Do Decision-Makers Without Taste Stop Commissioning Creators? Checking the Problem of Image-Generating AI Through Three Links (Commissioning, Habituation, and the Creator Base) and Relating It to Research on AI Slop

The causal chain in the post

On October 9, 2026, the manga artist nakamachi_keiji posted the following on X (nakamachi_keiji 2026).

The problem with image-generating AI is not that “AI makes impressive work and takes creators’ jobs.” I think it is that “people who lack taste but hold decision-making authority believe that what they had AI make is tasteful and stop commissioning creators,” “society becomes accustomed to low quality,” and “as a result, the base of creators narrows and the culture of making declines.” (translated from the Japanese original)

The post first rejects a familiar scenario and then links an alternative scenario through three steps. The rejected scenario is that jobs disappear because AI output surpasses human work. In its place, the post proposes the following chain. The first link concerns commissioning: decision-makers without the ability to judge aesthetic quality believe that AI output is good and stop outsourcing work. The second link concerns habituation: society comes to regard low quality as normal. The third link concerns the creator base: the layer of creators who can obtain work becomes thinner, and the culture of making itself declines.

The post is a short opinion that ends with “I think,” not a claim presented with evidence. Even so, each of the three links contains a testable proposition. These propositions are whether commissions decreased, whether decision-makers judge differently from experts, whether people’s standards shift, and whether the layer of creators becomes thinner.

This note sets studies that support each link against studies that do not, and then relates the chain to research on AI slop. The answer that remains is that the entry of the chain (fewer commissions and jobs) and part of its exit (fewer uploads by existing creators) have been measured, whereas the mechanism in the middle, “decision-makers without taste,” has not been measured directly. Furthermore, within the measured range, AI images are hard to identify even for people with experience in art, so the framing that divides people by whether they have taste does not match the research.

Scope and method

• Scope: peer-reviewed papers, academic books, and research reports that correspond to the three links and to the rejected scenario. On the side of judgment, we covered expertise in aesthetic judgment, the identification and evaluation of AI-generated works, the decline in evaluations when AI involvement is disclosed, and overestimation of one’s own work and of one’s performance when using AI. On the side of markets and culture, we covered demand for creative work under generative AI, human uploads on artwork-sharing platforms, employment and participation of young workers and newcomers, repeated viewing and preference, shifting baselines, homogenization of works, degradation of models trained on generated data, the theory of markets in which quality cannot be judged, and the critique of kitsch.
• Collection: we collected about 60 candidates from academic databases (OpenAlex, Crossref, Semantic Scholar, arXiv, NBER, and SSRN) and checked bibliographic data, retractions, corrections, and the availability of peer-reviewed versions. No retracted or withdrawn sources were found. Among the sources used in the existing notes on AI slop and their sister notes, those relevant to this question were used after their full texts were obtained again.
• Verification: for 28 of the 39 sources, we obtained the full text (published version, author version, working paper version, or preprint) and checked numbers and claims against the relevant passages. For the remaining 11, the full text could not be obtained through three or more routes, so they are described within the scope of their abstracts and attributed with “according to the abstract.” Working papers and preprints for which no peer-reviewed version was found are marked “not peer reviewed.”
• Terms: we read the post’s “taste” as the ability to judge and choose the quality of works. In research, this ability is measured as expertise in art, art education, or experience in art. A “decision-maker” is the person who decides whether to outsource work, which corresponds to the client or employer in research.

The rejected scenario: did jobs decrease, and was it because AI output was better?

The volume of work

A decline in demand for outsourced image creation has been measured on a global freelancing platform. According to the abstract of Demirci, Hannane, and Zhu (2025), job posts related to image creation fell by 17% after image-generating AI appeared. The same study reported that, after ChatGPT, automation-prone job posts related to writing and coding fell by 21% within eight months, and that the remaining posts were more complex and offered higher pay. According to the abstract of Hui, Reshef, and Zhou (2024), after the release of ChatGPT, DALL-E 2, and Midjourney, freelancers in highly affected occupations experienced reductions in both the number of jobs and earnings. Teutloff et al. (2025) stated that demand for skills that AI can substitute (such as writing and translation) fell by 20% to 50% relative to the counterfactual trend.

However, in the same period, the number of commissions increased in another market. Zhang, Yuan, and Xiong (2026) examined a Chinese commission platform for illustrations before and after the leak of a generative model specialized in anime-style images (NovelAI’s model), comparing the anime-style market with the wallpaper market. In the anime-style market, the average price fell by 64%, the number of orders rose by 121%, and total revenue rose by 56%. The increase came from cheap orders for personal use, whereas the price, volume, and revenue of commercial orders remained relatively stable. After the leak, 97% of orders were won by creators who had registered before the leak.

