Imagine scrolling a vertical feed that looks and feels like TikTok — but almost everything you watch is made by AI, and your face (and voice) can be swapped into other people’s clips with a few taps. That’s the product OpenAI quietly rolled out this week: Sora, a mobile app paired with a much-improved video-and-audio model called Sora 2. The company is pitching it as a creative tool — the kind of moment that could make video generation feel as obvious and useful as ChatGPT did for text — but it also raises the obvious questions about consent, deepfakes and what social media looks like when cloning faces becomes easy.

What is Sora (and what is Sora 2)?

Sora is two things at once. It’s the name of OpenAI’s updated video-and-audio generator — Sora 2 — and the name of an iPhone app that stitches that model into a short-form social experience. The model adds synchronized audio, more realistic physics (things like accurate motion and believable interaction with the environment), and much tighter control over how a scene unfolds. OpenAI’s demo reels lean hard on things that used to be awkward in AI video — beach volleyball, skateboard tricks, bodies moving together — and show clips that look, at a glance, convincingly real.

The app itself is invite-only for now and is available only in the United States and Canada. When new people get access, they receive a small handful of invites to share with friends; there’s no Android build yet. Internally, OpenAI employees have described the debut as a potential “ChatGPT moment for video generation,” the kind of product that could make a technology suddenly feel obvious to millions.

The hook: cameos and remix culture

Sora’s social twist is a feature called Cameos. You record short reference material — a little head movement and a voice sample — and you can then give other Sora users permission to “cast” you in AI-generated clips. If you allow it, your likeness becomes a reusable asset other people can slot into scenes: a cameo can be placed into a news desk sketch, a movie-style action beat, or a surreal clip where you’re suddenly doing backflips on a rooftop. OpenAI says the person whose likeness is used is treated as a “co-owner” of the resulting video, and they can always delete the video or revoke access for others.

There’s also a Remix button for anything on your feed: tap it and Sora lets you tweak someone else’s clip, swap characters, change the tone or extend the scene — assuming the original creator set their cameo permissions to allow that. It’s a deliberate attempt to graft the remix-and-duet culture of short-form video onto generative AI.

Limits, guardrails and a confusing detail

OpenAI says it’s put guardrails in place: you can’t generate explicit porn, extremist content, or other “extreme” material, and the app blocks depictions of public figures unless the public figure uploaded a cameo and consented to be used. OpenAI also plans to embed watermarks and metadata so that AI-generated clips are traceable as synthetic — a standard the company and other industry players have been promoting as a mitigation step.

One small — but important — detail is inconsistent in reporting: some outlets and internal docs describe Sora clips as being 10 seconds long, while OpenAI’s existing Sora documentation (and tools released earlier) reference up to 20 seconds for generated clips. That’s not just trivia: format constraints shape creativity, moderation effort, and how easily content spreads. Expect the length rules to become clearer as the app moves out of its initial invite phase.

Why people are excited — and why others aren’t

From a maker’s perspective, Sora 2 is a leap. It pairs believable motion with synchronized voice and sound effects, making short scenes far more convincing than the first generation of AI clips. For creators, the promise is obvious: imagine making a sketch where your friend plays half the cast, or a travel clip in which your likeness joins a stunt that would be expensive or dangerous in real life. OpenAI even hinted that ChatGPT Pro users may get early access to a more capable Sora 2 model and that the company plans to expose features via API later.

On the other hand, this is exactly the technology that has defenders and critics talking past each other. If it’s easy to make a clip of someone saying something they never said, the permission model — and the practicalities of policing a huge feed of user-generated AI clips — becomes a public-safety issue. Journalists and researchers have flagged failures in content moderation and the persistent risk that bad actors could try to game opt-outs or spread disinformation. OpenAI emphasizes consent and technical markers, but technical fixes rarely solve social problems on their own.

The legal and copyright thicket

Sora also touches on copyright and rights-holder questions. Early reporting suggests OpenAI will allow generation referencing copyrighted characters in some cases unless rights-holders opt out, and that the company is aware the industry will push back: film and music companies have already been lobbying for clearer guardrails and commercial terms around AI use. If a hit TV show suddenly appears in thousands of AI remixes — some benign, some not — the legal fights will follow fast.

The bottom line

Sora is a tidy experiment in wrapping a powerful generative model in social features that make it immediately usable. That design makes sense — tools become meaningful when they fit into existing social habits — but it also raises questions that will need social, legal and technological answers, not just product patches. For now, Sora is an invite-only glimpse at what short-form social media could look like when faces, voices and scenes are no longer fixed in pixels but are playthings of generative models. Whether the cameo-and-consent approach will be enough to keep the fun from bleeding into harm is the big question everyone watching this rollout will be trying to answer.