reductions while preserving decision equivalent topic outputs. The results identify human in the loop governance as a practical lever to align GenAI productivity with environmental efficiency.
Figures & tables
| RQ | Objective | Motivation |
| RQ1 | Quantify how alternative prompting strategies change the computational footprint of a decision-equivalent survey workflow, measured by runtime and CO 2 e. | GenAI lowers the cost of producing and iterating code, which can increase throughput. The environmental cost of this productivity gain is not well measured at the workflow level, where repeated execution and redundant computation can arise. |
| RQ2 | Test whether generic green language is sufficient to reduce footprint, or whether operational constraints and explicit decision rules are required to deliver stable reductions in runtime and emissions. | In practice, researchers often rely on high level instructions when interacting with LLMs. If such instructions do not reliably change what the system computes, then meaningful footprint reductions require governance through explicit constraints, not only normative appeals. |
| RQ3 | Assess practical output equivalence across prompting strategies by testing whether reduced footprint comes with any substantive change in topic model outputs. | Footprint reductions are only meaningful if the analytical object is preserved. Because topic models are not identified across runs, we need a transparent equivalence check to separate genuine efficiency gains from changes in the underlying clustering of documents. |
| RQ4 | Conceptualise and test prompts as decision policies that allocate discretion between the researcher and the model, shaping the breadth of search and the amount of computation executed in AI-assisted research. | The key margin is not whether an LLM is used, but how human machine interaction is structured. Prompt design can be treated as part of computational governance, since it determines stopping rules, search scope, and which outputs are computed and stored. |
Table 1: Research questions, objectives, and motivations.
Figure 1: Identification and screening pipeline for the Green AI literature map. Counts reflect our arXiv retrieval and filtering protocol.
Figure 2: Model-selection diagnostics for the Green AI corpus: topic coherence ( , higher is better) and held-out perplexity (lower is better) for LDA models estimated over . Each point corresponds to one fitted model on the preprocessed abstract corpus; the main specification uses based on the coherence plateau and interpretability checks.
| Topic | Label | Share of tokens (%) |
| 1 | Carbon footprint and environmental impact of training | 28.1 |
| 2 | System-level energy optimisation (hardware and infrastructure) | 12.1 |
| 3 | Sustainability narratives, frameworks, and meta-research | 14.9 |
| 4 | Energy consumption and environmental methods | 8.8 |
| 5 | Inference and architecture optimisation for efficiency | 13.7 |
| 6 | Green algorithms and task specific learning | 11.2 |
Table 2: Green AI topic clusters from the LDA model and their prevalence. “Share of tokens” reports the fraction of all preprocessed abstract tokens assigned to each topic under the fitted model (topics relabelled for presentation; see Appendix Table 8 ).
Figure 3: Topic co-occurrence in the Green AI corpus. Cells report the frequency with which two topics appear within the same document (based on document-level topic mixtures from the LDA model), highlighting which themes tend to co-appear in abstracts.
Figure 4: Yearly evolution of Green AI themes. For each year, we count documents by their dominant LDA topic (highest topic share within the document) using the model, showing how the composition of the retained arXiv corpus shifts over time.
| Topic | Focus | Representative references |
| 1 | Carbon footprint and environmental impact of training | ( Strubell et al. 2019 ; Henderson et al. 2020 ; Schwartz et al. 2019 ) |
| 2 | System-level energy optimisation (hardware and infrastructure) | ( Wang et al. 2025 ; Bian et al. 2023 ; Kadian and others 2025 ) |
| 3 | Sustainability narratives, frameworks, and meta-research | ( Schwartz et al. 2019 ; Verdecchia et al. 2023 ; Singh et al. 2024 ) |
| 4 | Energy consumption and environmental methods (measurement and reporting) | ( Lacoste et al. 2019 ; Courty et al. 2023 ; Henderson et al. 2020 ) |
| 5 | Inference and architecture optimisation for efficiency | ( Jegham et al. 2025 ; Li et al. 2024 ; Kadian and others 2025 ) |
| 6 | Green algorithms and task specific learning | ( Schwartz et al. 2019 ; Zhong et al. 2024 ; Ferdous et al. 2025 ) |
Table 3: Green AI topic clusters: focus and representative references (illustrative, not exhaustive).
Figure 5: Conceptual mechanism: prompt design allocates discretion between researcher and system, shaping search scope, stopping, and which outputs are computed, which determines executed compute and therefore runtime and CO 2 e. Output equivalence is required to interpret footprint reductions as efficiency gains.
Figure 6: Measurement boundary of the benchmark. We measure downstream workflow compute in Colab (runtime and CodeCarbon CO 2 e) and exclude upstream LLM inference for code generation/filtering as well as background machine activity unrelated to the instrumented workflow.
Figure 7: Schematic workflow of the LDA pipeline and measurement scope. Candidate topic counts are evaluated by fitting LDA models and computing coherence diagnostics; the stopping rule and the choice of depend on the prompting strategy (e.g. exhaustive grid, coarse search, or decision-driven early stopping). Measurement starts at preprocessing and covers model search/selection and post-estimation computation (runtime and CodeCarbon CO 2 e estimates).
