Introduction
Machine learning has spent a decade being told to account for its carbon, and a decade not doing so.
Calls for transparency (Strubell et al., 2019; Schwartz et al., 2020; Henderson et al., 2020) have been met with instruments: model cards (Mitchell et al., 2019), datasheets (Gebru et al., 2021), repository-level emissions fields, and most recently sustainability-oriented model cards (Jouneaux & Cabot, 2025), carbon-aware evaluation metrics (Santosh et al., 2026), and proposals for mandatory venue disclosure (Halgamuge et al., 2026). And yet where reporting has been measured, it is sparse and may even have stalled (Castaño et al., 2023; Luccioni & Hernández-García, 2023). The problem this project explores therefore not "how much does AI emit?" but searching question of why, given so many instruments, does environmental disclosure in AI keep failing? And what would actually make it stick?
The project will explore, therefore, a hypothesised "footprint gap" as a sociotechnical and epistemic failure, as a product of how machine learning constructs what counts as responsible disclosure, who is treated as accountable, and how transparency norms emerge, circulate, and quietly die. On this account the proliferation of competing reporting instruments is itself a symptom, not a solution. The project will explore why disclosure has not taken hold and what a credible regime of environmental accountability for AI would require.
This reframing is what makes the study original. Existing studies have audited reporting on a single hub (Castaño et al., 2023, on Hugging Face) or surveyed emissions across model samples (Luccioni & Hernández-García, 2023), and several studies have proposed new disclosure standards (Jouneaux & Cabot, 2025; Santosh et al., 2026; Halgamuge et al., 2026). This thesis, by contrast, treats those very instruments as its objects of study - asking why they under-deliver - and brings critical data studies, science-and-technology studies, and critical infrastructure studies to bear on a problem the technical literature has approached principally through the lens of engineering.
Methodology, Challenges and Expected Outcomes
The evidence base will be assembled by a purpose-built technical pipeline. The candidate will harvest ML research outputs and documentation at scale from sources such as arXiv, the ACL Anthology, OpenReview, OpenAlex, the Hugging Face Hub API, and linked GitHub repositories, then parse heterogeneous formats - PDFs (via tools such as GROBID), Markdown model cards, YAML metadata, and README prose - into a normalised representation. The core task is information extraction: detecting whether energy or carbon is reported at all, locating it, and harmonising inconsistent quantities (kWh, CO₂e, GPU-hours, hardware, PUE, energy mix, instrumentation such as CodeCarbon) into a common schema, combining rule-based detection, fine-tuned classifiers/NER, and LLM-assisted structured extraction validated against a human-annotated gold set built with active learning. Anticipated technical challenges include: extracting a sparse, low-base-rate signal from noisy, semi-structured text at scale; distinguishing genuine measurement from boilerplate mentions; constraining LLM hallucination to produce auditable, reproducible extractions; and evaluating extraction quality in the absence of ground truth. The pipeline must also account for its own footprint, providing chance to apply efficient, right-sized methods reflexively.
In addition to the evidence base and its implementation, the anticipated contributions to knowledge are threefold.
A longitudinal account of environmental reporting practice that reaches beyond model hubs to the published literature and community norms, mapping not only what is reported but how disclosure instruments are taken up, adapted, ignored, or abandoned.
A theoretical framework for AI environmental accountability that explains the recurring failure of disclosure as a knowledge-and-governance phenomenon, rather than cataloguing it.
A critically-grounded account of what credible accountability would actually demand - institutional, epistemic, and infrastructural - against which any standard (existing or future) can be assessed.
References:
References
Strubell, E., Ganesh, A. & McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. *Proc. ACL 2019*. https://doi.org/10.18653/v1/P19-1355
Schwartz, R., Dodge, J., Smith, N.A. & Etzioni, O. (2020). Green AI. *Communications of the ACM* 63(12). https://doi.org/10.1145/3381831
Henderson, P. et al. (2020). Towards the Systematic Reporting of the Energy and Carbon Footprints of Machine Learning. *Journal of Machine Learning Research* 21(248). https://jmlr.org/papers/v21/20-312.html
Mitchell, M. et al. (2019). Model Cards for Model Reporting. *Proc. FAT\* 2019*. https://doi.org/10.1145/3287560.3287596
Gebru, T. et al. (2021). Datasheets for Datasets. *Communications of the ACM* 64(12), 86–92. https://doi.org/10.1145/3458723
Luccioni, S., Jernite, Y. & Strubell, E. (2024). Power Hungry Processing: Watts Driving the Cost of AI Deployment? *Proc. ACM FAccT 2024*. https://doi.org/10.1145/3630106.3658542
Luccioni, A.S. & Hernández-García, A. (2023). Counting Carbon: A Survey of Factors Influencing the Emissions of Machine Learning. arXiv:2302.08476. https://arxiv.org/abs/2302.08476
Castaño, J., Martínez-Fernández, S., Franch, X. & Bogner, J. (2023). Exploring the Carbon Footprint of Hugging Face's ML Models: A Repository Mining Study. *Proc. ACM/IEEE ESEM 2023*. https://arxiv.org/abs/2305.11164
Jouneaux, G. & Cabot, J. (2025). Towards Sustainability Model Cards. arXiv:2507.19559. https://arxiv.org/abs/2507.19559
Santosh, K. et al. (2026). AI-CARE: Carbon-Aware Reporting Evaluation Metric for AI Models. arXiv. *(carbon-aware reporting metric; treated here as object of study)*
Halgamuge, M. et al. (2026). Emissions transparency in AI research should be mandatory. *Sustainable Computing: Informatics and Systems*. *(mandatory-disclosure policy framework; treated here as object of study)*
Digital Humanities Climate Coalition (2024). The Cloud and the Climate: Navigating AI-Powered Futures. https://doi.org/10.5281/zenodo.13850067 — and (2022) Researcher Guide / Toolkit: https://doi.org/10.5281/zenodo.6451499 ; https://sas-dhrh.github.io/dhcc-toolkit/