Digital Overconsumption and Waste: A Closer Look at the Impacts of Generative AI

Steve DiPaola; Vanessa Utz

REVIEW 3 major objections 6 minor 10 references

Digital Overconsumption and Waste: A Closer Look at the Impacts of Generative AI

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Generative AI systems currently add to digital waste because most of the images they produce are never used for a practical purpose, the paper argues.

desk verdict Plausible framing piece on generative AI and digital overconsumption, but its key statistics are self-cited from unpublished work, so the central claim is not independently checkable as written. read the letter →

arxiv 2505.18894 v1 pith:DHMZ7OXN submitted 2025-05-24 cs.AI

Vanessa Utz , Steve DiPaola This is my paper · ORCID

classification cs.AI

keywords generativeAIdigitalwasteoverconsumptionCO2emissionsenergyconsumptionuserbehaviorcommon-poolresourcedilemma

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

Generative AI image tools have grown so quickly that their daily output is estimated in the tens of millions, and this paper argues that a large share of that output is a new form of digital waste. The authors' prior survey data show users generating more than 1,500 images per week on average, with roughly 40% using the tools purely for entertainment and fewer than 15% calling themselves professional users. Because most generated images are never looked at again, and because the electricity and CO2 costs of generation are real, the paper characterizes the situation as digital overconsumption. It frames this as a common-pool resource dilemma: users consume a shared environmental resource without knowing its size or their personal share of it. The paper's proposed response is education and an urgent public discussion, together with more research into the psychological pull of these tools.

What carries the argument

The argument is carried by two linked concepts. The first is digital waste, defined as the emissions, waste, natural-resource extraction, and other environmental harms tied to the creation, use, and maintenance of data infrastructures; this gives the paper a category for treating generated images that are never used again as environmental harm. The second is the common-pool resource dilemma, a framework in which a shared resource has an unknown total size, access to it is unequal, and individuals must decide how much to consume without seeing the full resource costs; this turns individual generative-AI use into a collective overconsumption problem. A third, supporting mechanism is uses-and-gratifications theory, which the paper invokes to explain why users keep generating in large quantities.

What would settle it

An independent usage study that logs a representative sample of generative-AI sessions, records whether each generated image is ever reopened or used, and meters the actual energy draw per image would settle the central claim. If real per-user weekly generation turned out to be far lower than 1,500, or if most images were made for concrete utilitarian tasks, the digital-overconsumption conclusion would lose its empirical support.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that generative AI systems currently contribute negatively to the production of digital waste through the energy they consume and the CO2 they emit, and that a significant portion of the generated images are not created or used for a direct utilitarian purpose. The supporting evidence offered is an energy estimate of 1.92–9.29 TWh per year for these systems, an average of over 1,500 generated images per user per week, and survey findings that most users had not considered the energy implications, that a large share are skeptical of climate data, and that less than 15% of users are professionals. From this, the authors conclude that the mass adoption of generative AI replicates harmful overconsumption in the digital space and that users are making consumption decisions without a full picture of the resources involved.

Load-bearing premise

The argument rests on the authors' own survey statistics — an average of over 1,500 generated images per user per week and roughly 40% of users generating purely for entertainment — being representative of the entire generative-AI user base.

Editorial extensions

If this is right

If most generated images are never used again, then a substantial portion of generative AI's electricity use and emissions is discretionary, so reducing non-utilitarian generation becomes a direct mitigation lever.
Because many users are skeptical of climate data or pessimistic about others' behavior, simply publishing energy figures is unlikely to change user behavior; the paper points to education campaigns and potential legislative intervention.
Framing overconsumption as a common-pool resource dilemma means individual decisions add up to a collective environmental cost, so solutions may require shared norms, transparency, or governance rather than individual choice alone.
As generative systems move from still images toward video and virtual reality, the same pattern would likely scale up in energy use and in the escapism-related risks the paper flags, making early intervention more valuable.

Reading between the lines

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

Extending the paper's logic, the same common-pool framing applies to other generative AI workloads such as text generation and video synthesis, whose mass adoption is still young enough that the waste patterns described here might be preempted.
The reported average of 1,500 images per user per week may be dominated by a small number of heavy users; if true, interventions targeting the heaviest users could reduce waste far more cheaply than broad education campaigns.
The paper treats 'never looked at again' as the waste signal, but a sharper test would compare the marginal energy cost of each image with the value the user actually gets from it; that comparison could distinguish waste from legitimate creative exploration.
If users already distrust researcher-provided climate data, an interface-level intervention, showing an estimated CO2 cost per image at the moment of generation, might be a more direct test of whether transparency changes behavior.

