Truace
Subscribe
70/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

A review of 171 studies finds smaller datasets and simpler models cut AI energy use

Study title: CRISP-ML(QS): Embedding Sustainability in AI-ML Development Lifecycle

Evidence type
Systematic review or meta-analysis
Sample
n = 171
Sources
1 sourcePeer-reviewed
Published
3 min read
Truvace record
TRV-2026-2542cite or check it

Study title: CRISP-ML(QS): Embedding Sustainability in AI-ML Development Lifecycle

Researchers at the Indian Institute of Management Lucknow reviewed 171 papers on the environmental sustainability of AI and sorted the findings by stage of the machine-learning development lifecycle. They propose CRISP-ML(QS), an extension of an existing quality-assurance process that adds an energy and carbon check at each stage, from data preparation to deployment and monitoring.

Read the study abstract

Artificial Intelligence (AI) in contemporary industry and society is omnipresent, with its applications permeating across enterprise functions such as manufacturing, supply chain and research & development, and enterprise goals such as efficiency, innovation, and sustainability. However, an important question that has been raised, quite often recently, is: Is AI environmentally sustainable? This study proposes an extension to the CRoss-Industry Standard Process for Machine Learning with Quality Assurance (CRISP-…

The gain

In a review of 171 papers, studies it cites report that smaller training datasets cut machine-learning energy use by as much as 92% and simpler models cut energy and emissions by 23% to 99% at similar accuracy.

In brief

The reviewed studies report large savings from ordinary engineering choices. One found energy gains of up to 92% from shrinking the training dataset; another found that picking a simpler model over a complex one can cut energy use and emissions by 23% to 99% while keeping similar accuracy. The authors say the main obstacle is a gap between green-AI research and what industry actually adopts.

The savings are not uniform. In the same review, trimming features cut Random Forest energy by up to 74.81% with a 0.06% drop in F1-score, but the same step in KNN saved only 0.92% of energy while cutting F1-score by 98.05%. The framework covers only the environmental dimension and software-level choices, not hardware.

Main points

  1. The review screened 2,124 database records plus 1,397 found through citation chaining and included 171 papers (102 from databases, 69 from citation chaining).
  2. Reducing features in a Random Forest cut energy use by up to 74.81% with only a 0.06% drop in F1-score, according to a cited study; in KNN the same step saved 0.92% of energy and cost 98.05% of F1-score.
  3. For speech recognition, a cited study found a state-of-the-art transformer emitted half its total training CO2 just to gain a 0.3 reduction in word error rate.
  4. Where a cloud job runs matters: carbon-free energy share and emissions can differ by about 5 to 10 times even within the same country and organization, per a cited study.
  5. Pruning, compression and efficient architecture search appear throughout: a cited energy-aware pruning method cut AlexNet energy by 3.7 times with under 1% top-5 accuracy loss.
  6. A cited analysis of Meta's AI infrastructure put power capacity across experimentation, training and inference at roughly 10:20:70, so inference draws the most.
  7. Software-based energy estimates differ from hardware power measurements by an average of 20%, which can lead to wrong optimization decisions.

The full story

The authors start from the scale of the problem. They cite work saying training one AI model can emit five times the CO2 of an average US car over its lifetime, and that training GPT-3 (175 billion parameters) used 10,000 GPUs and emitted 552 metric tons of carbon dioxide. They argue regulation and practical frameworks lag behind these environmental needs.

To build the framework, they searched Scopus, ScienceDirect and Google Scholar for papers with titles on green or sustainable AI, with studies available up to 3 February 2026. Two reviewers screened independently, and the included studies were mapped to six stages: business and data understanding, data preparation, model building, evaluation, deployment, and monitoring and maintenance.

At the model-building stage the review lists many levers. Python implementations used more energy than C in one study, and TensorFlow and PyTorch differed in training and inference efficiency. Switching the optimizer from SGD with momentum to LARS roughly halved the training workload in another. A cited hyperparameter method was more than 3 times as energy-efficient as Bayesian optimization and random search.

For evaluation, the authors say models should be judged on energy as well as accuracy, and they catalogue metrics such as Accuracy Per Consumption and the Green AI quotient, plus tools like CodeCarbon-style trackers, Carbon Tracker and Green Algorithms. At deployment, a cited study found running inference near data sources with fog computing reduced latency and power use compared with remote cloud data centers.

Sources

1 on file
  1. International Journal of Mathematical, Engineering and Management Sciencesdoi:10.33889/ijmems.2026.11.5.083
    Peer-reviewed2026-09-29

Talk it over

Four questions to start onWho gets this benefit, and who was left out of the study or story?

Start with a question

A review of 171 studies finds smaller datasets and simpler models cut AI energy usehttps://truvace.com/item/crisp-ml-qs-embedding-sustainability-in-ai-ml-development-lifecycle

  1. Who gets this benefit, and who was left out of the study or story?

    The gain In a review of 171 papers, studies it cites report that smaller training datasets cut machine-learning energy use by as much as 92% and simpler models cut energy and emissions by 23% to 99% at similar accuracy.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a systematic review or meta-analysis with a sample of 171, from International Journal of Mathematical, Engineering and Management Sciences.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat The framework covers only the environmental dimension and software-level choices, and the savings come from separate studies rather than one controlled test; effects differ by algorithm.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Climate