The results from these two markets narrow the scope of the post’s claim that decision-makers “stop commissioning.” Whether the number of commissions falls or rises differs by market, and what occurred in both markets was a fall in the price of cheap work.

Did differences in quality affect the decline?

The post rejected the view that attributes job losses to the quality of AI output. The closest evidence on this point is the result of Hui et al. (2024). According to the abstract, workers’ quality, measured by past performance and employment, did not moderate the decline in employment. Rather, there was suggestive evidence that top freelancers were affected more. The finding that high-quality workers were not protected does not contradict the post’s view. However, this study did not separately measure whether clients judged AI output to be better than human work or chose it because it was cheaper.

The results of Zhang et al. (2026) also point to a comparison of price rather than of quality. The authors attributed the stability of commercial orders to the fact that the level required for commercial use exceeds the capabilities of current generative AI. If cheap personal orders increased while expensive commercial orders did not change, then generative AI first entered work with low quality requirements.

The first link: do decision-makers believe that AI output is good?

Can AI-generated works be identified?

People are poor at identifying what AI has made. Porter and Machery (2024) showed readers who were not poetry experts poems by well-known poets and poems generated by AI. Accuracy was 46.6%, slightly below chance (16,340 judgments). Readers tended to judge AI poems as human-written, and AI poems were rated more highly for qualities such as rhythm and beauty. For images, according to the abstract of Gangadharbatla (2022), participants in online survey experiments could not accurately identify AI-generated artwork and tended to associate representational art with humans and abstract art with machines.

Identifying and preferring move independently. van Hees et al. (2025) showed pairs of DALL-E 2 paintings and human paintings, asking one group to choose the painting they liked and another group to choose the painting made by AI. When the creator was not disclosed, the preference group chose the AI paintings significantly more often than chance (d = 0.48). The identification group identified the AI paintings significantly better than chance, but the effect was small (d = 0.31). Participants’ experience in art was not significantly correlated with either their preference for AI paintings or their identification accuracy.

There is also no result showing that people with art education are particularly good at identification. Huang, Shen, and Gan (2026) showed six pairs of human and AI paintings to 356 participants with different levels of art education. Overall accuracy was 54.9% (95% confidence interval 52.5% to 57.3%), slightly above chance. For two of the six pairs, more than 60% of participants judged the human painting to be AI-generated, and accuracy fell below chance. Art education was associated with performance, but only the group with “some art background” performed significantly above chance. In Study 1 of Chamberlain et al. (2018), observers often distinguished computer-generated from human-made works, so how easy identification is depends on the kind of work and the generation technique.

What do experts judge to be good?

Expertise changes what people prefer more than their ability to identify. According to the abstract of Leder et al. (2014), art experts showed more attenuated emotional reactions than non-experts and liked negative artworks more. The same pattern appeared for non-art pictures, and the authors interpreted this as a sign that experts may process visual stimuli differently in general. If experts and non-experts prefer different works, a decision-maker who chooses a different picture from an expert is not necessarily making an error of judgment.

For material close to what clients judge, AI output received higher evaluations. Zhang and Gosline (2023) had professional content creators from a leading global consulting firm and ChatGPT-4 produce advertising content for products and persuasive content for campaigns, and online participants rated their quality. Content produced by AI alone and content in which AI made the final decision were rated more highly than content produced by human experts alone. Disclosing the creator narrowed the gap but did not reverse it. The material was text, not images.

These results diverge from the post’s phrase “people who lack taste” in two respects. First, because difficulty in identification extends to people with experience in art, the weakness of decision-makers’ judgment cannot be explained only by a lack of taste. Second, because many viewers prefer AI works when the creator is not disclosed, a decision-maker who chooses AI output may be choosing in line with the viewers’ preferences. The post’s “low quality” is quality measured by the creators’ standard, which is not the same as quality measured by the viewers’ preferences.

Does what one had AI make look better?

The post’s “believe” refers to overestimating what one had AI make. This mechanism has solid support. Norton, Mochon, and Ariely (2012) showed that people who assembled IKEA boxes, folded origami, or built Lego sets valued their amateur creations about as highly as experts’ creations and expected others to share their opinion. In Kruger and Dunning (1999), people who scored in the bottom quartile on tests of humor, grammar, and logic were actually near the 12th percentile but estimated themselves to be at the 62nd.

