The Hidden Costs of AI: A Review of Energy, E-Waste, and Inequality in Model Development

Jenis Winsta

REVIEW 3 major objections 5 minor 16 references

The Hidden Costs of AI: A Review of Energy, E-Waste, and Inequality in Model Development

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This review argues that AI's growth carries four hidden systemic costs — energy use, e-waste, compute inequality, and cybersecurity's power demand — that current practices ignore.

desk verdict A readable primer on AI's known sustainability costs, undone by a likely fabricated reference supporting its cybersecurity section. read the letter →

arxiv 2507.09611 v1 pith:4PUKFYEH submitted 2025-07-13 cs.AI cs.CY

Jenis Winsta This is my paper · ORCID

classification cs.AIcs.CY

keywords AIenergyconsumptionelectronicwastecomputedividecybersecurityoverheadGreenResponsiblecarbonemissionssustainablecomputing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This review argues that the environmental and social costs of AI — energy consumption, electronic waste, unequal access to computing, and the energy demanded by cybersecurity — are systemic consequences of how models are designed and deployed, not accidental side effects. It brings together published studies to show that training large models can emit hundreds of tons of carbon dioxide, that fast hardware turnover feeds a rising e-waste stream expected to reach 74.7 million metric tons by 2030, and that frontier development is concentrated in a handful of well-resourced regions. The author proposes that sustainability, transparency, and equity be treated as core metrics in AI evaluation, equal in weight to accuracy and speed, and that reporting practices must change to make these costs visible.

What carries the argument

Central to the argument is a lifecycle perspective that treats a model's footprint as spanning training, deployment, hardware retirement, and security operations. The review operationalises this perspective with four named mechanisms: the compute divide (the gap between institutions that can afford frontier-scale training and those that cannot), the Jevons paradox (efficiency gains spur more usage), the PUE blind spot (a data-center metric that ignores idle server waste), and the zero-trust overhead (encrypted and logged traffic raising data-center energy use by up to 30%).

What would settle it

An independent measurement campaign on a production data center that compares power draw under a perimeter-only security model and a zero-trust model would settle the 30% overhead claim. At the same time, a full lifecycle carbon audit of a frontier large language model that tracks training, fine-tuning, serving, and hardware retirement would test the review's central assertion that these costs are systemic and comparable in magnitude to one another.

Watch

Extended reading notes

Core claim

The paper's central claim is that AI's hidden costs form a connected system: the same design choices that push accuracy higher — larger models, denser hardware, and relentless experimentation — drive emissions, hardware churn, and compute concentration, while the security layers added to protect these systems consume additional power. The review synthesizes evidence from lifecycle assessments, e-waste monitors, infrastructure reports, and governance research to argue that Responsible AI must expand beyond model fairness to encompass the full physical footprint of development. It contends that the absence of standard reporting on compute hours and emissions lets these costs remain invisible, and that making them visible is a prerequisite for change.

Load-bearing premise

The load-bearing premise is that the published measurements the review cites are accurate; in particular, the 30% cybersecurity energy overhead rests on one source that the review does not make checkable, so that specific figure is the weakest anchor for the overall claim.

Editorial extensions

If this is right

If the review's synthesis is correct, emissions reporting for AI models should become as routine as accuracy reporting.
Treating e-waste as a systemic AI cost would shift procurement toward longer-lived hardware and repairable data-center designs.
Acknowledging the compute divide implies that open-weight models alone are insufficient; policy must fund distributed and public compute infrastructure.
Accounting for cybersecurity energy would make secure-by-design and green-by-default joint requirements for AI systems.
Redefining progress in AI to include sustainability could redirect research funding toward efficiency and edge inference.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An untested corollary of the lifecycle framing is that model-serving (inference) and fine-tuning may already dominate training energy in aggregate; a lifecycle audit that separates these phases would show where mitigation effort pays off most.
The compute-divide argument implies that safety and alignment research, which also requires frontier-scale compute, is similarly concentrated; democratizing compute would therefore also democratize who gets to shape AI governance.
Applying the Jevons-paradox logic to AI suggests that making models cheaper to run will increase total usage unless paired with absolute caps or pricing on carbon; the review implies but does not state this policy consequence.
The cybersecurity section's 30% overhead figure is a fragile quantitative anchor; measuring real zero-trust deployments could either strengthen or weaken the whole 'hidden burden' narrative.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The paper is a review of AI's environmental and social costs—energy, e-waste, compute inequality, and cybersecurity overhead. The first three themes are competent restatements of established work: Strubell et al., Patterson et al., Schwartz et al., Forti et al., Lehdonvirta et al. It is a readable survey if you need a primer.

