In the early 18th century, if you lived in London and you wanted to slander an opponent or rival, you could head to Grub Street, a slum just north of the city named for a garbage ditch.
Amid the narrow lanes teeming with hack writers and failed scribblers, you might find someone like Samuel Boyse — a drunken Irish poet who had fled Edinburgh because of his debts. His specialty, doggerel. He would happily compose you some for a shilling or two. If you had deeper pockets, he could even gin up a pamphlet that really laid into your target.
Boyse lived in a garret. He struggled with alcoholism. Once he found himself in such dire straits that he was forced to pawn the shirt off his back to buy the paper necessary for his craft. (Ah, the life of the writer!) Boyse’s situation was not unique — a writer as highly regarded as Samuel Johnson is alleged to have spent some time on Grub Street — although Boyse was one of the more prolific practitioners of his craft.
Grub Street produced mountains of what we today might call “slop”: single verses, pamphlets, periodicals, newspapers — all for sale in both its production and consumption. Inevitably, perhaps, this drew comparison of such writers to strumpets, and critics equated writing for pamphleteers to whoring. No respectable person would take part in such a dreadful enterprise … except that lots of respectable people did, even while decrying it to their equally titillated friends.
What I’m saying is, slop is hardly a new phenomenon. Neither is the moral panic around it. AI-generated content is the latest iteration of an age-old trend: Whenever technology dramatically lowers the cost of producing media, it unleashes a flood of cheap, derivative work — and a chorus of critics convinced that culture has reached a new low. After Grub Street came the penny press, then pulp fiction, television (a “vast wasteland” if you’re feeling fancy, or, more colloquially, the boob tube), followed by blogging, social media, and now AI.
Each time, the cost of execution falls, and the masses jump in. Each time, criticism comes hot and fast. And each time, the answer from the self-proclaimed cognoscenti is obvious: Claims to taste become their refuge, and taste achieves apotheosis. While AI may differ in degree and perhaps even in kind, the cultural response to falling production costs has recognizable historical precedents.
But this immediately raises the question many are debating: What exactly is taste? If taste consists of stable criteria that experts consistently apply, then a modern AI can almost certainly parrot taste. A model trained on enough examples of Bauhaus design can infer the principles that make an object feel Bauhaus. A model trained on a literary agent’s acquisitions can approximate that agent’s editorial preferences. Taste as the consistent application of discoverable rules is precisely the kind of pattern-recognition problem modern AI excels at.
If, on the other hand, “taste” isn’t reducible to stable criteria, then we’ve been using the word far too casually. The idea that taste simply means knowing what’s “good” borders on the tautological, while reducing it to personal preference — I like this; I don’t like that — is too subjective to explain why some people repeatedly identify promising ideas, products, or creators before everyone else.
So what are we really talking about when we say someone has taste? Prediction? Judgment? Social position? Experience? Luck? “Taste” is sometimes all of these: it’s a folk category into which we’ve collapsed separate mechanisms. AI forces us to disaggregate them. Drawing on social science can help, not in giving us a single theory of taste but in teasing disparate ideas apart. Once we identify the separate strands, the question shifts from whether machines have taste to which mechanisms that produce good human judgment remain scarce when production becomes cheap.
Success is socially contingent
One body of research that may help illuminates how songs become hits.
In the early 2000s, Columbia University sociologists Duncan Watts and Matt Salganik, together with a team of researchers, created an artificial online “music lab.” Thousands of participants were invited to listen to and download songs by unknown bands. The researchers divided listeners into eight separate groups, each with a different condition of how the songs were presented. In some, participants could see how many times each song had already been downloaded; in others, they could not. The team also manipulated the information in some “worlds” (what they called their experimental conditions), showing, for instance, a particular song at #1 when it was not.
When social information was hidden, downloads were evenly distributed. But when people could see what others were choosing, outcomes diverged. In these manipulated conditions, a song that became a runaway hit in one world performed poorly in another.
The larger lesson is that once quality clears a basic threshold, success is heavily shaped by social influence and path dependence. This makes it almost impossible, even for informed observers, to predict in advance which products will rise to the very top. Put another way, experts in a domain — TV executives, for example — are pretty good at filtering out obvious failures before they happen. But they cannot reliably predict breakout successes among the remaining strong candidates. Or as Salganik has said, the success of Harry Potter was a fluke.
Success is not necessarily dictated by quality alone. Therefore taste cannot simply mean predicting success.
