The claim
In May 2026 Emily Segal, a co-founder of the culture and trends consultancy Nemesis and of K-HOLE, posted a definition on X: tasteslop is “slop made out of things considered to be tasteful.” Her examples run from a generated brand moodboard of Rimowa luggage, a skinny-neck kettle and a Dieter Rams book, to Muzak, to a pair of split-toe tabi shoes. The argument has five parts.
Taste is a relation, not a property: “a socially validated relation between objects, people, histories, scenes, and timing.” The same object can be tasteful, vulgar, camp or slop depending on the route by which it arrives. Taste needs three things, discernment, pattern recognition and idiosyncrasy, the last being a personal tie that others cannot clone. Copying what you think is tasteful is a mark of bad taste, because what becomes iconic becomes obvious. Taste needs tension, and tasteslop “has no such tension within it.” And models make all this worse, because they are very good at picking out the rote signs of taste and very bad at knowing when those signs have stopped meaning anything.
She does not say a machine cannot find good things. She asks whether an agent that combs reviews and turns up an out-of-the-way restaurant has taste, and leaves it open. Her objection is to taste with no one to answer to. In a line she quotes from Greg Fong, “Nobody can have taste unless somebody else can see it,” and a model “can only look for data indexes of taste.”
These are claims about people and markets, and several of them can be checked. We took them as four questions.
Do models converge on one taste?
Yes, within a model and across models. The Artificial Hivemind study collected 26,000 real open-ended queries, such as “Write a metaphor about time.” On 100 of them, each of 25 models wrote 50 answers per query at temperature 1.0, and in 79% of cases the answers to one query had an average pairwise similarity above 0.8, by a sentence-embedding measure. For the time metaphor, the 50 answers from each of the 25 models fell into two clusters, “time is a river” and “time is a weaver.” In Nanyang Technological University’s test of taste, models from OpenAI, Anthropic and DeepSeek played 554,940 survey respondents, who were asked which of 17 kinds of music they like. Real respondents said yes to 24.18% of them. The model respondents said yes to more, by 22 points (OpenAI), 22 (DeepSeek) and 16 (Anthropic), and most of all for the less popular genres, where the share liked rose by 32, 29 and 20 points.
Ideas converge too. In a study of 36 people, those who used ChatGPT to come up with ideas produced less semantically distinct ideas from one person to the next than with another idea tool, and felt less responsible for them. Our own count of idea families found the most common idea took 32 to 65% of a model’s thirty answers to one brief (Models write alike, not necessarily worse).
One result shows why this is slop and not only sameness. In a study with more than 100 NLP researchers, blind reviewers judged a model’s research ideas more novel than the experts’ own (p < 0.05), and slightly weaker on feasibility. The same paper found the model’s ideas lacked diversity. An idea can read as new to the reviewer who sees it while the set of ideas repeats.
The claim holds with one addition. What models share is a consensus, and a consensus can be right about the average. In a study of machine taste in K-pop, vision and audio models agreed with one another more on visuals than on sound, and their visual consensus rankings correlated positively with views on YouTube, with the strength varying by artist group. Sameness among models is not the same as being wrong. It means the model’s taste is the shared taste.
Do the signs of taste lose their worth as they spread?
Where taste signals identity, yes. Berger and Le Mens followed first names in France and the United States for more than a hundred years. Names adopted fast died faster, and expecting parents shied away from names whose popularity had just jumped, because fads are seen as short-lived. In four studies, Berger and Heath found that people diverge from the majority in domains that signal who they are, such as music or hairstyles, and not in domains that do not, such as backpacks or stereos. In identity domains they avoided what majorities preferred and dropped preferences they discovered they shared with them.
The pull is not toward distance for its own sake. Askin and Mauskapf measured nearly 27,000 songs on Billboard’s Hot 100. Songs that sounded too much like what was already there did worse than songs with some differentiation. In three studies of consumer products, typicality and novelty were jointly and about equally effective in explaining preference, and each hid the other’s effect until the other was held constant (Hekkert and colleagues). In 17.9 million papers, the highest-impact work combined novelty with otherwise conventional ties to prior work (Uzzi and colleagues). The pattern is a conventional base with one step that is different.