Using AI changes the form of this misestimation. In Study 1 of Fernandes et al. (2026), 246 participants used ChatGPT to solve 20 logical reasoning problems from the Law School Admission Test; their performance improved by three points compared with a norm population, but they overestimated their performance by four points. The usual pattern in which less able people overestimate more disappeared with AI use, and participants with more knowledge of AI judged their performance less accurately. Study 2 (452 participants) replicated these results. Steyvers et al. (2025) showed that users overestimated the accuracy of LLM answers and that longer explanations raised their confidence even when accuracy did not change.

However, the material in these studies was assembly, logical problems, and knowledge questions, not aesthetic judgment. No study was found that measured whether people in charge of commissioning rate the quality of images they had AI make more highly than experts do.

Does knowing that a work is AI-made lower its evaluation?

A person who had AI make an image knows that the image is AI-made. It has been shown repeatedly that evaluations drop when people are told that the creator is AI. Bellaiche et al. (2023) randomly attached a “human-created” or “AI-created” label to paintings made by AI and showed that paintings labeled as human-created were rated more highly on all four criteria: liking, beauty, profundity, and worth. According to the abstract of Millet et al. (2023), across four experiments (1,708 participants), the same artwork was perceived as less creative, evoked less awe, and was preferred less when labeled as AI-made. According to the abstract of Raj, Berg, and Seamans (2026), in experiments with 27,491 participants, disclosing AI involvement in creative writing lowered its evaluation, and the decline persisted across interventions proposed in prior research.

This result works in the opposite direction from the post’s “believe.” Which tendency is stronger in the context of commissioning decisions, the tendency to value one’s own creations highly or the tendency to devalue works known to be AI-made, has not been measured. Measuring situations in which the two tendencies conflict within the same person is the next step in testing the first link.

The second link: does society become accustomed to low quality?

Repeated viewing and preference

People come to like what they see often. According to the abstract of Montoya et al. (2017), a meta-analysis of 268 curve estimates from 81 articles found that the effect of repeated exposure follows an inverted-U shape, first raising liking and then lowering it as exposures increase, and that this shape appeared for visual stimuli. The abstract of Reber, Schwarz, and Winkielman (2004) traced the effects of symmetry, prototypicality, and repetition to processing fluency. In this framework, the more fluently a perceiver can process an object, the more beautiful it feels.

This mechanism has also been observed in the formation of an artistic canon. Cutting (2003) counted how often 132 Impressionist images appeared in books in a university library (4,232 appearances in 980 books) and showed that adults’ preferences correlated with these frequencies. Children’s preferences did not correlate with the frequencies, and the author concluded that repeated exposure helps maintain an artistic canon.

From this mechanism, the inference that a flood of AI images would make their style familiar and therefore liked is plausible. However, no study was found that measured whether exposure to AI images lowered people’s standards of quality. In addition, what repeated viewing produces is “preference for a familiar style,” not “preference for low quality.” As the previous section showed, who holds the standard for judging quality as low differs between creators and viewers.

Shifting baselines

The concept closest to the post’s “become accustomed” is the shifting baseline syndrome from ecology. According to the abstract of Soga and Gaston (2018), generations without knowledge of past conditions accept the conditions in which they grew up as normal, and the threshold at which they accept environmental degradation keeps falling. The authors argued that self-reinforcing feedback loops accelerate this decline. However, this concept was proposed for environmental degradation, and no empirical study that applied it to aesthetic standards was found.

Homogenization as the measured outcome

What has been measured about generative AI and the quality of works is homogenization rather than a decline in quality. Zhou and Lee (2024) analyzed more than 4 million artworks by more than 50,000 users and reported that adopting image-generating AI increased productivity by 25% and the likelihood of receiving a favorite per view by 50%. At the same time, the average novelty of artwork content declined, while the novelty of the most novel works increased. Doshi and Hauser (2024) showed that short stories written with AI ideas were individually rated as better but became more similar to one another. Agarwal, Naaman, and Vashistha (2025) showed in an experiment with 118 participants from India and the United States that AI writing suggestions pulled the writing of Indian participants toward Western styles. Padmakumar and He (2024) showed that writing with a model tuned to follow instructions made writers’ texts more similar to one another, whereas writing with the model before such tuning did not.

In light of these results, the post’s “become accustomed to low quality” matches the research better when restated as “become accustomed to similar works.” In Zhou and Lee (2024), peer evaluations (favorites) actually rose, and there is no evidence that viewers perceived a decline in quality.

When the cost of making works fell, did quality fall?