Figure 8: Generic prompt used to generate a baseline LDA pipeline with AI-assisted coding.
| Held constant in the experiment | May change in applied settings | Expected implication for footprint and interpretation |
| Task and workflow structure (LDA based survey pipeline) | Other workflows (regressions, simulations, machine learning, data cleaning, writing with code execution, agentic pipelines) | Absolute levels will differ, but the mechanism should strengthen when workflows involve wider search spaces and more optional outputs, since discretion has more scope to expand compute. |
| Execution environment (Google Colab free tier) | Local machines, dedicated servers, cloud instances with different hardware and background load | Baseline runtime and emissions may shift, and noise may change, but decision rules should still reduce redundant execution when compute is unconstrained. |
| Model used for coding (ChatGPT 5.2) | Other models and tools (different LLMs, code copilots, multi agent tools) | More capable or more agentic systems can increase variance in compute when unconstrained. Governance via explicit rules should remain relevant as capability rises. |
| Measurement approach (runtime and CodeCarbon CO 2 e) | Alternative carbon accounting assumptions, metered energy, different carbon intensity estimates | CO 2 e levels can vary with assumptions, which motivates reporting runtime alongside emissions. Relative differences across prompting strategies should be more stable than levels. |
Table 4: Experimental controls and external validity considerations.
Figure 9: Stylised efficiency frontier for prompting strategies. The efficient frontier represents the maximum attainable breadth of exercise (practical scope) for a given computational cost (runtime or CO 2 e). Weak guidance (green soft) may leave outcomes below the frontier with little change relative to the naive baseline. Constraint-based prompting reduces cost but can also reduce breadth of exercise when the prompt narrows what is computed. Decision-driven prompting can further lower cost by restricting computation to decision-relevant outputs. Points are illustrative and not estimated.
| Runtime | ||
| Strategy | Absolute savings (s) | Relative (%) |
| Green soft | 1.03 | 0.05 |
| Green computational constraints | 1011.07 | 44.89 |
| Decision-driven stewardship | 1251.39 | 63.26 |
| Carbon emissions | ||
| Strategy | Absolute savings (g CO 2 e) | Relative (%) |
Table 5: Average paired savings in runtime and estimated emissions relative to a within-block naive baseline ( paired runs per strategy).
Figure 10: Absolute runtime (seconds) by paired repetition. Each panel shows a blocked comparison between a freshly re-run naive baseline (Naive) and the corresponding prompting strategy (Strategy) for paired runs; lines connect within-pair outcomes.
Figure 11: Absolute emissions (g CO 2 e) by paired repetition. Each panel shows a blocked comparison between a freshly re-run naive baseline (Naive) and the corresponding prompting strategy (Strategy) for paired runs; lines connect within-pair outcomes.
| Comparison (vs. naive) | Average matched Jaccard |
| Green soft | 1.000 |
| Green computational constraints | 1.000 |
| Decision-driven (chosen ) | 1.000 |
Table 6: Sanity check of topic-output consistency relative to the naive baseline. Topics are compared at using the top 10 words per topic. Because topic labels are not identified across runs, topics are aligned using the Hungarian algorithm to maximise total similarity across topics. Similarity is measured by the Jaccard index over sets of top words. Under fixed seeds and identical specifications, is expected and is interpreted as a verification that the analytical pipeline is unchanged across prompting strategies.
| Productivity (research output) | Efficiency (computational footprint) | |
| Primary margin | Lower time and cognitive costs for producing, refactoring, and debugging code; faster iteration | Runtime, energy use, and CO 2 e of executed workflows, including optional search and post-estimation outputs |
| What GenAI changes | Reduces the marginal cost of iteration and enables more rapid prototyping of survey pipelines | Can widen the realised search scope and expand computed artefacts when discretion is not constrained |
| Main opportunity | Faster construction and updating of computational surveys under scale | Scope to reduce avoidable computation through explicit decision rules and stopping criteria |
| Main risk (if unmanaged) | “More iterations” may shift effort toward exploration and complexity without proportional analytical gains | “More automation” may translate into redundant execution, inflated runtime, and higher emissions |
| What our benchmark suggests | Prompting strategies preserve the ability to generate end-to-end pipelines | Generic green language has limited and unstable effects; operational constraints and decision-driven prompts can materially reduce runtime and estimated CO 2 e in this workflow |
| Role of human judgement | Shapes research direction, interpretation, and what is read in depth | Encodes search scope, stopping behaviour, and which outputs are decision-relevant (computational governance) |
Table 7: GenAI-assisted survey workflows: productivity gains and computational-efficiency risks.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 12: Yearly evolution of the retained Green AI corpus. Bars report the number of arXiv records classified as Green AI (titles and abstracts, after filtering) by publication year, illustrating the recent acceleration in relevant preprints.