Popularity is also noisy. In an artificial market of 14,341 people downloading unknown songs, showing people what others had chosen made success more unequal and less predictable. The best songs rarely did poorly and the worst rarely did well, but any other result was possible (Salganik and colleagues). Simulations of recommendation systems find the same loop in software. When a system is trained on choices made by people it has already steered, behaviour becomes more alike without becoming more useful (Chaney and colleagues).
So Segal’s third claim holds in the domains where taste is a signal, and it takes a sharper form: what pays is a recognisable base with a step that has not been taken. We add one inference of our own. The signs of a famous person’s taste are the most widely copied signs there are, so a framework of that person is where this loss shows first.
Who answers for the taste?
The Nanyang study is the closest test of the claim that models have the signs without the social body. It compared model respondents with real ones on the relations among tastes and on how tastes line up with age, class, gender and race. It found the relations among genres “completely distorted”. Models pushed musical taste toward the young, revived class links that are out of date, and caricatured gender and race. They had learned what groups are said to like and not how likes hang together in a life.
Studies of persona prompting find the same from the side of the prompt. GPT-3.5 and GPT-4 portraits of racial groups contained more stereotypes than portraits written by people given the same prompts (Cheng and colleagues). Across four models and 3,200 participants and 16 identities, models misportrayed groups and flattened them, and prompting with an identity essentialised it (Wang and colleagues). These studies are about demographic groups. The mechanism, a model writing from a label and not from a life, is the one at work when the label is a famous name.
A model that scores taste has the same limits. Models used as evaluators score their own outputs higher than others’ even where people judge them of equal quality, and in fine-tuned models the better they recognise their own text the stronger the preference (Panickssery and colleagues). Scores from 56 language models, six reward models and four model judges tracked human ratings less closely on pairs of answers of similar quality and on answers where annotators disagreed (Jiang and colleagues). We found the same with captions: on the best captions a strong model agreed with the crowd as often as one random voter did (A model judge is one voter). A model’s score can screen out the weak. It cannot stand in for the person who has to live with the choice.
How much of taste is shared, and how much is only yours?
This is where Segal’s third pillar becomes a number, and the answer depends on what is being judged. For faces, private taste was found to be about as strong as shared taste (Hönekopp). For abstract art it was far stronger: 75% of taste, against 40% for faces (Leder and colleagues). Across five kinds of picture, preferences for faces and landscapes held a high share of shared taste, while exterior architecture, interior architecture and artworks showed strong individual differences. The authors’ reading is that what is shared is conserved in naturally occurring domains, which matter in much the same way to everyone, and what is individual lies in artifacts of human culture, which lack that uniform relevance (Vessel and colleagues).
The lab’s own recheck on 554 people rating art found 23% of rating variance shared, 44% stable and personal, and 33% noise. Eurovision jurors were close to that (22, 43, 35), and for jokes the shared part was 12%. On people’s ratings of answers from 21 language models to everyday questions it went the other way: a shared selector explained 20% of the variance in ratings, and a personal selector added 0.2% to it.
The two ends meet in image generation. When six users compared pictures, those refined toward a shared aesthetic score ranked below pictures left unrefined, and those refined toward one person’s own taste ranked above them (Maerten and colleagues; What a model can learn about one person). The personal model’s gain over the best shared scorer was small on average and largest on the pairs where people disagreed.
So idiosyncrasy can be measured, and it is largest in artifacts of human culture such as art and architecture. In art, most of what is stable in a person’s taste is theirs alone. In answers to everyday questions, what a personal selector added was close to nothing. A model sits closest to the shared taste, and where taste is shared, slop costs the least.
A test on Framework’s own ideas
Segal’s warning reaches any page that writes ideas for a named person. If the sign of a person’s taste is the idea most people would guess they propose, how much of what the page writes for that person is that idea? We fixed the question, the measures and the bar for changing the page before running anything.
