{"id":"1909dd4a-86e0-4143-9433-3b50ab3b320c","arxiv_id":"2505.18894","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Generative AI is framed as a driver of digital overconsumption and waste, with user surveys cited to show low awareness of energy use.","lead":"This workshop paper argues that generative AI tools encourage digital overconsumption and waste, and that most users are unaware of the energy costs. It proposes education as a solution while relying on the authors' own prior surveys for its key numbers.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central overconsumption claim rests on self-cited unpublished survey statistics with no instrument, sample size, or error bars; until the raw data or an independent replication is available, 'most AI-generated images are non-utilitarian waste' is not verifiable.","rationale":"I read the paper as a workshop position piece whose novel contribution is the framing of generative-AI output as digital overconsumption and waste. For that framing to be more than an analogy, the paper needs credible evidence that a large share of AI-generated content is low-utility. That evidence is currently two self-cited survey statistics with no measurement details. The reader's conditional verdict identifies this same weak spot, and I agree; my stress-test adds that even the inference from 'casual user' to 'non-utilitarian output' is not licensed without a direct measure of utility, and that the mean-versus-median distinction matters for a population with heavy users. I find no internal inconsistency in the argument, and I am not holding the paper to the standard of a full empirical study; it is a position paper awaiting stronger data. Hence the verdict should remain conditional rather than being upgraded or rejected.","tokens_in":3454,"tokens_out":3829,"duration_ms":36130,"concrete_test":"Request from the authors the survey instruments, de-identified raw responses, and recruitment details for [3] and [4]. If the data are released, recompute the weekly-output and use-purpose statistics with bootstrap confidence intervals and a mean-versus-median breakdown; if the median is far below 1500 images per week or the 95% confidence interval for 'entertainment-only' is wide, the average-based overconsumption claim is overstated. If the data are not released, run a preregistered replication with a representative sample of generative-AI users recruited outside the authors' own communities and compare the estimates; a substantially lower mean output or entertainment share would require revising the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim—that mass adoption of generative AI produces large volumes of non-utilitarian digital waste—rests on two numbers from the authors' own prior work: an average of over 1500 generated images per user per week, and roughly 40% of respondents using the tools solely for entertainment (Section 2). Reference [4] is described as a manuscript under review; [3] is in press. Neither the survey instrument, sampling frame, sample size, response rate, nor confidence intervals are reported in this paper, so it is impossible to tell whether these figures describe the general generative-AI user population or a self-selected convenience sample. The argument also conflates 'not professional' with 'not utilitarian': casual, hobbyist, or exploratory use can serve legitimate personal or creative purposes, and the paper provides no measurement of how many generated images are 'never looked at again.' Even the energy-cost half of the claim rests on estimates in [3] that the paper itself describes as spanning roughly 1.92–9.29 TWh, a fivefold range caused by missing usage and hardware data. The overconsumption narrative therefore cannot be evaluated independently from the current text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":3763,"tokens_out":3913,"duration_ms":36147,"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":[{"comment":"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":"Section 2"},{"comment":"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":"Section 2"},{"comment":"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.","section":"Section 2"}],"minor_comments":[{"comment":"There is a typo: 'over 1500 generated images per users' should be 'per user'.","section":"Section 2"},{"comment":"The chemical formula should be 'CO2' rather than 'C02' in the sentence about emissions from ML training.","section":"Section 1"},{"comment":"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'.","section":"Section 4"},{"comment":"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":"References"},{"comment":"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'.","section":"Section 4"},{"comment":"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.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This is a short workshop paper that largely summarizes the authors' prior work. The main new elements are the 'digital overconsumption and waste' framing and the proposed common-pool resource interpretation. The heavy reliance on an unpublished self-citation ([4]) is a concern for a journal-level review; if the survey data cannot be shared, the authors may need to reframe the paper explicitly as a position piece rather than a paper with empirical support for its central claim. The recommendation of major_revision reflects the need to make the evidence base assessable, not a judgment on the plausibility of the argument."