| Internal ID | Presentation Topic | Matching cues (high-probability tokens) |
| 0 | 1 | model, training, carbon, footprint, performance |
| 1 | 2 | energy, system, network, efficiency, consumption |
| 2 | 3 | framework, research, paper, sustainable, ai |
| 3 | 4 | energy, consumption, environmental, method, efficiency |
| 4 | 5 | inference, deep, accuracy, computational, task |
| 5 | 6 | task, learning, efficiency, computational, performance |
Table 8: Mapping between internal LDA topic IDs (0–6) and presentation topics (1–7).
| Topic | Top tokens with probabilities |
| 1 | 0.024 model , 0.016 ai , 0.010 training , 0.010 carbon , 0.009 learning , 0.007 system , 0.007 performance , 0.007 footprint , 0.006 data , 0.006 computational |
| 2 | 0.015 energy , 0.011 system , 0.010 ai , 0.008 model , 0.007 performance , 0.007 network , 0.006 efficiency , 0.006 consumption , 0.006 data , 0.005 learning |
| 3 | 0.017 ai , 0.009 data , 0.008 energy , 0.006 system , 0.006 network , 0.006 framework , 0.005 research , 0.005 paper , 0.005 approach , 0.005 sustainable |
| 4 | 0.019 ai , 0.017 model , 0.013 energy , 0.009 data , 0.006 efficiency , 0.006 consumption , 0.005 learning , 0.005 environmental , 0.005 approach , 0.005 method |
| 5 | 0.018 model , 0.009 network , 0.009 inference , 0.008 deep , 0.007 learning , 0.007 accuracy , 0.007 computational , 0.007 task , 0.007 energy , 0.006 performance |
| 6 | 0.014 energy , 0.012 model , 0.007 efficiency , 0.007 learning , 0.007 task , 0.005 network , 0.005 computational , 0.005 performance , 0.005 consumption , 0.004 accuracy |
Table 9: Top tokens per topic and estimated probabilities in the final model. Topics are ordered according to the presentation mapping in Table 8 .
Figure 13: Topic co-occurrence network for the Green AI map. Nodes are topics and edge weights reflect document-level co-occurrence frequency (topics that jointly receive non-trivial weight within the same abstract), offering a graph view of the heatmap in Figure 3 .
Figure 14: Word clouds for the Green AI topics. Word size is proportional to the estimated topic–word probability from the LDA model fitted on preprocessed abstracts (lowercased, stopwords removed, lemmatised). Topic numbering follows the presentation mapping in Table 8 .
| Strategy vs. naive | ||||
| Green soft | 1.03 | 28.83 | 0.08 | 0.940 |
| Green computational constraints | 1011.07 | 8.11 | 278.85 | |
| Decision-driven stewardship | 1251.39 | 50.00 | 55.97 |
Table 10: Paired -tests relative to the naive baseline ( , ): runtime differences (seconds).
| Strategy vs. naive | ||||
| Green soft | 0.010 | 0.063 | 0.36 | 0.734 |
| Green computational constraints | 5.149 | 0.041 | 278.81 | |
| Decision-driven stewardship | 2.532 | 0.101 | 56.04 |
Table 11: Paired -tests relative to the naive baseline ( , ): emissions differences (grams CO 2 e).
| Strategy | Mean | Median | Boot CI95 (mean) | Boot CI95 (median) | Pos. | |
| Soft | 1.03 | 2.98 | [-23.83, 20.40] | [-44.99, 31.33] | 3/5 | 1.0000 |
| Comp. constraints | 1011.07 | 1011.16 | [1004.62, 1017.11] | [999.72, 1020.54] | 5/5 | 0.0625 |
| Decision-driven | 1251.39 | 1244.57 | [1214.45, 1292.22] | [1197.23, 1323.99] | 5/5 | 0.0625 |
Table 12: Robustness checks for paired runtime savings relative to the within-block naive baseline ( paired repetitions). We report mean and median paired savings, non-parametric bootstrap 95% confidence intervals (percentile, resamples), and an exact two-sided sign test based on the number of positive paired savings.
| Strategy | Mean | Median | Boot CI95 (mean) | Boot CI95 (median) | Pos. | |
| Soft | 0.010 | 0.007 | [-0.040, 0.055] | [-0.086, 0.071] | 3/5 | 1.0000 |
| Comp. constraints | 5.149 | 5.150 | [5.116, 5.180] | [5.091, 5.197] | 5/5 | 0.0625 |
| Decision-driven | 2.532 | 2.518 | [2.457, 2.615] | [2.423, 2.679] | 5/5 | 0.0625 |
Table 13: Robustness checks for paired emissions savings relative to the within-block naive baseline ( paired repetitions). We report mean and median paired savings, non-parametric bootstrap 95% confidence intervals (percentile, resamples), and an exact two-sided sign test based on the number of positive paired savings. Emissions are expressed in grams of CO 2 e.
More in this topic
AI-Assisted Scientific Research
AI-assisted scientific research uses machine learning and AI agents to support discovery, analysis, experimentation, and scientific communication.
See all 251 papers