In past technological changes, there are cases in which quality did not fall when the cost of making works fell. Waldfogel (2012) built indices of the quality of new works from critics’ retrospective lists and from sales and airplay data for the period after Napster, when the cost of producing and distributing recorded music fell. The quality of new music did not fall after Napster and rose according to two usage-based indices. Aguiar and Waldfogel (2018) analyzed the tripling of the number of new music products between 2000 and 2008 brought about by digitization. Because it is hard to predict which cultural works will succeed before release, an increase in the number of works also increases the number of successful works, which raises consumer benefits substantially.

These two studies concern the distribution of music, not generative AI. Even so, they show that the inference “when the cost of making works falls, the quality of culture falls” did not always hold in past cases.

The third link: does the creator base narrow, and does the culture of making decline?

Did creators stop uploading?

A decline in uploads by existing creators has been measured on two platforms. Kim, Jin, and Lee (2026, not peer reviewed) compared illustrators and comic artists on Pixiv, an artwork-sharing platform for anime and manga, before and after the launch of NovelAI’s image generator on October 3, 2022. Fewer than 0.5% of existing creators adopted AI, and uploads by illustrators who did not use AI fell by 10.1%. The decline was larger, at 14.3%, for creators who attached links to commercial websites, and it was especially large for the top 1% of creators by upload volume. After the launch, 22% of illustration posts were AI works, compared with 4% of comic posts. As mechanisms, the study showed that bookmarks per post (viewers’ attention) decreased and that creators whose main IP (such as Pokémon) was more heavily invaded by AI works reduced their uploads more.

Lin (2024, not peer reviewed) showed that after DeviantArt introduced an image-generation feature in November 2022 and set artworks to be included in training data by default, digital artists who did not use AI published 21% fewer artworks than craft artists. Among artists who also posted on Instagram, only uploads to DeviantArt decreased, and the quality of their published works (responses) did not change. The author interpreted this decline as artists holding back their uploads because of concern about their works being used for training.

What these two studies show is a decline in uploads by creators who were already active, driven by competition for attention and concern about training. This mechanism differs from the post’s decline driven by decision-makers who stop commissioning.

Did the entry-level layer become thinner?

No study was found that measured whether the entry-level layer in creative fields (apprentices, newcomers, and those who build skills through cheap work) became thinner. In other fields, declines among young workers and newcomers have been measured. Brynjolfsson, Chandar, and Chen (2026, not peer reviewed) reported from US payroll data that employment of workers aged 22 to 25 in occupations highly exposed to AI was 19% lower than that of their peers in less exposed occupations, while experienced workers showed no comparable gap (version of August 2026). Burtch, Lee, and Chen (2024) showed that visits and questions on Stack Overflow decreased after ChatGPT and that the decrease was concentrated among newer users. In developer communities on Reddit, activity did not decrease, and the authors attributed this to social ties buffering the decline.

There are also cases in which the path to acquiring skills breaks down. According to the abstract of Beane (2019), in robotic surgery, the role of trainees in operations was limited and approved methods of learning stopped working. Only a minority of trainees who engaged in informal, norm-challenging practices (shadow learning) became competent, and the supply of experts decreased relative to demand. The scenario in which cheap work is replaced by AI and newcomers lose opportunities to build skills is plausible from this mechanism, but it has not been tested in creative fields.

There are also results in the opposite direction. Brynjolfsson, Li, and Raymond (2025) showed in a customer support setting that AI recommendations raised productivity by 15% on average, with larger effects for less skilled and less experienced workers. According to the abstract of Gao, Liu, and Ye (2024), when freelancers specializing in image creation adopted generative AI, the popularity of their works and the diversity of their content increased. Whether the creator base narrows or widens depends on whether it is counted as “people who draw by hand” or “people who release works, including with AI.”

How can a decline in the culture of making be measured?

No study was found that directly measured a “decline in the culture of making.” The closest arguments concern the training data of generative models. Shumailov et al. (2024) showed that indiscriminately using model-generated content to train the next generation causes defects in which the tails of the original distribution disappear, and they called this model collapse. Alemohammad et al. (2024) showed that when image-generation models are repeatedly trained on their own outputs, their quality or diversity declines progressively unless each generation has enough fresh real data.

These results show that generative models depend on the layer of human creators. Greenberg (1939) named, as the precondition of kitsch (a substitute for genuine culture, mechanical and made by formulas), the availability close at hand of a fully matured cultural tradition. If the layer of human works becomes thinner, the new works from which generative models learn also become fewer. However, this is an argument about model performance, not evidence that human culture declines.

Relating the chain to research on AI slop

Externalized verification in commissioning

Willison (2024), who established the term AI slop, did not place the core of slop in poor quality but in sharing unreviewed generated content with others. Building on this definition, the sister note AI Slop: Reading It as Outsourced Verification, Not Low Quality read slop not as a property of things but as the allocation of the burden of verification between producers and recipients.