The fourth theme is where it fails. Section V asserts that standard cybersecurity measures have 'notable energy overheads' and that AI-based security systems carry a 'dual energy burden,' citing [13]. That reference is likely fabricated: it claims Journal of Cleaner Production vol. 123 (2023), but JCLP's PII prefix is S0959-6526, not S0195-..., and volume 123 would date to roughly 2006. The DOI doesn't resolve. This is not a minor typo—it is the only support for the paper's distinct fourth pillar. There is also a separate misattribution: Section II attributes a '2021 report' from Stanford CRFM to [4], which is actually Brundage et al. 2020. So the citation record is unreliable.

The reader's take is slightly off on one detail: the 'up to 30%' energy overhead in Section V actually cites [7] (IEA EDNA), not [13]. But that does not rescue the section; [13] still carries the qualitative cybersecurity claims.

What is new? Very little. That is acceptable for a review, but a review's value rests on trustworthy sourcing. With a likely fabricated reference, the synthesis cannot be trusted as-is.

If the author verifies or replaces [13] and corrects the CRFM citation, the paper could serve as an accessible primer for policymakers. As it stands, it is not ready for publication. My recommendation: desk reject, with an invitation to resubmit after fixing the reference problems.

Referee Report

3 major / 5 minor

Summary. The manuscript is a literature review that argues AI development carries hidden environmental and social costs in four areas: energy consumption in model training and deployment, electronic waste from hardware turnover, inequality in compute access, and the energy burden of cybersecurity systems. It draws on well-known sources such as Strubell et al. (2019), Patterson et al. (2021), and Schwartz et al. (2020), as well as institutional reports, and concludes by advocating for Green AI principles, regulatory oversight, and equitable compute infrastructure. The paper is structured as a narrative review with sections devoted to each theme, a solutions and gaps section, and a conclusion.

Significance. If its sources were trustworthy, the paper would provide a useful synthesis of four interrelated concerns in Responsible AI: energy use, e-waste, compute inequality, and cybersecurity-related energy overhead. The first three themes are supported by credible and verifiable references, and the paper correctly highlights research gaps such as the lack of standardized emission reporting and the concentration of frontier AI development. However, the review's value as a synthesis is critically dependent on the accuracy of its bibliographic record, and the manuscript contains a citation that appears to be fabricated, as well as several misattributions and identifier errors. Because this is a review paper, citation integrity is load-bearing; the paper does not contribute new experiments or derivations, and its reliability rests entirely on the fidelity of its secondary sources.

major comments (3)

[Section V, reference [13]] Reference [13] is cited as the sole source for the claims that standard protection systems 'incur notable energy overheads' and that AI-based security systems introduce a 'dual energy burden' with energy-inefficient 24/7 IoT deployment. This reference, G. V. Smith, 'Comparative eco-efficiency assessment of cybersecurity solutions,' Journal of Cleaner Production, vol. 123, pp. 45-56, 2023, with ScienceDirect PII S0195925523000628, appears to be fabricated. The PII prefix does not match the journal's actual ISSN (0959-6526, PII prefix S0959-6526), and volume 123 of that journal would have been published around 2006, not 2023. Without a verifiable source, the fourth pillar of the review—the hidden energy burden of cybersecurity—is unsupported. Note that the 'up to 30%' figure in the same section is actually attributed to [7], the IEA EDNA report, not [13]; nevertheless, the qualitative cybersecurity-overhead claims rest on [13].
[Section II, reference [4]] The text states that the Stanford Center for Research on Foundation Models (CRFM) 'expanded this conversation by framing energy consumption as a systemic issue' in 'their 2021 report,' citing reference [4]. However, [4] is Brundage et al. 2020, 'Toward Trustworthy AI Development,' which is not a CRFM report and was not published in 2021. The following paragraph then credits Brundage et al. with a different claim about 'high-precision compute tracking,' using the same reference. This conflates two distinct works, and the actual CRFM 2021 report cited in the text does not appear in the bibliography. This is a serious citation error that undermines confidence in the accuracy of the related attributions.
[Section IV, reference [10]] The text claims that 'scholars such as Timnit Gebru and Abeba Birhane have framed this divide as a form of AI colonialism' and cites reference [10], which is Mohamed, Orife, Png, and Birhane, 'Decolonial AI.' Abeba Birhane is indeed an author of [10], but Timnit Gebru is not. If the authors intended to reference a separate work by Gebru on this topic, that work is not cited, and the sentence as written misattributes a view to a scholar who is not a source of the cited reference.

