OpenAI wants your face on its feed

Copyright opts out, likeness opts in. The TikTok clone constrained by compute reality

OpenAI released Sora 2 on September 30 alongside a TikTok-style app and a promise of “responsible” design. The company’s own safety memo touts watermarking, teen protections, and consent-first “cameos.” Yet on copyright, OpenAI is taking an opt-out approach that’s already prompted Disney to block its catalog, according to Reuters and the Wall Street Journal. The tension is obvious: your likeness requires a yes; famous IP is in by default unless rights holders say no.

What’s actually new

Sora 2 generates 10-second clips with synchronized audio and more realistic physics. In demos, a missed basketball shot finally bounces instead of teleporting through the hoop. Lip-synced speech lands where it should. The headline feature is Cameo: record a brief video and audio sample once, then let others—if you grant permission—drop your likeness into scenes. In press briefings, OpenAI employees described cameos as making the cameo subject a “co-owner” of the final video, with the right to delete or revoke at any time. Public figures are off-limits unless they voluntarily upload a cameo.

Access starts in the U.S. and Canada on iOS, invite-only, with four extra invites per user. An Android version is in development, but there’s no date. Sora 2 is free with usage limits; ChatGPT Pro users get a higher-quality “Sora 2 Pro” via the web and, later, the app. An API is promised “in the coming weeks.”

The social bet, not just the model

This isn’t merely a model release; it’s a network play. The app ships a vertical feed, Remix for trends, likes, comments, and DMs. OpenAI says the ranking favors people you know and content likely to inspire creation over raw time-spent. Teen accounts default to scroll limits and stronger privacy, with parental controls through ChatGPT. Every output carries visible watermarks and C2PA metadata; OpenAI also claims internal tools can trace Sora videos back to source.

The distinction the company draws is stark: a consent gate for faces, provenance for content, and safety classifiers across video frames and audio transcripts. That’s the theory. Feeds have their own gravity.

Copyright vs. consent

OpenAI is running two playbooks at once. For likeness, consent is a hard stop: no cameo, no generation. For copyright, the reported policy runs the other way: include copyrighted characters or styles unless owners explicitly opt out. The Wall Street Journal adds that OpenAI won’t accept blanket, portfolio-wide exclusions; studios are being routed to a takedown-style process. That asymmetry reveals priorities: minimize legal risk on deepfakes of real people, keep the content firehose open for everything else.

Studios are already testing those boundaries. Disney, per Reuters, has invoked the opt-out. Others had been alerted in advance, receiving instructions on how to report violations. Expect more friction here than in any watermarking debate, because this is distribution, not training.

The safety promise meets the slop problem

OpenAI stresses “layered defenses” and says it has tightened policies relative to image generation, especially now that video includes motion and sound. Cameos use liveness checks and audio challenges. Music prompts are scanned to deter imitations of living artists. Human review backs up automated filters.

Skeptics focus less on filters than on incentives. As the AP framed it, AI video is compelling precisely because it is implausible yet realistic. If feeds reward attention, and AI content is cheaper and faster to produce than human video, the risk isn’t one bad clip slipping through. It’s a flood. Meta rolled out its own AI video feed days earlier. Google is wiring Veo into YouTube. The common pattern isn’t hard to spot.

TikTok’s opening, OpenAI’s pitch

Timing helps. TikTok still faces U.S. scrutiny over ownership and data, and Washington mood music hasn’t softened. OpenAI arrives with a domestic stack, restraint language about consent and wellbeing, and a promise to bias the feed toward friends. CEO Sam Altman adds a test: if most users don’t feel their lives are better after six months, OpenAI will “make significant changes—or discontinue the service.” That’s a high bar, and it creates a paper trail. But it also kicks the can: who measures “better,” and when?

The compute ceiling

There’s a practical limit that marketing can’t ignore. OpenAI says Sora launches with “generous limits,” and that its only current plan for excess demand is letting people pay to generate an extra video. Translation: compute is the choke point. Video inference is expensive; virality multiplies costs linearly. A social product needs network effects; the infrastructure needs throttles. Those pressures don’t align.

The rollout explains the reality. Slow access ramp. Short clips. A higher-quality tier kept to paying users on the web, for now. If usage spikes, limits tighten; if growth lags, the pitch weakens. That’s a tough tightrope, even before the copyright fights begin.

The bottom line

Sora 2 is a technical step forward and a strategic escalation. OpenAI is late to public video generation but early to a social layer that centers identity and remixing. The consent-for-faces and opt-out-for-IP split is the real story—and the real risk. If the app becomes a hit, it will be because the feed is fun, not because the watermark is visible. If it misfires, compute costs and rights friction will be the reasons.

Why this matters

• The incentive flip: Cheap AI video plus engagement-driven feeds tilts platforms toward volume, not craft—regardless of watermarking or safety rhetoric.
• Cost vs. growth: OpenAI must scale a social app while rationing compute, all as studios test the limits of opt-out copyright. That tension will shape what Sora can actually be.

OpenAI Sora 2: Permission for faces, opt-out for copyright

OpenAI’s Sora 2 demands permission for your face but lets Spider-Man appear unless studios opt out. The split policy arrives as year-end conversion deadlines loom and competitors launch rival video feeds—testing fair use through product, not precedent.

Implicator.aiMaria Garcia