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe 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.\n\nWhat 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.\n\nWhere 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.\n\nStill, 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.\n\nI 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.\n\nBest,\n[You]","headline":"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.","tokens_in":4169,"tokens_out":1574,"would_cite":false,"duration_ms":16853,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["generative AI","digital waste","digital overconsumption","CO2 emissions","energy consumption","user behavior","common-pool resource dilemma"],"falsifier":"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.","tokens_in":3202,"feed_emoji":"🗑️","tokens_out":8138,"duration_ms":69863,"temperature":0.7,"pith_summary":"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.","feed_headline":"Most AI images are never used again — and that's digital waste","feed_subtitle":"Generative AI users create 1,500+ images a week, mostly for fun, unaware of the CO2 cost.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the scale of adoption, over 10 million daily users, that makes individual image generation a collective environmental issue.","marker":"[1]"},{"why":"Provides the daily global output estimate of over 20 million images that anchors the overproduction claim.","marker":"[2]"},{"why":"Contains the authors' earlier energy estimates of 1.92–9.29 TWh per year and the survey data on entertainment use and iteration counts.","marker":"[3]"},{"why":"Is the still-under-review study that yields the average of over 1,500 generated images per user per week and the finding that most users had not considered energy consumption.","marker":"[4]"},{"why":"Defines digital waste as emissions and resource extraction from data infrastructures, the category the paper applies to AI-generated images.","marker":"[6]"},{"why":"Supplies the common-pool resource dilemma framework used to argue that generative-AI overconsumption is a collective problem with incomplete information.","marker":"[10]"}],"fun_headline_variants":["AI's digital waste: 1,500 images per user weekly, many unused","Generative AI overconsumption: energy-hungry image creation","AI image overproduction: most users ignore CO2 impact, study says","Digital overconsumption: AI users generate images without climate awareness","Generative AI's hidden waste: energy spent on images rarely used"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["AI's digital waste: 1,500 images per user weekly, many unused","Generative AI overconsumption: energy-hungry image creation","AI image overproduction: most users ignore CO2 impact, study says","Digital overconsumption: AI users generate images without climate awareness","Generative AI's hidden waste: energy spent on images rarely used"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000315,"raw_usage":{"total_tokens":1716,"prompt_tokens":803,"completion_tokens":913,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":419,"completion_tokens_details":{"reasoning_tokens":819}},"tokens_in":419,"tokens_out":913,"duration_ms":7916,"temperature":1.0,"reasoning_tokens":819,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:22:21.258576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"StabilityAI raises seed round at $1 billion value","cited_arxiv_id":null,"evidence_quote":"Supplies the scale of adoption, over 10 million daily users, that makes individual image generation a collective environmental issue."},{"cited_title":"You might actually fin d a muse in the machine (or Midjourney)","cited_arxiv_id":null,"evidence_quote":"Provides the daily global output estimate of over 20 million images that anchors the overproduction claim."},{"cited_title":"Climate implications of diffusion-based generative visual AI systems and their mass adoption","cited_arxiv_id":null,"evidence_quote":"Contains the authors' earlier energy estimates of 1.92–9.29 TWh per year and the survey data on entertainment use and iteration counts."},{"cited_title":"Manuscript","cited_arxiv_id":null,"evidence_quote":"Is the still-under-review study that yields the average of over 1,500 generated images per user per week and the finding that most users had not considered energy consumption."},{"cited_title":"Data waste","cited_arxiv_id":null,"evidence_quote":"Defines digital waste as emissions and resource extraction from data infrastructures, the category the paper applies to AI-generated images."},{"cited_title":"What can be done to reduce overconsumption? Ecological Economics, 32(1) :27-41, 2000","cited_arxiv_id":null,"evidence_quote":"Supplies the common-pool resource dilemma framework used to argue that generative-AI overconsumption is a collective problem with incomplete information."}],"review_version":1}