Position: Stop Preaching and Start Practising Data Frugality for Responsible Development of AI

Sophia N. Wilson, Andrew Millard, Guðrún Fjóla Guðmundsdóttir, Raghavendra Selvan, Sebastian MairICML 2026 · OpenReview

Abstract

This position paper argues that the machine learning community must move from preaching to practising data frugality for responsible artificial intelligence (AI) development. For long, progress has been equated with ever-larger datasets, driving remarkable advances but now yielding increasingly diminishing performance gains alongside rising energy use and carbon emissions. While awareness of data frugal approaches has grown, their adoption has remained rhetorical, and data scaling continues to dominate development practice. We argue that this gap between preach and practice must be closed, as continued data scaling entails substantial and under-accounted environmental impacts. To ground our position, we provide indicative estimates of the energy use and carbon emissions associated with the downstream use of ImageNet-1K. We then present empirical evidence that data frugality is both practical and beneficial, demonstrating that coreset-based subset selection can substantially reduce training energy consumption with little loss in accuracy, while also mitigating dataset bias. Finally, we outline actionable recommendations for moving data frugality from rhetorical preach to concrete practice for responsible development of AI.

Introduction

In the current AI development paradigm, progress is frequently equated with scaling—using ever-larger datasets to brute-force performance gains. However, this "hyper-datafication" yields diminishing returns while driving up energy consumption and carbon emissions. While the machine learning (ML) community is increasingly aware of these environmental costs, the response has largely been rhetorical. This paper addresses the "value-action gap" between preaching sustainability and the actual daily practices of AI researchers, proving that data frugality is not just an ethical ideal, but a highly practical technical approach.

The Environmental Toll of Data Scaling

To quantify the hidden costs of standard benchmarks, we estimated the downstream usage of ImageNet-1K. By analyzing ICLR papers (2017–2022) and extrapolating to the broader literature, we estimate that at least 46,179 models have been trained from scratch on ImageNet-1K since 2017. This equates to approximately 5.46 GWh of energy and 2,429 tonnes of CO2e—roughly the annual carbon footprint of over 500 people. Furthermore, the redundant local storage of this dataset alone costs an additional ~360 MWh/year (~160 tCO2e).

Empirical Evidence for Data Frugality

Data frugality is often perceived as a trade-off against performance, but empirical evidence suggests otherwise. By employing state-of-the-art coreset selection methods such as Dynamic Uncertainty (Dyn-Unc) and InfoMax, researchers can identify and remove redundant or uninformative training samples without compromising model quality.

Accuracy vs. Pruning Ratio for Swin-T and ResNet-34 architectures, demonstrating that significant data reduction can be achieved with minimal impact on performance.

As shown in Figure 3, pruning 25% of ImageNet-1K resulted in zero drop in Top-1 accuracy for architectures like Swin-T and ResNet-34. Crucially, these reductions translate to real-world efficiency: training a ResNet-34 on a 25% pruned subset reduced actual wall-clock training time by 32% and energy consumption by 29% per epoch. This highlights that compute efficiency does not automatically guarantee energy efficiency; time and energy must be explicitly measured and reported.

Beyond Efficiency: Mitigating Dataset Bias

Beyond saving energy, carefully curating a smaller dataset can actively remove systemic biases. By intentionally balancing a coreset, models can be prevented from learning spurious correlations that are often present in massive, uncurated datasets.

Accuracy vs. Training Budget on biased Colour-MNIST, showing how smart sampling (Balanced and Reweighted) outperforms standard baselines.

Using a biased toy dataset (Colour-MNIST) where digits are 99% correlated with specific colors, we demonstrated that algorithmic data sampling can correct distributional flaws. As seen in Figure 6, reweighted and balanced sampling strategies allow models to significantly outperform baseline models on biased data, approaching the performance ceiling of perfectly aligned datasets even with limited training budgets.

Recommendations and Future Work

To bridge the gap between preaching and practice, we propose the following actionable steps:

  1. Shift Metrics: Move from aggregate performance metrics to "capability-per-resource" or "Data-Pareto" reporting.
  2. Mandatory Reporting: Conferences should make resource and energy reporting mandatory to incentivize efficient development.
  3. Shared Infrastructure: Institutions should fund shared dataset storage to prevent redundant downloads and enforce "data sunset laws" to clear out unused, dark data.

While data pruning works exceptionally well for discriminative tasks, future work should explore its impact on generative models, which may rely more heavily on long-tail variations. Ultimately, the community must move toward a culture where data is no longer treated as a free, infinite input, but as a resource to be used with care and frugality.

Summary updated

Abstract

This position paper argues that the machine learning community must move from preaching to practising data frugality for responsible artificial intelligence (AI) development. For too long, progress has been equated with ever-larger datasets, driving remarkable advances but now yielding increasingly diminishing performance gains alongside rising energy use and carbon emissions. While awareness of data frugal approaches has grown, their adoption has remained rhetorical, and data scaling continues to dominate development practice. We argue that this gap between preach and practice must be closed, as continued data scaling entails substantial and under-accounted environmental impacts. To ground our position, we provide indicative estimates of the energy use and carbon emissions associated with the downstream use of ImageNet-1K. We then present empirical evidence that data frugality is both practical and beneficial, demonstrating that subset selection methods can substantially reduce training energy consumption with little loss in accuracy, while also mitigating dataset bias. Finally, we outline actionable recommendations for moving data frugality from rhetorical preaching to concrete practice for responsible development of AI.
BibTeX
@inproceedings{wilson2026position,
  author = {Sophia N. Wilson and Andrew Millard and Guðrún Fjóla Guðmundsdóttir and Raghavendra Selvan and Sebastian Mair},
  title = {{Position: Stop Preaching and Start Practising Data Frugality for Responsible Development of AI}},
  booktitle = {International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=Ybp1OTiNaH},
}
Figure