The post’s scenario is the form that this reading takes in commissioning. In conventional commissioning, a person who checks and takes responsibility for the quality of the work (the creator) stood between the decision-maker and the viewers. When decision-makers have AI make images and use them themselves, this position becomes empty, and the task of checking moves to the decision-makers’ own judgment. The studies in the first link show, with other material, that this transferred judgment is hard to rely on. People’s own creations look good to them (Norton et al. 2012), people overestimate their own performance when using AI (Fernandes et al. 2026), and AI works are hard to identify even for people with experience (van Hees et al. 2025, Huang et al. 2026). The sister note left the question of “how to put a price on the work of taking on verification.” The post’s scenario depicts the situation in which that price becomes zero because decision-makers consider their own check sufficient.

Buyers who cannot judge quality, and the withdrawal of good products

In economics, the post’s “people who lack taste but hold decision-making authority” correspond to buyers who cannot judge quality. Akerlof (1970) showed that in the used-car market, when buyers cannot tell good cars from bad ones, both sell at the same price, sellers of good cars withdraw from the market, and bad cars drive out good ones. The finding of Hui et al. (2024) that workers with strong past records were not protected from the decline does not contradict this mechanism.

However, what matters in the lemons mechanism is not buyers’ taste but the asymmetry of information about quality. As the first link showed, AI images are hard to identify even for people with experience in art. If so, educating decision-makers would not change much about the inability to identify. Akerlof named quality signals such as guarantees, brand names, and licensing as institutions that counteract this kind of market failure. Given the studies showing that evaluations drop when a work is disclosed as AI-made (Bellaiche et al. 2023, Raj et al. 2026), the label “made by a human” already works as a substitute quality signal. As the sister note pointed out, this signal works only while the origin is reported truthfully.

The lineage of judging quality by “taste”

The post judges low quality by whether one has “taste.” The sister note Slop Seen Through Design History: The Profession Began as a Countermeasure to Shoddy Mass Production showed that this way of speaking has precedents. In 1852, Henry Cole and his colleagues opened an exhibition of products they judged to be in bad taste (the Chamber of Horrors) to educate consumers’ taste, but it closed within two weeks after manufacturers protested. Greenberg (1939) described kitsch as a substitute for “those who, insensible to the values of genuine culture,” still want culture. Later design history and sociology showed that judgments of bad taste also served as a device for reproducing class distinctions. The post’s “society becomes accustomed to low quality” takes the same form as this lineage.

Placed in this lineage, the post’s claims are best handled in two parts. One part depends on who judges quality, and it includes “people who lack taste” and “low quality.” The studies in the first link show that many viewers prefer AI works when the creator is not disclosed, so this part presupposes the creators’ standard. The other part does not depend on judgments of taste, and it includes the removal of the person who checks quality from the production process, the decline in uploads by existing creators, and the homogenization of works. This part has been measured, and, like the sister note’s definition of AI slop as the allocation of the burden of verification, it stands even after passing through the criticism of the class character of taste.

Judgment on the three links

What can be said about the post’s claims

Of the post’s chain, the parts best supported by research are that cheap outsourced work is replaced by AI or falls in price, and that uploads by creators who do not use AI decrease. The former has been measured on a global freelancing platform (Demirci et al. 2025) and on a Chinese commission platform (Zhang et al. 2026), and the latter on Pixiv (Kim et al. 2026) and DeviantArt (Lin 2024). However, the mechanisms of the latter are competition for attention and concern about training, not the end of commissions.

The “mistaken belief of decision-makers without taste” at the core of the post is plausible from studies of overestimating one’s own creations and of overconfidence when using AI, but it has not been measured in commissioning. Moreover, the difficulty of identifying AI works extends to people with experience in art, and many viewers prefer AI works when the creator is not disclosed. Therefore, a claim framed as blaming decision-makers’ lack of taste does not match the research. What matches the research is a claim framed as follows: the person who checks and takes responsibility for quality is removed from the production process, and no one bears that check anymore.

“Becoming accustomed to low quality” and “the decline of the culture of making” have not been measured so far. What has been measured is that works become more similar to one another and that models degrade when trained on their own output. When music was digitized, quality did not fall even though production costs fell, and as the number of works increased, so did the number of successful works (Waldfogel 2012, Aguiar and Waldfogel 2018). Whether generative AI follows the same path as this case, or follows a different path because the layer of people who draw by hand becomes thinner, is not yet known.