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 short version: this is a workshop position paper, not a research preprint. It makes a plausible argument that generative AI enables digital overconsumption and waste, and it applies two existing frameworks (uses and gratifications theory, common pool resource dilemma) to that problem. It does not pretend to offer new measurements, and it is honest about the uncertainty in its own energy estimates (1.92–9.29 TWh range). That is a plus.

What it does well: it names a real issue—users generating large volumes of images without thinking about the climate cost—and it connects the dots to prior work on addictive social media and data waste. The proposal to treat overconsumption as a common pool resource dilemma is a useful framing. The paper also explicitly flags that users are skeptical or pessimistic about changing behavior, which is a fair and non-hysterical observation.

Where it is soft: the load-bearing numbers come from the authors’ own prior work. The 1500+ images per user per week, the 40% entertainment-only use, and the 50+ iterations are quoted from [3] and [4], with no survey instrument, sample size, recruitment method, or error bars. Worse, [4] is described as still under review, so I cannot check it. That means the central overconsumption claim is not independently verifiable from this text. Additionally, the leap from “only 15% are professional users” to “most output is non-utilitarian” is too quick. Casual creative exploration can be personally valuable, and the paper gives no evidence about how many images are truly never looked at again.

Still, the paper is honest about its scope. It says it is outlining previous work and sparking discussion, not settling the empirical question. The logic is coherent and the tone is measured. If the authors release the underlying data, the argument would be much stronger; as it stands, it is a hypothesis in need of support.

I would send this to a workshop venue, not a top journal, and with the explicit instruction that the authors make their data and methods available before publication. It deserves referee time because the topic matters and the framing is useful—but the referee should push hard on transparency.

Best, [You]

Referee Report

3 major / 6 minor

Summary. This workshop position paper argues that the mass adoption of generative AI systems such as Midjourney and Stable Diffusion contributes to digital overconsumption and waste through large volumes of low-utility image generation and associated energy use and CO2 emissions. It summarizes the authors' previous estimates of annual global energy consumption (1.92–9.29 TWh) and survey findings (over 1500 generated images per user per week, roughly 40% entertainment-only use, over 50 iterations per satisfying result, and fewer than 15% professional users), then frames overconsumption as a common-pool resource dilemma and proposes education and awareness as a solution pathway. The paper is primarily a synthesis of prior work rather than a new empirical study, and it closes with suggestions for future research on cognitive reward systems and the societal implications of generative AI in immersive environments.

Significance. The paper raises a timely and important issue: the environmental cost of low-utility AI-generated content, and it connects this to established frameworks such as digital waste, common-pool resource dilemmas, and uses-and-gratifications theory. The main conceptual contribution is reframing generative AI's environmental impact in terms of overconsumption and waste, not just raw energy numbers, and the proposed educational solution pathway is plausible and worth discussing. However, the empirical foundation for the central overconsumption claim consists of self-cited, partially unpublished survey statistics with no reported methodology, sample size, or confidence intervals, so the strength of the evidence is currently not assessable from the paper alone. The authors also make unsupported empirical assertions about 'never looked at again' images. If the underlying data can be made available or independently replicated, the paper's core message would be considerably strengthened; as it stands, the paper is better read as a position statement than as a definitive empirical account.

major comments (3)

[Section 2] The paper's central empirical claim—that most AI-generated images are non-utilitarian waste—rests on statistics from the authors' own prior work ([3], [4]) without any methodological details. Specifically, the 1500+ images per user per week, the 40% entertainment-only figure, the 50+ iterations, and the less-than-15% professional-user figure are presented without sample size, survey instrument, sampling frame, response rate, or confidence intervals, and reference [4] is an unpublished manuscript that is not accessible to readers. This makes the key numbers unverifiable from the current text. The authors should either include the full survey methodology and summary statistics in the paper or in an accessible appendix, or explicitly label these as preliminary and self-reported findings from an unreviewed study.
[Section 2] The sentence 'Most images that are created using these tools are currently never looked at again after the initial creation' is an unsupported empirical claim. No evidence or citation is provided, and it does not follow from the preceding survey results about iterations and user categories. The authors should either provide data to support this statement or remove it (or qualify it as a conjecture).
[Section 2] The inference from 'less than 15% professional users' and 'around 40% use solely for entertainment' to 'a significant portion of the generated images are not created and used for a direct utilitarian purpose' conflates non-professional and entertainment use with non-utilitarian use. Entertainment, hobbyist exploration, and creative play can have personal or creative value to the user, and the paper provides no definition or measure of 'direct utilitarian purpose' or of whether generated images are ever re-viewed or used. The claim therefore does not logically follow from the cited statistics and should be reworded or supported with additional evidence.

minor comments (6)

