Sora adds Cameo guardrails to stop deepfake misuse

Letting anyone on the internet generate videos of your digital self is a wild premise. Waking up to find your avatar delivering political propaganda or swearing in a clip is even worse.

Following a chaotic debut week where Sora shot to the top of the App Store despite being invite-only, OpenAI is rushing to add guardrails. The app became an overnight sensation largely thanks to "Cameos," which let users upload biometric data to place their likeness into AI videos. But the chaos followed just as fast, with users flooding social feeds with SpongeBob, Pikachu, and deepfakes of Sam Altman mocking OpenAI's own copyright rules.

Now, OpenAI is hitting the brakes on both likeness abuse and intellectual property scraping.

Setting ground rules for your digital twin

OpenAI's Bill Peebles announced that Sora users can now set custom rules for their Cameos. If you want to make your face available to others without becoming an unwilling meme, you can now add natural-language restrictions under your cameo preferences.

For instance, you can tell Sora to never put you in videos involving political commentary, or explicitly ban specific words from your avatar's mouth. Peebles noted the team is working on making these restrictions more robust, with more fine-grained controls coming down the road.

Alongside likeness controls, OpenAI is overhauling safety across the board. Peebles said the team is making Sora's watermark clearer and more visible, while patching loopholes to catch generations that slip past the filters. He admitted overmoderation is frustrating, but argued that a conservative approach is necessary while society adjusts to hyperrealistic video.

There was also a bizarre administrative fix: previously, deleting your Sora account would accidentally wipe out your ChatGPT account as well. That bug has finally been squashed.

Retreating from the copyright wild west

Perhaps the biggest shift comes from Sam Altman himself. Before launch, The Wall Street Journal reported that OpenAI told Hollywood studios they had to explicitly opt out if they didn't want their IP showing up in Sora generations.

Altman is now walking that back. In a blog post, he signaled that Sora will adopt an opt-in model for copyrighted characters, giving studios granular control over how their characters appear, or allowing them to block Sora entirely.

Altman framed this as an embrace of "interactive fan fiction," noting that many rights holders want to participate on their own terms. Crucially, he also acknowledged that OpenAI has to make money from video generation, hinting at revenue-sharing models with IP holders rather than just charging users for compute during peak hours.

Why it matters

For everyday creators, natural-language Cameo restrictions are the only reason anyone with a reputation to protect would consider using this feature. Without prompt-level guardrails against defamation or abuse, public likeness sharing is a nonstarter. It makes Sora's avatars significantly safer for regular people to play with.

For the media industry, OpenAI's retreat from an opt-out stance to an opt-in model is a massive concession. Hollywood wasn't going to tolerate Pikachu roping Sam Altman into deepfakes forever. Proposing a revenue-sharing fan-fiction model is a clever olive branch that could turn bitter IP lawsuits into licensing revenue.

We think OpenAI acted quickly here, but the execution will be messy. Relying on prompt moderation to enforce Cameo restrictions and IP bans will inevitably produce edge cases and jailbreaks. Until Sora proves its filters can reliably block copyrighted characters and abusive likeness deepfakes, copyright owners and privacy advocates will remain skeptical.