Gaps

• Clients’ judgments: No study was found that measured why people in charge of commissioning in companies or as individuals adopted AI-generated images instead of outsourcing (judgment of quality, price, or speed). No peer-reviewed study was found that measured a shift from outsourcing to in-house production either.
• Evaluation of images one had AI make: No study was found that compared the quality of images people had AI make with experts’ evaluations. Situations in which overestimation of one’s own creations conflicts with the decline in evaluation caused by knowing that a work is AI-made have not been measured.
• Shifting standards: No study was found that measured whether continued exposure to AI images shifts people’s standards of quality.
• Entry-level creators: No study was found that measured how the path by which newcomers in illustration and manga build skills through cheap work changed after generative AI. Apart from the Pixiv study (Kim et al. 2026), no academic study of the Japanese manga, anime, or illustration industries was found.
• Historical analogies: In this search, no academic study was found that measured how cheap substitutes changed the base of a profession, such as photography and portrait painters or desktop publishing and designers. The lineage of design reform and the critique of kitsch is covered in Slop Seen Through Design History: The Profession Began as a Countermeasure to Shoddy Mass Production.
• Homogenization and creators’ income: How homogenization affects creators’ income and their willingness to continue has not been measured. The fulfillment of dreams through generative AI and the value of making are covered in Does a Dream Die When AI Fulfills It? Checking "Instant Dream Death," "Dream Deflation," the "Obsession Gap," "Passion Capitalism," and "Life Prototypes" Against the Academic Literature.

Unverified items

• No main claim remains unverified.
• Eleven sources were used only within the scope of their abstracts (Demirci et al. 2025, Hui et al. 2024, Gao et al. 2024, Gangadharbatla 2022, Leder et al. 2014, Millet et al. 2023, Raj et al. 2026, Montoya et al. 2017, Reber et al. 2004, Soga and Gaston 2018, Beane 2019). In the text, each is attributed to its authors with “according to the abstract.”
• Three sources are not peer reviewed (Kim et al. 2026, Lin 2024, Brynjolfsson et al. 2026), and each is marked “not peer reviewed” in the text. Zhang et al. (2026) has a peer-reviewed version (LNCS), and its text was read in both the arXiv version and the peer-reviewed version.
• Brynjolfsson, Chandar, and Chen (2026) is a working paper still under revision, and its numbers are from the version of August 2026. Earlier versions reported 13% (data as of July 2025) and 16% (data as of September 2025).
• Brynjolfsson, Li, and Raymond (2025) was read in the arXiv version (second version, November 2024), and its numbers were not checked against the published version.
• Chamberlain et al. (2018) was read in the manuscript that the authors deposited in an institutional repository. Fernandes et al. (2026) was read in the arXiv version (second version, January 2025), and its numbers were confirmed to match the abstract of the published version. Waldfogel (2012) and Aguiar and Waldfogel (2018) were read in their NBER working paper versions.
• Greenberg (1939) was read in a reprinted PDF.

Related notes

• AI Slop: Reading It as Outsourced Verification, Not Low Quality: A note that reads AI slop as the externalization of verification. The last section of this note applies that definition to commissioning.
• Slop Seen Through Design History: The Profession Began as a Countermeasure to Shoddy Mass Production: A sister note that traces how design as a profession began as a countermeasure against poor mass-produced goods and how judgments of bad taste also served as a class device.
• The History of Slop and Countermeasures: Why the Same Wager Keeps Failing: A note that traces the cycle of low-quality products and countermeasures since the printing press.
• Does a Dream Die When AI Fulfills It? Checking "Instant Dream Death," "Dream Deflation," the "Obsession Gap," "Passion Capitalism," and "Life Prototypes" Against the Academic Literature: A note that examines the falling price of making things with generative AI from the side of dreams and passion. It shares Zhang et al. (2026) and the Pixiv study with this note.

References

All sources were accessed on 2026-10-11.

The post

• nakamachi_keiji (Nakamachi, manga artist). (2026, October 9). The problem with image-generating AI is not that “AI makes impressive work and takes creators’ jobs”… [Post on X, in Japanese]. https://x.com/nakamachi_keiji/status/2108517562543923479