The data are 24 runs of Framework’s Ideas page: three people, Steve Jobs, Oprah Winfrey and Dieter Rams, on four everyday briefs (a coffee shop, a 40th birthday, a rental living room, a gift for a dad), two runs each. Each run has five sets of five drafts. For each person and brief, one call listed the three ideas most people would guess that person would propose. A separate model, blind to the set a draft came from, then marked each draft 2 if its central move was one of those ideas or a small variation, 1 if it borrowed a well-known motif of the person but made a different move, and 0 otherwise, in two shuffled orders. The share of drafts at 2 is the number below.
| Set | What the writer is asked for | Drafts at 2 |
|---|---|---|
| Plain | the good idea many entrants would reach, no person named | 54% |
| Framed | what the person would most likely do | 65% |
| Stretch | the person’s way of reasoning, not their signature forms | 52% |
| Far | the person’s method meeting outside material | 34% |
| Farthest | breaking an assumption nearly every proposal shares | 37% |
With no person named, 54% of drafts already land on the expected move, so part of what most people would guess a person proposes is simply what most people would guess for the brief. Naming the person lifts it to 65% in Framed, the set that Segal’s definition describes, the person’s known signs put to work again. Stretch sits at 52%, no further from the expected move than a page with no person in it. Far and Farthest fall to about a third. The people differ. With no name at all, 72% of the drafts for the Steve Jobs briefs already matched the expected Jobs idea. For Dieter Rams it was 32% and naming him raised it to 60%.
The cost of leaving the expected move shows in the page’s own scores, which were read together for Plain, Stretch and Farthest. Fit to the brief falls from 79 to 71 to 59, and unusual rises from 21 to 27 to 58, each on a scale of 0 to 100.
We then tried two changes to the writing of Stretch on the same 24 runs. In the first, the writer states in advance the idea most people would guess the person would propose, then has to build one that is not that idea or a variation of it. The share at 2 fell from 52% to 40%, a drop of 12 points (95% interval 4 to 21). Unusual rose by 5 points and fit fell by 4.5. But a model asked to match one draft from each of the three people to their names was right 35% of the time with the change, 50% for Stretch as it is, and 67% for Framed, where chance is 33%. Each figure rests on 48 matches, so the intervals are wide (35% is 23 to 54, and 50% is 29 to 71). The bar we had set asked that this match rate fall by no more than 5 points. It fell by 15, so the page keeps its writing as it is.
In the second, each person’s stated axioms were regrouped under “what is theirs”, and the writer was told to build on those. A call that sorts the axioms into what is theirs and what most thoughtful people in the field would say as well rated 14 of the 15 statements as shared (Steve Jobs 5 of 6, Oprah Winfrey 4 of 4, Dieter Rams 5 of 5). The writer in this test was not given those sorted lists, so the set only regrouped the axioms, and it matched Stretch (58% against 52%, a difference of 5 points, interval −5 to 14). It does not show what building on what is only theirs would do, for these pages or for a framework built from a person’s own short answers.
One more result came after the plan and was not a planned test. The page scores each draft for how unusual it is against the usual answers to the brief, and that score hardly sees the person’s expected move. Across 600 drafts the correlation between the two is −0.17. Drafts at 2 averaged 29 on unusual and drafts at 0 averaged 35, so a draft can be the person’s cliché and still read as moderately unusual.
We then tested a score that is told what is expected. The page’s scorer reads a list of usual answers to the brief as the zero of its scale. We added the three ideas most people would guess the person proposes to that list and scored the same 600 drafts again, in the same order. It did not see them any better. The correlation with the expected mark stayed at −0.23, against −0.20 within runs before, and a control that added another person’s three ideas gave −0.22. The extra lines lowered every score by two or three points and did nothing else. The ladder from Plain to Stretch to Farthest and the scores for fit held. Whether a draft’s central move is one of the expected ideas is a different question from how unusual it is, and it is the marker’s question, which does see them.
The writers, the reference, the marker and the page’s scorer are one model family, so the marker may share the writers’ idea of what is expected. Nobody was asked whether the sets read as different.
What it means for a framework
A framework is a written account of how one person reasons. These results are from one page writing for three well-known people, so what follows is what they support for that page.
The worry holds for the set closest to the person. Framed, which writes what the person would do, leans on the expected move more than a page with no person in it. A framework that only restates a person’s known moves gives what Segal describes, the signs without the relation. Far and Farthest sit near a third, and they pay for it in fit, which matches the studies of music and science, where what pays is a conventional base with one step that has not been taken.