Taste is brokerage
Another possibility is that what we call taste isn’t prediction at all. Maybe it’s the function of a particular social position: standing somewhere that gives you a unique vantage point to see how the world could fit together, or how ideas and products could translate from one arena to another.
This is why some of the most innovative employees in companies tend to be the ones most connected to different groups — or that’s what Ron Burt, another sociologist, found. Burt calls this “brokerage.” Brokerage isn’t the clairvoyance necessary to pick winners but a structural position that generates novel recombinations.
Burt’s conclusions grew out of decades of research mapping social networks inside organizations. Rather than focusing on individual brilliance or expertise, he asked how people’s positions within those networks affected their performance. First, Burt reconstructed who communicated with whom. He then compared those patterns with measures like promotion, compensation, and managers’ evaluations. He found that the most innovative employees were rarely embedded in a single, tightly knit community. Instead, innovators (might we call them those with taste?) occupied “structural holes”: the gaps between otherwise disconnected groups.
Because they moved in different circles, brokers encountered information that others didn’t and, even more importantly, saw problems from multiple perspectives. Their position allowed them to recombine ideas that no one network could have produced on its own. Here, innovation isn’t a matter of exceptional intelligence but of extraordinary position.
This suggests that taste may not be an aesthetic faculty at all but rather a structural advantage. We often imagine taste as residing inside someone’s head — as a refined sensibility or intuitive aesthetic judgment. Burt suggests that what looks like taste may instead arise from where someone sits in a social network. The people who seem to have an uncanny feel for what’s interesting, original, or promising are often those who spend their lives moving between otherwise disconnected social worlds. Their “taste” emerges not from mysterious intuition but from repeated exposure to combinations that no one else is able to see.
An LLM trained on, well, everything has, in a sense, access to all the networks simultaneously — so if structural holes were purely about information access, AI might excel. But the broker’s advantage comes from living inside multiple social worlds, with the embodied, relational, contextual knowledge that comes from their network position. We might say the broker’s advantage is tacit. Brokers don’t merely possess information from multiple communities; they actively participate in multiple communities. They accumulate tacit knowledge, trust, and experience, and respond to novel incentives in ways that LLMs simply can’t. You can’t simulate being an outsider-insider from training data alone. That’s a defensible human edge — for now, at least.
Granted, Burt’s work is about organizational careers and competitive strategy, not cultural products. But perhaps what we call taste in other domains may be structurally equivalent to Burt’s brokerage — the advantage of the person who’s genuinely of multiple worlds.
Judgment is trained
So where can we find human judgment? We create it.
LLMs can perform the tasks often associated with entry-level roles: drawing up contracts like a junior law associate, performing basic financial analysis like a junior consultant, coding like an engineer, and so on. It’s been suggested that this may allow organizations to slash costs and increase output by replacing those jobs with an LLM. It also grants expertise to the common person: No longer do you have to rely on Big Law and its hourly rates for basic (or sometimes even complex) tasks.
The human role in all of this, argues economist Christian Catalini, is verification: making sure the LLM’s output didn’t drift too far. Humans check. Humans correct. Humans employ taste?
But this opens up the question of where humans learn taste. Good (human) evaluators develop their abilities with apprenticeship, feedback, and exposure.
Back in the day, when I was a junior editor at Princeton University Press, I used to have to photocopy the book manuscripts that my boss had edited with his green fountain pen before I mailed them to the author. We needed a copy for reference and also in case the original was lost. While the copy machine churned away, I took the time to review his edits, read his queries, understand both his green-inked penmanship (which was a nontransferable job-specific skill that hasn’t helped later in my career but was, at the time, crucial, especially because few others — including the authors — could decipher his writing) and, more importantly, his editorial style. This informed my own editing (which my mentor reviewed in turn) and created the foundation from which I still work. (This experience also points to another often underappreciated truism: Jobs can be messy.)
In other words: At least one element of what we call taste — judgment — is a kind of discrimination one accumulates and refines over time. It is not innate. If we are to continue to create humans with taste, we have to create spaces where they can develop it.
The obvious objection is that firms investing in apprenticeships may lose in the short term to competitors who aggressively automate. Maybe! But the tradeoff resembles overcutting a forest: Chopping down every sapling may improve one quarter’s balance sheet, but it also ensures that you’ll have no mature trees later on. And of course it’s incumbent on anyone looking for a job to make sure they train themselves in the absence of those roles.
Practice complicates things
Pushing each of these three findings toward some kind of practice, none of them break — but I don’t think any of them means quite what the obvious reading suggests, either.
