Beyond slop, taste, and AI moral panic: What social science tells us

What it claims

The essay's practical argument is about how firms should restructure junior roles, not about whether to keep them. "Taste" is split into separable mechanisms, and each one is tested against what LLMs change:

• Taste as consistent, discoverable criteria (e.g., Bauhaus principles, a literary agent's acquisition preferences) is a pattern-recognition problem. The author expects models to reproduce it.
• Prediction of hits is not taste. AI increases the number of candidates that clear the "good enough" bar, so quality decides even less of what succeeds. Candidates that already have an audience gain the advantage.
• Brokerage is split in two. LLMs already do the "cheap half": they surface what is happening in an adjacent field to anyone who asks the right question. The defensible human edge is participating in worlds that were never written down anywhere a model could ingest, such as subcultures, informal practice, and arguments made in a room and never recorded.
• Apprenticeship is the one case where "protect it" holds. What trained judgment was never the grunt work itself. It was "someone catches what you got wrong, repeatedly, until you stop getting it wrong." So the recommended move is to redirect entry-level roles toward checking and correcting AI drafts under senior supervision, rather than eliminating them. The author says this keeps the training scaffolding while capturing some of the cost savings. The verification framing is credited to economist Christian Catalini, who argues that the human role is checking that LLM output hasn't drifted.
• The essay flags a tragedy-of-the-commons risk: if every firm cuts junior roles, senior judgment becomes scarce across the whole industry. It also concedes that firms investing in apprenticeships may lose in the short term to competitors who automate aggressively.

Evidence

All evidence is secondary or anecdotal. There is no original data.

• Watts & Salganik music lab (early 2000s): thousands of participants downloaded songs by unknown bands across eight "worlds."
• When download counts were hidden, downloads were evenly distributed.
• When counts were visible or manipulated (for example, showing a song at #1 when it wasn't), the same song could be a runaway hit in one world and perform poorly in another.
• The author's reading: experts reliably filter out obvious failures but cannot pick breakouts among the strong candidates that remain.
• Burt: decades of mapping who communicates with whom inside organizations, compared against promotion, compensation, and manager evaluations. The most innovative employees sat in "structural holes" between disconnected groups.
• Author anecdote: as a junior editor at Princeton University Press, the author learned by reviewing a senior editor's markups while photocopying them, then having their own edits reviewed in turn.

Limits

• This is an opinion essay. No study measures whether supervised verification of AI output actually trains judgment as well as drafting from scratch. The author even hedges that drafting "can still be invaluable."
• The author acknowledges that Burt's findings concern organizational careers and competitive strategy, not cultural products. Treating taste as brokerage is offered as a "perhaps."
• The page carries a16z's standard disclaimer: the views are the individual's, not the firm's, and the content is not advice.