[Section 2] There is a typo: 'over 1500 generated images per users' should be 'per user'.
[Section 1] The chemical formula should be 'CO2' rather than 'C02' in the sentence about emissions from ML training.
[Section 4] In the sentence about the common-pool source dilemma, 'access to the resources in not equally distributed' should be 'access to the resources is not equally distributed'.
[References] Reference [4] is listed merely as 'Vanessa Utz and Steve DiPaola. Manuscript' with no title, venue, or status; this is insufficient for a citation and should be completed or the reference should be removed.
[Section 4] The phrase 'There is no unilateral solution' could be misinterpreted; consider using 'There is no single solution' or 'There is no one-size-fits-all solution'.
[General] The paper would benefit from a short limitations paragraph acknowledging that the cited survey data are self-reported and preliminary, and that the energy estimates have a wide range.

Circularity Check

1 steps flagged · score 5.0 of 10

Overconsumption claim rests on self-cited, unpublished survey statistics; the paper has independent discussion content but its load-bearing empirical premise is not independently verifiable.

self citation load bearing [Section 2, 'Digital Overconsumption and Generative AI', paragraphs 1-2]
"Our data showed an average weekly number of over 1500 generated images per users [4]. Additionally, a brief survey that we conducted as part of our initial work on energy consumption of generative AI [3], indicated that around 40% of respondents use the tools solely for themselves as a manner of entertainment. ... These numbers show that a significant portion of the generated images are not created and used for a direct utilitarian purpose."

The paper's central empirical claim—that a significant portion of generated images are non-utilitarian digital waste—rests entirely on the authors' own prior statistics. Reference [4] is listed only as 'Manuscript' under review, with no survey instrument, sample size, recruitment method, or response rate disclosed, and reference [3] is an in-press self-citation. Neither source is independently checkable from this paper, and no external replication is cited. Removing [3] and [4] leaves the 'significant portion' conclusion with no support, so the load-bearing premise reduces to self-cited, inaccessible prior work rather than to a derivation from public data or an independent benchmark.

full rationale

This workshop paper is a discussion/position piece rather than a formal derivation: it contains no equations, no fitted parameters, and no prediction that is mathematically forced by its own inputs. Therefore the strongest circularity patterns (self-definitional equivalence, fitted-input-called-prediction, ansatz-smuggling, renaming) do not apply. The main circularity concern is pattern 3: the central overconsumption claim is justified by the authors' own previous papers, one of which is an unpublished manuscript. That is load-bearing self-citation, because the paper's quantitative foundation—1500+ generated images per user per week, roughly 40% entertainment-only use, and fewer than 15% professional users—comes exclusively from [3] and [4], with no independent verification or even a summary of the survey methodology. The paper is transparent that it is 'outlining previous work' and 'expanding' on it, which lowers the impression of hiding the dependence, but transparency does not make the underlying evidence independent. On the other hand, the paper does contain independent intellectual content: the application of uses-and-gratifications theory, the common-pool resource dilemma framing, the educational solution pathway, and the discussion of future VR-related risks. These do not reduce to the self-cited statistics. Weighing the load-bearing nature of the self-citations against the presence of independent discussion, a moderate score of 5 is appropriate: not a construction-level circularity (which would warrant 6+), but more than a minor self-citation (which would warrant 2 or lower).

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new free parameters or invented entities. Its central premises are borrowed theories and unverifiable survey data from the authors' previous work.

assumptions (4)

domain assumption The survey data from the authors' prior work [3,4] are representative of generative AI users.
Section 2 relies on statistics (1500 images/week, 40% entertainment use) to conclude overconsumption, but the sampling method and response rates are not described.
domain assumption Uses and gratifications theory applies to generative AI use.
Section 2 hypothesizes this as a possible explanation for the explosion of generative AI use, but it is a borrowed theory and is not empirically established here.
domain assumption The common pool resource dilemma is an appropriate framework for digital overconsumption.
Section 4 applies Brown and Cameron's framework without justifying that digital resources behave like common pool resources, which is a structural analogy rather than a proven equivalence.
ad hoc to paper Digital waste is a useful lens for generative AI's environmental impact.
The paper adopts Bietti and Vatanparast's definition of data waste and uses it as the framing device, but it does not compare this framing with alternative impact metrics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Digital Overconsumption and Waste: A Closer Look at the Impacts of Generative AI." pith.science (2026). https://pith.science/paper/DHMZ7OXN