Demand for work and markets

• Demirci, O., Hannane, J., & Zhu, X. (2025). Who is AI replacing? The impact of generative AI on online freelancing platforms. Management Science, 71(10), 8097–8108. https://doi.org/10.1287/mnsc.2024.05420
• Hui, X., Reshef, O., & Zhou, L. (2024). The short-term effects of generative artificial intelligence on employment: Evidence from an online labor market. Organization Science, 35(6), 1977–1989. https://doi.org/10.1287/orsc.2023.18441
• Teutloff, O., Einsiedler, J., Kässi, O., Braesemann, F., Mishkin, P., & del Rio-Chanona, R. M. (2025). Winners and losers of generative AI: Early evidence of shifts in freelancer demand. Journal of Economic Behavior & Organization, 235, 106845. https://doi.org/10.1016/j.jebo.2024.106845
• Zhang, K., Yuan, Z., & Xiong, H. (2026). The impact of generative artificial intelligence on market equilibrium: Evidence from a natural experiment. In Web and Internet Economics (Lecture Notes in Computer Science, pp. 609–626). Springer. https://doi.org/10.1007/978-3-032-08560-3_34 (arXiv version: https://arxiv.org/abs/2311.07071)
• Gao, M., Liu, J., & Ye, Q. (2024). Empowering innovation: An empirical study on the impact of generative AI on online freelancers’ performance. In ICIS 2024 Proceedings (41). https://aisel.aisnet.org/icis2024/diginnoventren/diginnoventren/41

Identification, preference, and expertise

• Porter, B., & Machery, E. (2024). AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. Scientific Reports, 14, 26133. https://doi.org/10.1038/s41598-024-76900-1
• Gangadharbatla, H. (2022). The role of AI attribution knowledge in the evaluation of artwork. Empirical Studies of the Arts, 40(2), 125–142. https://doi.org/10.1177/0276237421994697
• van Hees, J., Grootswagers, T., Quek, G. L., & Varlet, M. (2025). Human perception of art in the age of artificial intelligence. Frontiers in Psychology, 15, 1497469. https://doi.org/10.3389/fpsyg.2024.1497469
• Huang, Y., Shen, Q., & Gan, F. (2026). Judging AI-generated paintings: Art-education differences in source classification. Frontiers in Psychology, 17, 1903106. https://doi.org/10.3389/fpsyg.2026.1903106
• Chamberlain, R., Mullin, C., Scheerlinck, B., & Wagemans, J. (2018). Putting the art in artificial: Aesthetic responses to computer-generated art. Psychology of Aesthetics, Creativity, and the Arts, 12(2), 177–192. https://doi.org/10.1037/aca0000136
• Leder, H., Gerger, G., Brieber, D., & Schwarz, N. (2014). What makes an art expert? Emotion and evaluation in art appreciation. Cognition and Emotion, 28(6), 1137–1147. https://doi.org/10.1080/02699931.2013.870132
• Zhang, Y., & Gosline, R. (2023). Human favoritism, not AI aversion: People’s perceptions (and bias) toward generative AI, human experts, and human–GAI collaboration in persuasive content generation. Judgment and Decision Making, 18, e41. https://doi.org/10.1017/jdm.2023.37

Evaluations when a work is disclosed as AI-made

• Bellaiche, L., Shahi, R., Turpin, M. H., Ragnhildstveit, A., Sprockett, S., Barr, N., Christensen, A., & Seli, P. (2023). Humans versus AI: Whether and why we prefer human-created compared to AI-created artwork. Cognitive Research: Principles and Implications, 8, 42. https://doi.org/10.1186/s41235-023-00499-6
• Millet, K., Buehler, F., Du, G., & Kokkoris, M. D. (2023). Defending humankind: Anthropocentric bias in the appreciation of AI art. Computers in Human Behavior, 143, 107707. https://doi.org/10.1016/j.chb.2023.107707
• Raj, M., Berg, J. M., & Seamans, R. (2026). The artificial intelligence disclosure penalty: Humans persistently devalue AI-generated creative writing. Journal of Experimental Psychology: General, 155(4), 896–915. https://doi.org/10.1037/xge0001889

Overestimation of one’s own work and of performance with AI

• Norton, M. I., Mochon, D., & Ariely, D. (2012). The IKEA effect: When labor leads to love. Journal of Consumer Psychology, 22(3), 453–460. https://doi.org/10.1016/j.jcps.2011.08.002
• Kruger, J., & Dunning, D. (1999). Unskilled and unaware of it: How difficulties in recognizing one’s own incompetence lead to inflated self-assessments. Journal of Personality and Social Psychology, 77(6), 1121–1134. https://doi.org/10.1037/0022-3514.77.6.1121
• Fernandes, D., Villa, S., Nicholls, S., Haavisto, O., Buschek, D., Schmidt, A., Kosch, T., Shen, C., & Welsch, R. (2026). AI makes you smarter but none the wiser: The disconnect between performance and metacognition. Computers in Human Behavior, 175, 108779. https://doi.org/10.1016/j.chb.2025.108779
• Steyvers, M., Tejeda, H., Kumar, A., Belem, C., Karny, S., Hu, X., Mayer, L. W., & Smyth, P. (2025). What large language models know and what people think they know. Nature Machine Intelligence, 7(2), 221–231. https://doi.org/10.1038/s42256-024-00976-7