minor comments (5)

[Reference [8]] The arXiv identifier listed for Ahmed and Wahed is 'arXiv preprint arXiv:2009.10385,' but the URL in the same reference points to 'https://arxiv.org/abs/2010.15581,' which is the correct identifier for this paper. The printed identifier should be corrected.
[Reference [3]] The author list for the 'Green AI' paper is given as 'A. Schwartz, A. Dodge, N. Smith, and O. Etzioni,' but the correct names are Roy Schwartz, Jesse Dodge, Noah A. Smith, and Oren Etzioni. The initials in the reference are incorrect.
[Section V] The claim that zero-trust frameworks 'can raise energy use by up to 30%, depending on workload and cryptographic methods' is attributed to [7], the IEA EDNA report. The report should be checked to ensure that the 30% figure is accurately represented and not presented as a general result when it may apply only to specific test conditions.
[Figure 3] Figure 3 is labeled as an 'Illustrative comparison' and its source is given as 'Adapted from Lehdonvirta et al. [9], Ahmed & Wahed [8], and analysis by the author.' Since the figure includes an author-added analysis, the caption should explicitly state that the country categorization is illustrative and not based on a single quantitative metric, to avoid implying a rigor that the underlying sources do not provide.
[General] As a review article, the manuscript does not describe its literature search methodology, including databases, search terms, inclusion criteria, or a timeline. For a paper whose contribution is synthesis, a brief note on the review method would improve reproducibility and help readers judge the comprehensiveness of the four themes.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is a literature review whose claims are all attributed to external sources, with no internal derivation chain to reduce.

full rationale

This paper is a narrative literature review, not a derivation. It contains no equations, no fitted parameters, no predictive model, and no formal argument in which an output is constructed from its own inputs. Every substantive quantitative or qualitative claim—training emissions, e-waste statistics, compute divide characterizations, and cybersecurity energy overheads—is explicitly attributed to external references [1]–[15]. There is no self-citation: the author, Jenis Winsta, is an independent researcher and none of the cited works are authored by the author. The skeptical concern about reference [13] (a possibly fabricated citation with an invalid ScienceDirect PII) and the misattribution of a '2021 report' to [4] are serious bibliographic and correctness issues, but they are not circularity. An unsupported or even fabricated external source is an evidence-integrity problem, not a case where a claim reduces by definition to its own premise. Since the paper makes no internal prediction or fit, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The paper introduces no free parameters or invented entities. It rests on the assumption that its sources are reliable, which is questionable given the citation errors and a likely fabricated reference.

assumptions (1)

domain assumption The cited sources are accurate and exist as listed.
The review's quantitative claims, such as emissions figures and the 30% cybersecurity energy overhead, rest on the integrity of the cited references. This assumption is violated by reference [13], which appears fabricated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Hidden Costs of AI: A Review of Energy, E-Waste, and Inequality in Model Development." pith.science (2026). https://pith.science/paper/4PUKFYEH