Leaving the expected move costs some of the person. A reader matched a draft to its person 67% of the time for Framed, 50% for Stretch and 35% for the version that was told to avoid the expected idea, against 33% by chance. In this test much of what makes a draft read as that person is the move people expect of them. That is Segal’s point that taste has to be seen by someone, in a number: a draft that avoids the signs is also harder to attribute. Our bar for the page was to cut the expected move without losing that, and the writing change did not meet it.
What is only theirs is what the page has to carry. A model writes toward the shared taste without being told, and with no person named it already lands on the expected move more than half the time. Where taste is shared, a model is close to right already. In art the stable part of a person’s taste is mostly their own, and that part is what a framework can add. A model rated 14 of the 15 stated axioms of the three frameworks tested as ones most thoughtful people in the field would also state, so what is only a person’s own is not in those statements.
A score is not the person. The page’s scores come from a separate model, scored blind, and a model judge is one voter. On the best captions it agreed with the crowd as often as a random voter did, and judges favour their own writing. The unusual score hardly sees what is expected of a person, and giving it that person’s expected ideas did not change that. A high number is a reason to look at an idea and not a verdict on it.
Sources
- Segal, Tasteslop, Nemesis Memos, posted May 5, 2026 (first on X), read November 18, 2026. The Greg Fong quotation is from this post.
- Jiang and colleagues, Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond), arXiv 2510.22954, version 2.
- Ma, Zhang, Ang and Chen, Tastes without distinction: silicon samples and the synthetic construction of tastes, arXiv 2606.30085, version 2.
- Anderson, Shah and Kreminski, Homogenization Effects of Large Language Models on Human Creative Ideation, Creativity and Cognition, 2024.
- Si, Yang and Hashimoto, Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers, arXiv 2409.04109, 2024.
- Wang, Pan, Cui, Tian and Huang, Understanding Machine Taste: Aesthetic Agreement Across Foundation Models in K-pop, arXiv 2610.17731, 2026.
- Berger and Le Mens, How adoption speed affects the abandonment of cultural tastes, Proceedings of the National Academy of Sciences 106(20), 2009.
- Berger and Heath, Where Consumers Diverge from Others: Identity Signaling and Product Domains, Journal of Consumer Research 34(2), 2007.
- Askin and Mauskapf, What Makes Popular Culture Popular? Product Features and Optimal Differentiation in Music, American Sociological Review 82(5), 2017.
- Hekkert, Snelders and van Wieringen, ‘Most advanced, yet acceptable’: Typicality and novelty as joint predictors of aesthetic preference in industrial design, British Journal of Psychology 94(1), 2003.
- Uzzi, Mukherjee, Stringer and Jones, Atypical Combinations and Scientific Impact, Science 342(6157), 2013.
- Salganik, Dodds and Watts, Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market, Science 311(5762), 2006.
- Chaney, Stewart and Engelhardt, How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility, arXiv 1710.11214 (RecSys 2018).
- Cheng, Durmus and Jurafsky, Marked Personas: Using Natural Language Prompts to Measure Stereotypes in Language Models, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, 2023.
- Wang, Morgenstern and Dickerson, Large language models that replace human participants can harmfully misportray and flatten identity groups, Nature Machine Intelligence 7(3), 2025.
- Panickssery, Bowman and Feng, LLM Evaluators Recognize and Favor Their Own Generations, arXiv 2404.13076, 2024.
- Hönekopp, Once more: Is beauty in the eye of the beholder? Relative contributions of private and shared taste to judgments of facial attractiveness, Journal of Experimental Psychology: Human Perception and Performance 32(2), 2006.
- Leder, Goller, Rigotti and Forster, Private and Shared Taste in Art and Face Appreciation, Frontiers in Human Neuroscience 10, 2016.
- Vessel, Maurer, Denker and Starr, Stronger shared taste for natural aesthetic domains than for artifacts of human culture, Cognition 179, 2018.
- Maerten, Verwiebe, Karthik, Prabhu and Wagemans, Personalizing Text-to-Image Generation to Individual Taste, arXiv 2604.07427, 2026.
- Countless Lab, Models write alike, not necessarily worse; A model judge is one voter; What a model can learn about one person.
- Countless Lab, tasteslop tests on Framework’s Ideas page, November 18, 2026: 24 runs, 186 and 96 new model calls in the two tests (73 and 48 more from cache), each plan written before its run.