Repeated viewing, preference, and shifting baselines

• Montoya, R. M., Horton, R. S., Vevea, J. L., Citkowicz, M., & Lauber, E. A. (2017). A re-examination of the mere exposure effect: The influence of repeated exposure on recognition, familiarity, and liking. Psychological Bulletin, 143(5), 459–498. https://doi.org/10.1037/bul0000085
• Reber, R., Schwarz, N., & Winkielman, P. (2004). Processing fluency and aesthetic pleasure: Is beauty in the perceiver’s processing experience? Personality and Social Psychology Review, 8(4), 364–382. https://doi.org/10.1207/s15327957pspr0804_3
• Cutting, J. E. (2003). Gustave Caillebotte, French Impressionism, and mere exposure. Psychonomic Bulletin & Review, 10(2), 319–343. https://doi.org/10.3758/BF03196493
• Soga, M., & Gaston, K. J. (2018). Shifting baseline syndrome: Causes, consequences, and implications. Frontiers in Ecology and the Environment, 16(4), 222–230. https://doi.org/10.1002/fee.1794

Homogenization and falling production costs

• Zhou, E., & Lee, D. (2024). Generative artificial intelligence, human creativity, and art. PNAS Nexus, 3(3), pgae052. https://doi.org/10.1093/pnasnexus/pgae052
• Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290
• Agarwal, D., Naaman, M., & Vashistha, A. (2025). AI suggestions homogenize writing toward Western styles and diminish cultural nuances. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (pp. 1–21). ACM. https://doi.org/10.1145/3706598.3713564
• Padmakumar, V., & He, H. (2024). Does writing with language models reduce content diversity? In The Twelfth International Conference on Learning Representations (ICLR 2024). https://arxiv.org/abs/2309.05196
• Waldfogel, J. (2012). Copyright protection, technological change, and the quality of new products: Evidence from recorded music since Napster. The Journal of Law and Economics, 55(4), 715–740. https://doi.org/10.1086/665824
• Aguiar, L., & Waldfogel, J. (2018). Quality predictability and the welfare benefits from new products: Evidence from the digitization of recorded music. Journal of Political Economy, 126(2), 492–524. https://doi.org/10.1086/696229

Creators’ uploads, the entry-level layer, and skill paths

• Kim, S., Jin, G. Z., & Lee, E. (2026). Does generative AI crowd out human creators? Evidence from Pixiv (NBER Working Paper No. 34733). National Bureau of Economic Research. https://doi.org/10.3386/w34733
• Lin, S. (2024). Hiding from generative AI [Working paper, version August 2024]. Rotman School of Management, University of Toronto. https://tse-fr.eu/sites/default/files/TSE/documents/conf/2025/digital/lin.pdf
• Brynjolfsson, E., Chandar, B., & Chen, R. (2026). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence [Working paper, revised August 2026]. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
• Burtch, G., Lee, D., & Chen, Z. (2024). The consequences of generative AI for online knowledge communities. Scientific Reports, 14, 10413. https://doi.org/10.1038/s41598-024-61221-0
• Beane, M. (2019). Shadow learning: Building robotic surgical skill when approved means fail. Administrative Science Quarterly, 64(1), 87–123. https://doi.org/10.1177/0001839217751692
• Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044

Models trained on generated output

• Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631(8022), 755–759. https://doi.org/10.1038/s41586-024-07566-y
• Alemohammad, S., Casco-Rodriguez, J., Luzi, L., Humayun, A. I., Babaei, H., LeJeune, D., Siahkoohi, A., & Baraniuk, R. G. (2024). Self-consuming generative models go MAD. In The Twelfth International Conference on Learning Representations (ICLR 2024). https://arxiv.org/abs/2307.01850

Quality information, kitsch, and slop

• Akerlof, G. A. (1970). The market for “lemons”: Quality uncertainty and the market mechanism. The Quarterly Journal of Economics, 84(3), 488–500. https://doi.org/10.2307/1879431
• Greenberg, C. (1939). Avant-garde and kitsch. Partisan Review, 6(5), 34–49. https://www.paduan.dk/Kunsthistorie%202008/Tekster/CLEMENT%20GREENBERG-AVANT-GARDE%20AND%20KITSCH.pdf
• Willison, S. (2024, May 8). Slop is the new name for unwanted AI-generated content [Blog post]. https://simonwillison.net/2024/May/8/slop/ (used as primary evidence of the term’s origin, not as an academic source)