@misc{pith2026250518894,
  author       = {Pith},
  title        = {Pith review of: Digital Overconsumption and Waste: A Closer Look at the Impacts of Generative AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHMZ7OXN}},
  note         = {Machine review of arXiv:2505.18894}
}

read the original abstract

Generative Artificial Intelligence (AI) systems currently contribute negatively to the production of digital waste, via the associated energy consumption and the related CO2 emissions. At this moment, a discussion is urgently needed on the replication of harmful consumer behavior, such as overconsumption, in the digital space. We outline our previous work on the climate implications of commercially available generative AI systems and the sentiment of generative AI users when confronted with AI-related climate research. We expand on this work via a discussion of digital overconsumption and waste, other related societal impacts, and a possible solution pathway

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 9 canonical work pages

[3]

Climate implications of diffusion-based generative visual AI systems and their mass adoption

Vanessa Utz and Steve DiPaola . Climate implications of diffusion-based generative visual AI systems and their mass adoption. International Conference on Computational Creativity, in press

work page
[4]

Manuscript

Vanessa Utz and Steve DiPaola. Manuscript

work page
[1]

StabilityAI raises seed round at $1 billion value

Mureji Fatunde and Crystal Tse. StabilityAI raises seed round at $1 billion value. Bloomberg, https://www.bloomberg.com/news/articles/2022-10- 17/digital-media-firm-stability-ai-raises-funds-at-1-billion- value, 2022

work page 2022
[2]

You might actually fin d a muse in the machine (or Midjourney)

Adrian Pennington. You might actually fin d a muse in the machine (or Midjourney). NABAmplify. https://amplify.nabshow.com/articles/ic-the-muse-in-the- machine/, 2022

work page 2022
[5]

Luccioni and Alex Hernandez-Garcia

Alexandra S. Luccioni and Alex Hernandez-Garcia. Counting carbon: A survey of factors influencing the emissions of machine learning. arXiv:2302.08476v1, 2023

arXiv 2023
[6]

Data waste

Elettra Bietti and Roxana Vatanparast. Data waste. Harvard International Law Journal, 61:1-11, 2020

work page 2020
[7]

Uses and gratification research

Elihu Katz, Jay Blumler and Michael Gurevitch. Uses and gratification research. Public Opinion Quarterly, 37(4): 509- 523, 1973

work page 1973
[8]

On the psychology of TikTok use: A first glimpse from empirical findings

Christian Montag, Haibo Yand and Jon Elhai. On the psychology of TikTok use: A first glimpse from empirical findings. Frontiers in Public Health, 9:1-6, 2021

work page 2021

Show all 10 references

[9]

Watch, share or create: The influence of personality traits and user motivation on TikTok mobile video usage

Bahiya Omar and Wang Dequan. Watch, share or create: The influence of personality traits and user motivation on TikTok mobile video usage. International Association of Online Engineering, 14(4):121-136, 2020

2020
[10]

What can be done to reduce overconsumption? Ecological Economics, 32(1) :27-41, 2000

Paul Brown and Linda Cameron. What can be done to reduce overconsumption? Ecological Economics, 32(1) :27-41, 2000

2000

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [3]

Climate implications of diffusion-based generative visual AI systems and their mass adoption

Vanessa Utz and Steve DiPaola . Climate implications of diffusion-based generative visual AI systems and their mass adoption. International Conference on Computational Creativity, in press

work page

[2] [4]

Manuscript

Vanessa Utz and Steve DiPaola. Manuscript

work page

[3] [1]

StabilityAI raises seed round at $1 billion value

Mureji Fatunde and Crystal Tse. StabilityAI raises seed round at $1 billion value. Bloomberg, https://www.bloomberg.com/news/articles/2022-10- 17/digital-media-firm-stability-ai-raises-funds-at-1-billion- value, 2022

work page 2022

[4] [2]

You might actually fin d a muse in the machine (or Midjourney)

Adrian Pennington. You might actually fin d a muse in the machine (or Midjourney). NABAmplify. https://amplify.nabshow.com/articles/ic-the-muse-in-the- machine/, 2022

work page 2022

[5] [5]

Luccioni and Alex Hernandez-Garcia

Alexandra S. Luccioni and Alex Hernandez-Garcia. Counting carbon: A survey of factors influencing the emissions of machine learning. arXiv:2302.08476v1, 2023

arXiv 2023

[6] [6]

Data waste

Elettra Bietti and Roxana Vatanparast. Data waste. Harvard International Law Journal, 61:1-11, 2020

work page 2020

[7] [7]

Uses and gratification research

Elihu Katz, Jay Blumler and Michael Gurevitch. Uses and gratification research. Public Opinion Quarterly, 37(4): 509- 523, 1973

work page 1973

[8] [8]

On the psychology of TikTok use: A first glimpse from empirical findings

Christian Montag, Haibo Yand and Jon Elhai. On the psychology of TikTok use: A first glimpse from empirical findings. Frontiers in Public Health, 9:1-6, 2021

work page 2021