The inseparable link between EDI and AI's technical stack
When it comes to how AI is trained, bias, unjust outcomes, and a lack of representation can occur in the data used to create models. Recent work from UCL finds that when people interact with biased AI systems, they not only see biased outputs (for example, underestimating women’s performance), but they also become more biased themselves, creating a feedback loop between human and machine decision‑making. These are not just abstract issues; they are deeply ingrained in the technical infrastructure and procedures. RSE teams should therefore monitor not just model outputs but user behaviour post-deployment.
From requirements collecting to deployment, every phase of the software development lifecycle needs to be seen through an EDI lens. This involves defining representativeness criteria, demographic, regional, linguistic, etc., before testing begins, not after. When AI is incorporated into existing software, the algorithms often inherit or amplify systemic biases embedded in operational procedures or legacy code. Auditing the entire integrated system is necessary for true equity.
Any AI model's DNA is its training data. Racial, gender, socioeconomic, and regional biases are frequently overrepresented or underrepresented in datasets. For instance, models for image identification that were mostly trained on data from Western nations performed badly in other places. Given that the "dataset problem" is fundamentally a global supply chain and EDI issue, the human labour involved in labelling and cleaning data, typically low-wage workers in the Global South, must also abide by fair labour practices and ethical sourcing norms.
Environmental cost of AI/software scale and green computing
The exponential growth in AI model size and training complexity presents a critical challenge to sustainability. The energy demands of modern AI are staggering. Recent analysis by Epoch AI estimates that a typical ChatGPT query using GPT‑4o consumes on the order of 0.3 watt‑hours of electricity, a seemingly small cost that scales rapidly when multiplied across billions of daily queries, tying environmental responsibility (Green Computing) directly to ethical considerations (AI ethics).
RSE teams should treat efficiency as a first-class engineering requirement rather than an afterthought. In practice, this means: selecting the smallest model that meets accuracy requirements; applying compression, quantisation, or distillation before scaling up; and benchmarking energy consumption alongside performance at each development stage. The question "Do we need this model size?" should be asked explicitly at the design stage rather than retrospectively.
Infrastructure decisions carry equal weight. Teams should locate training workloads in regions served by verified renewable energy where possible, and treat carbon cost as a line item in project planning. Where renewable infrastructure is unavailable, carbon offsetting should be budgeted from the outset.
Hardware procurement is the third lever. The GPUs and TPUs required for AI training contribute significantly to e-waste, and the rare-earth minerals they depend on, cobalt and lithium, are frequently sourced under harsh labour and environmental conditions. RSE teams should extend hardware lifetimes, audit procurement chains for ethical sourcing standards, and require providers to provide transparent recycling documentation before purchase.
Green computing as an EDI strategy
The least obvious of the three relationships is perhaps the most important: leaner systems are more equitable systems.
Technical (green) choices about performance, energy use, and resource consumption (efficiency) directly shape who can participate and who benefits (equity). For example, if an AI system is so compute‑hungry that only well-resourced institutions with big clusters or financial power can run it, then people in less‑resourced universities or in the Global South are effectively excluded.
Making a system leaner is therefore also making it more equitable. Reduced energy consumption lowers cloud costs and broadens access. RSE teams should favour infrastructure providers with verified renewable energy commitments, and redirect budget freed from over-engineered systems toward accessibility tooling, training, and inclusion initiatives.
Applying the responsible stack in practice: the responsible stack
The three relationships explored above are not sequential steps; they are simultaneous lenses on every decision. Consider a research team choosing a cloud provider for a training run. Through a Green Computing lens: which regions offer verified renewable energy? Through an EDI lens: will the resulting mode/softwarel run on hardware available to a typical under-resourced institution? Through an AI ethics lens: is the chosen model size actually necessary, or is the team defaulting to a larger model because the cost is grant-funded and the environmental consequence is invisible?
A few practical principles follow from holding all three lenses at once:
- Ask "Do we need this model size?" explicitly at the design stage, and document the answer.
- Include projected carbon cost in grant applications as a named line item, not a footnote.
- Define representativeness criteria for training data before collection begins, not during evaluation.
- Report energy consumption in research outputs alongside standard computational metrics.
- Publish model weights, training scripts, and evaluation data in a FAIR-compliant repository, so that others can build on the work without repeating its carbon cost.
References
Josh You (2025), "How much energy does ChatGPT use?". Published online at epoch.ai. Retrieved from https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
UCL (2024). Bias in AI amplifies our own biases. [online] UCL News. Available at: https://www.ucl.ac.uk/news/2024/dec/bias-ai-amplifies-our-own-biases
Ahmad, A., Vallès, Y., & Idaghdour, Y. (2026). Bias in AI systems: integrating formal and socio-technical approaches. Frontiers in big data, 8, 1686452. https://doi.org/10.3389/fdata.2025.1686452
Ananthram, A., Stengel-Eskin, E., Vondrick, C., Bansal, M. and McKeown, K. (2024). See It from My Perspective: Diagnosing the Western Cultural Bias of Large Vision-Language Models in Image Understanding. arXiv (Cornell University). [online] doi:https://doi.org/10.48550/arxiv.2406.11665.
Du, M. and Okolo, C. T. (2025) ‘Reimagining the future of data and AI labor in the Global South’, Brookings Institution. Available at: https://www.brookings.edu/articles/reimagining-the-future-of-data-and-ai-labor-in-the-global-south/
Chandhiramowuli, S. and Anwar, M. A. (2024) ‘Who makes “AI”? Uncovering the workers behind AI in the global south’, School of Social and Political Science, University of Edinburgh, Edinburgh, 5 April. Available at: https://www.sps.ed.ac.uk/news-events/event/who-makes-ai-uncovering-workers-behind-ai-global-south
University of Cambridge (?) ‘Energy-efficient supercomputing helps drive breakthrough research’. Research Computing Services. Available at: https://www.hpc.cam.ac.uk/energy-efficient-supercomputing
Wilding, V. (?) ‘Green gold rush: the ethical landscape of rare earth mining’, Green Economy. Available at: https://www.greeneconomy.co.uk/news-and-resources/insights/green-gold-rush-the-ethical-landscape-of-rare-earth-mining/
Acknowledgements
This guide was written by Michael Umeokoli, and reviewed by Jyoti Bhogal.