In 1942 the musicians' union banned its members from all commercial recording — its answer to "canned music" destroying live work. In 1966 The Beatles released Revolver, an album built from studio techniques that could not be performed live at all. The medium the union tried to stop had become an instrument in its own right. Patreon co-founder Jack Conte's argument, made at SXSW in March 2026, is that this arc has now run four times in 130 years — and that "AI slop" is simply what the newest medium looks like while it's still copying the formats of the old one.
Conte assembles the same arc across film, recorded music, sound cinema, and synthesizers. Phase one imitates the medium that came before — 1890s film crews pointed static cameras at stage plays, and critics correctly called the result a grey copy of theatre. Phase two begins when someone discovers what the medium does that nothing before it could. For film, Conte dates that moment to 1902: Georges Méliès using cuts and effects in A Trip to the Moon to make something theatre could not stage.
The new medium reproduces the old medium's formats, and is judged against them — and usually judged worse. Cheaper theatre. Canned music. Slop.
Necessary — it's how the medium gets learned.
Someone stops copying and builds around what the medium alone can do. The cut. The studio album. The synth line no instrument could play.
This is where the value was, every time.
Slop is the output of a new medium being used to copy the formats of the old one. That turns it from an aesthetic complaint into a timing signal: slop-dominance means the medium is in its imitation phase — which says nothing about its ceiling and everything about where it sits on the arc.
The dominant deployment pattern is substitution: the same workflow, the same artifacts, the same org chart — with a model producing the document a person used to produce. Cheaper output in the old format, judged against the old format, and usually judged a grey copy. That is the enterprise imitation phase, and most organisations are in it.
It means designing the operation around what the medium natively does — parallelism, persistence, tool use, and a marginal cost per unit of work approaching the cost of inference — rather than around the shape of the pre-AI workflow. The harness model is one candidate for that native form: a single operator commanding a fleet of agents, with human specialists in the loop for judgment, instead of headcount-shaped teams using AI as a faster typewriter.
Featured: The Builder Is the Operator →An organisation cannot skip the imitation phase — it is how teams learn the medium, and funding it as a learning stage is rational. The risk is mistaking it for the destination: automating the old workflow so thoroughly, and so expensively, that the native redesign never gets budget. Imitation locked in is just rented efficiency on yesterday's process.
Conte's second argument is a governance item: AI vendors are signing licensing agreements with large rights-holders while ingesting individual creators' work without consent or compensation, defended as fair use — his slogan for the fix is consent, credit, and compensation.
The documented record cuts both ways, which is exactly the point. Major deals are real — Vox Media licensed its archive to OpenAI (2024), and Warner Music reached agreements with AI music firms after litigation. And so are the lawsuits — Disney's most prominent public move has been suing Midjourney, and voice professionals filed class actions against ElevenLabs and others in 2026. Rights-holders are licensing and litigating simultaneously, and the fair-use question is unresolved in most jurisdictions. For any company deploying generative output at scale, that makes training-data provenance a live legal and reputational exposure — on the same register as data residency and POPIA: an input-supply-chain question, answered before procurement, not after a claim.
Deeper: Intellectual property — training data & the law →Is this substitution inside the old workflow, or a redesign around the medium's native properties? Substitution is fine — as a funded learning stage. The question to ask in every review: is anyone budgeting the native move, or has the stage-play film become the strategy?
Every generative deployment gets a named owner for training-data and output-provenance exposure, with vendor licensing posture as a selection criterion — before procurement, the way residency already is.
Slop in a competitor's output is timing intelligence: it marks where they sit on the arc — still in the imitation phase — and therefore how much room remains for the first native move in your sector. The pattern has run four times in 130 years. The imitation phase ends; the question is only who ends it.
Where durable ownership sits once the imitation phase closes — rent the frontier model, own the layers that compound.
Read briefing →The harness model in full — the candidate "native form" this briefing points at: one operator, a fleet of agents, specialists in the loop.
Read briefing →Slop isn't the ceiling of the medium — it's the timestamp on the phase. Fund imitation as tuition, budget the native move, own your provenance, and read your competitors' slop for what it is: a market that hasn't had its Méliès moment yet.