Take the socially contingent hit first. Salganik’s conclusion doesn’t change: Success still isn’t dictated by quality alone, so taste still can’t simply mean predicting success. What AI adds is volume — more candidates clearing the “good enough” bar than any one curator can sort through. In other words, TV execs still filter the same way. It just means there are a lot more shots on goal, and quality alone determines even less of which shots land. Success depends on luck — or, more precisely, social or network position — in this environment more than ever. Whichever candidates already have an audience are the ones most likely to catch a wave before the field crowds them out.
Brokerage holds up better. An LLM can already do the cheap half of brokerage — surfacing what’s happening in an adjacent field to anyone who asks, which used to require actually knowing someone in that field. (Someone does still need to know to ask the right questions.) The structural hole that really matters isn’t “knows two worlds” in the informational sense. It’s inhabiting worlds that haven’t been written down somewhere an LLM could ingest them: subcultures, informal practice, arguments that happen in a room and never make it onto the record, and so on. If brokerage keeps paying off, it will be because someone sought out exactly that kind of undocumented and undocumentable territory.
Apprenticeship is the one where the standard advice — protect it — holds up. The apprenticeship model assumed junior people learned by doing grunt work under a senior person’s mentorship, the way I learned editing by photocopying my mentor’s marked-up pages and then having him review my own work, once I was allowed to start editing on my own. If AI now does that grunt work, the easy conclusion is that the apprenticeship disappears along with the job that trained it.
This may be backward: The process that mattered was “someone catches what you got wrong, repeatedly, until you stop getting it wrong.” That can survive if the junior person’s job becomes checking and correcting the AI’s draft under supervision, rather than drafting something from scratch themselves (although I’d argue that can still be invaluable). What changes isn’t whether firms keep junior humans around but rather what those humans spend their day doing. Firms that simply dispense with entry-level roles will tear out a lot of scaffolding with it. Firms that redirect entry-level roles toward supervised verification can maintain their scaffolding while still capturing some savings. (There is a Tragedy of the Commons possibility here, though, where all the firms dispense with junior jobs and simply lack senior judgment across the board later on, like the timber company that harvests all the young trees.)
***
Boyse pawned his shirt for paper and died on Grub Street. Samuel Johnson, who allegedly walked the same streets, didn’t stay. Not because he had better taste but perhaps because he had better scaffolding.
In the debate over AI, “taste” has become a kind of totem, invoked when someone wants to identify something humans can do but machines can’t. But taste doesn’t have to be mysterious. It is what good judgment looks like after it has been shaped by environments that reward experimentation, networks that expose us to novel combinations, and institutions that cultivate discrimination.
In the age of AI, the scarce resource isn’t taste. What threatens to become scarce is the social infrastructure that produces the people we describe as having it.
***
Acknowledgments: Many thanks to Drew Coffman, Sonal Chokshi, Steph Zinn, and especially Robert Hackett for helpful comments and edits.
***
The views expressed here are those of the individual AH Capital Management, L.L.C. (“a16z”) personnel quoted and are not the views of a16z or its affiliates. Certain information contained in here has been obtained from third-party sources, including from portfolio companies of funds managed by a16z. While taken from sources believed to be reliable, a16z has not independently verified such information and makes no representations about the current or enduring accuracy of the information or its appropriateness for a given situation. In addition, this content may include third-party advertisements; a16z has not reviewed such advertisements and does not endorse any advertising content contained therein.
This content is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. You should consult your own advisers as to those matters. References to any securities or digital assets are for illustrative purposes only, and do not constitute an investment recommendation or offer to provide investment advisory services. Furthermore, this content is not directed at nor intended for use by any investors or prospective investors, and may not under any circumstances be relied upon when making a decision to invest in any fund managed by a16z. (An offering to invest in an a16z fund will be made only by the private placement memorandum, subscription agreement, and other relevant documentation of any such fund and should be read in their entirety.) Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z, and there can be no assurance that the investments will be profitable or that other investments made in the future will have similar characteristics or results. A list of investments made by funds managed by Andreessen Horowitz (excluding investments for which the issuer has not provided permission for a16z to disclose publicly as well as unannounced investments in publicly traded digital assets) is available at https://a16z.com/investments/.
Charts and graphs provided within are for informational purposes solely and should not be relied upon when making any investment decision. Past performance is not indicative of future results. The content speaks only as of the date indicated. Any projections, estimates, forecasts, targets, prospects, and/or opinions expressed in these materials are subject to change without notice and may differ or be contrary to opinions expressed by others. Please see https://a16z.com/disclosures for additional important information.