Forum

Opt-Out Was Never a Design Strategy: Lessons From Sora 2’s First Week

A gray concrete block wall with a band of yellow and black diagonal hazard stripes along its base

On 30 September, OpenAI released Sora 2, a video model that generates synchronized sound along with images, and a new social app built around it. Within days the app was among the most talked-about launches of the year; CNBC reported that it passed a million downloads in under five days.

The launch also produced one of the fastest policy reversals in the short history of generative media. The way it unfolded is a case study in why consent cannot be added after shipping.

Two kinds of permission

Sora’s app treated two things very differently. For people’s faces and voices, it introduced “cameos”: a user records a verification video, creates a likeness and decides who may use it. That is opt-in by design, and it was presented as a core feature.

For copyrighted characters, the initial approach was looser. In the days after launch, the app filled with clips featuring well-known animated and film characters, and reports described a system in which rights holders were expected to ask for their characters to be excluded.

On 3 October, Sam Altman published an update on his blog promising to give rights holders “more granular control over generation of characters, similar to the opt-in model for likeness but with additional controls”, and said OpenAI would try sharing revenue with rights holders who want their characters used. On 6 October the Motion Picture Association went further, saying that it remains OpenAI’s responsibility, “not rightsholders'”, to prevent infringement on its service.

Why opt-out failed so quickly

The episode shows something designers know from every other field: defaults are decisions. A system where permission is assumed unless withdrawn puts the burden on the people with the least information and the least time. A small studio or an independent illustrator cannot monitor every new app for misuse of their characters.

The cameo feature proves that a better default was available. OpenAI built a careful consent flow for likeness, with verification and controls, because faces are obviously personal. Characters and artworks are personal to the people who made them too. Treating them as raw material until someone objects was a product choice, not a technical necessity.

What consent by design looks like

For designers building or commissioning AI products, the Sora week offers a short checklist.

  • Make permission the default state for anything identifiable: faces, voices, characters, signature styles.
  • Design the permission flow as carefully as the generation flow. If it is buried in settings, it is not consent.
  • Show provenance in the product. Viewers should be able to tell that a clip is generated, and from what.
  • Plan for revocation. People change their minds; systems should make withdrawal real, not symbolic.
  • Treat rights holders as users. Give them dashboards, controls and records, not only takedown forms.

What it means for creators using these tools

For artists and designers using video generators, the episode is a reminder to stay well clear of other people’s characters, however easy the tool makes it. Policies are changing week by week, and what an app allows today may be what a client is liable for tomorrow. Original characters, licensed assets and clearly documented processes remain the safest ground.

Competition entrants face the same expectation. The International AI Award rules require entrants to own or have permission for the work they submit, and they weigh ethical responsibility as part of judging. A striking video built on someone else’s intellectual property does not meet that bar, no matter how good it looks.

The broader lesson

Sora 2 is an impressive technical release, and its cameo system is a thoughtful piece of consent design. The trouble came from applying that thoughtfulness to one category of personal material and not another. The next wave of creative AI products will be judged less on what they can generate than on how well they respect the people whose work and identity they touch. That is a design problem, and it belongs at the start of the process.