@misc{pith2026250709611,
  author       = {Pith},
  title        = {Pith review of: The Hidden Costs of AI: A Review of Energy, E-Waste, and Inequality in Model Development},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4PUKFYEH}},
  note         = {Machine review of arXiv:2507.09611}
}

read the original abstract

Artificial intelligence (AI) has made remarkable progress in recent years, yet its rapid expansion brings overlooked environmental and ethical challenges. This review explores four critical areas where AI's impact extends beyond performance: energy consumption, electronic waste (e-waste), inequality in compute access, and the hidden energy burden of cybersecurity systems. Drawing from recent studies and institutional reports, the paper highlights systemic issues such as high emissions from model training, rising hardware turnover, global infrastructure disparities, and the energy demands of securing AI. By connecting these concerns, the review contributes to Responsible AI discourse by identifying key research gaps and advocating for sustainable, transparent, and equitable development practices. Ultimately, it argues that AI's progress must align with ethical responsibility and environmental stewardship to ensure a more inclusive and sustainable technological future.

Figures

Figures reproduced from arXiv: 2507.09611 by the authors.

**Figure 1.** CO2 emissions from selected AI models, illustrating the environmental cost of large-scale experimentation. (Adapted from [1], [2]) on model architecture, hardware, datacenter efficiency, and geographic location. Training GPT-3, for instance, required 1,287 megawatt-hours (MWh) of energy and emitted 552 tons of CO2e equivalent, an amount comparable to multiple transcontinental flights. They advocated for adopting spa… view at source ↗

**Figure 2.** Global e-waste generation from 2014 to projected levels [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Illustrative comparison of countries by access to ad [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages

[13]

Comparative eco-efficiency assessment of cybersecurity solutions,

G. V . Smith, “Comparative eco-efficiency assessment of cybersecurity solutions,” Journal of Cleaner Production , vol. 123, pp. 45–56, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/abs/ pii/S0195925523000628

work page 2023
[4]

Toward trustworthy ai development: Mechanisms for supporting verifiable claims,

M. Brundage, S. Avin, J. Wang, and et al., “Toward trustworthy ai development: Mechanisms for supporting verifiable claims,” 2020, arXiv preprint arXiv:2004.07213.3. [Online]. Available: https: //arxiv.org/abs/2004.07213

arXiv 2020
[7]

Energy efficiency metrics for data centres,

V . Maagøe, “Energy efficiency metrics for data centres,” October

work page
[10]

Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence

S. Mohamed, A. Orife, M.-T. Png, and A. Birhane, “Decolonial ai: Decolonial theory as sociotechnical foresight in artificial intelligence,” Philosophy and Technology , vol. 35, no. 4, 2022. [Online]. Available: https://arxiv.org/abs/2007.04068

work page Pith review arXiv 2022
[1]

Energy and policy considerations for deep learning in nlp,

E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for deep learning in nlp,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , 2019, pp. 3645–3650. [Online]. Available: https://aclanthology.org/P19-1355/

work page 2019
[2]

Carbon emissions and large neural network training,

D. Patterson, J. Gonzalez, Q. Le et al. , “Carbon emissions and large neural network training,” 2021. [Online]. Available: https://arxiv.org/abs/2104.10350

arXiv 2021
[3]

Green ai,

A. Schwartz, A. Dodge, N. Smith, and O. Etzioni, “Green ai,” Communications of the ACM , vol. 63, no. 12, pp. 54–63, 2020. [Online]. Available: https://doi.org/10.1145/3381831

doi:10.1145/3381831 2020
[5]

The global e- waste monitor 2020: Quantities, flows and the circular economy potential,

V . Forti, C. P. Bald ´e, R. Kuehr, and G. Bel, “The global e- waste monitor 2020: Quantities, flows and the circular economy potential,” 2020, united Nations University (UNU), International Telecommunication Union (ITU), and ISW A. [Online]. Available: https://ewastemonitor.info/gem-2020/

work page 2020

Show all 16 references

[6]

Clicking clean: Who is winning the race to build a green internet?

G. Cook, “Clicking clean: Who is winning the race to build a green internet?” 2017. [Online]. Available: https://www.greenpeace.de/ publikationen/20170110 greenpeace clicking clean.pdf

2017
[8]

The de-democratization of ai: Deep learning and the compute divide in artificial intelligence research,

N. Ahmed and M. Wahed, “The de-democratization of ai: Deep learning and the compute divide in artificial intelligence research,” 2020, arXiv preprint arXiv:2009.10385. [Online]. Available: https: //arxiv.org/abs/2010.15581

2020 arXiv
[9]

Compute north vs. compute south: The uneven possibilities of compute-based ai governance around the globe,

V . Lehdonvirta, B. Wu, and Z. Hawkins, “Compute north vs. compute south: The uneven possibilities of compute-based ai governance around the globe,” in Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES) , 2024. [Online]. Available: https://ojs.aaai.org/in...

2024
[11]

Decentralized, locally-tailored technology,

Distributed AI Research Institute, “Decentralized, locally-tailored technology,” 2024, dAIR Institute. [Online]. Available: https://www. dair-institute.org/blog/decentralized

2024
[12]

Creating trustworthy ai,

Mozilla Foundation, “Creating trustworthy ai,” 2020. [Online]. Available: https://assets.mofoprod.net/network/documents/ Mozilla-Trustworthy AI.pdf

2020
[14]

[Online]

Distributed AI Research Institute, “About,” 2024. [Online]. Available: https://www.dair-institute.org/about/

2024
[15]

The evolution of hugging face and its role in democratizing ai,

ODSC, “The evolution of hugging face and its role in democratizing ai,” 2024, medium. [Online]. Available: https://odsc.medium.com/ the-evolution-of-hugging-face-and-its-role-in-democratizing-ai-76f19af6d374

2024
[2022]

Available: https://www.iea-4e.org/wp-content/uploads/ 2022/10/EDNA-Studies-Metrics-for-data-centre-efficiency-Final.pdf

[Online]. Available: https://www.iea-4e.org/wp-content/uploads/ 2022/10/EDNA-Studies-Metrics-for-data-centre-efficiency-Final.pdf

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [13]

Comparative eco-efficiency assessment of cybersecurity solutions,

G. V . Smith, “Comparative eco-efficiency assessment of cybersecurity solutions,” Journal of Cleaner Production , vol. 123, pp. 45–56, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/abs/ pii/S0195925523000628

work page 2023

[2] [4]

Toward trustworthy ai development: Mechanisms for supporting verifiable claims,

M. Brundage, S. Avin, J. Wang, and et al., “Toward trustworthy ai development: Mechanisms for supporting verifiable claims,” 2020, arXiv preprint arXiv:2004.07213.3. [Online]. Available: https: //arxiv.org/abs/2004.07213

arXiv 2020

[3] [7]

Energy efficiency metrics for data centres,

V . Maagøe, “Energy efficiency metrics for data centres,” October

work page

[4] [10]

Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence

S. Mohamed, A. Orife, M.-T. Png, and A. Birhane, “Decolonial ai: Decolonial theory as sociotechnical foresight in artificial intelligence,” Philosophy and Technology , vol. 35, no. 4, 2022. [Online]. Available: https://arxiv.org/abs/2007.04068

work page Pith review arXiv 2022

[5] [1]

Energy and policy considerations for deep learning in nlp,

E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for deep learning in nlp,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , 2019, pp. 3645–3650. [Online]. Available: https://aclanthology.org/P19-1355/

work page 2019

[6] [2]

Carbon emissions and large neural network training,

D. Patterson, J. Gonzalez, Q. Le et al. , “Carbon emissions and large neural network training,” 2021. [Online]. Available: https://arxiv.org/abs/2104.10350

arXiv 2021

[7] [3]

Green ai,

A. Schwartz, A. Dodge, N. Smith, and O. Etzioni, “Green ai,” Communications of the ACM , vol. 63, no. 12, pp. 54–63, 2020. [Online]. Available: https://doi.org/10.1145/3381831

doi:10.1145/3381831 2020

[8] [5]

The global e- waste monitor 2020: Quantities, flows and the circular economy potential,

V . Forti, C. P. Bald ´e, R. Kuehr, and G. Bel, “The global e- waste monitor 2020: Quantities, flows and the circular economy potential,” 2020, united Nations University (UNU), International Telecommunication Union (ITU), and ISW A. [Online]. Available: https://ewastemonitor.info/gem-2020/

work page 2020