How do you make a video with Sora 2 responsibly?

Sora 2 is OpenAI's newest model for generating video with sound. Released on September 30, it produces short clips with dialogue and sound effects, follows instructions that span several shots, and handles physical motion better than earlier video models. Most people reach it through the new Sora app for iPhone, which is invite-only and rolling out first in the United States and Canada.

The realism is the point, and it is also the risk. A believable clip of a real person saying something they never said can harm them. So making a video responsibly has two parts: craft a good clip, and respect three things that OpenAI has built into the product: consent for likeness, provenance labels, and the content rules.

The rules you are working within

OpenAI's launch safety post sets out the main protections:

• Provenance. Every Sora video carries a visible watermark and embedded C2PA metadata, an industry standard for recording where a file came from. OpenAI also keeps internal tools that can trace a video back to Sora.
• Likeness by consent. Real people appear through cameos, a feature where a person records themselves once and then decides who may use their likeness. Public figures are blocked unless they have made a cameo themselves.
• Content filters. Prompts and finished clips are checked, including several frames and the audio transcript, to block content such as sexual material or self-harm promotion. Music that copies the style of living artists, or reproduces existing songs, is also blocked.
• Your control over sharing. A video reaches the public feed only if you choose to post it, and you can remove it later.

Removing the watermark or stripping the metadata defeats the purpose of these signals. Leave them in place, and when you repost a clip elsewhere, say plainly that it is AI-generated.

A worked example: a 10-second clip for a garden club

Suppose you run a community garden and want a short, playful clip to announce spring planting day. No real person needs to appear, which keeps consent simple.

Two shots, about 10 seconds total, bright morning light, gentle handheld camera.

Shot 1: Close-up of a cartoon-style earthworm in a tiny straw hat popping out
of dark soil between rows of seedlings. Birds chirping, a soft breeze.

Shot 2: Wide view of a small community garden with wooden raised beds and a
hand-painted sign that reads "Planting Day". The earthworm waves.
Cheerful whistled melody, no lyrics.

No people, no brand names or logos.

Then work through these steps:

1. Generate and watch the draft in full, with the sound on. Check for garbled sign text, odd motion and anything you would not want attached to your group's name.
2. Revise one thing at a time, for example "make the sign text larger and clearer", rather than rewriting the whole prompt.
3. Decide whether it should be public at all. Making a clip does not publish it; it reaches the Sora feed only if you choose to post it.
4. Label it when you share it outside the app: "Made with Sora, an AI video tool."

The prompt does several responsible things on purpose: it describes an invented character instead of a person, asks for original music rather than a famous song, and excludes logos that belong to someone else.

Checks before you share

A short checklist covers most problems. Could a viewer mistake this for real footage of a real event? Does it show anyone who has not agreed to appear? Does it borrow someone else's characters, music or brand? Would you be comfortable if the person or group shown saw it? If any answer worries you, do not share it.

TechCrunch notes that the Sora app's recommendations can draw on your activity, an approximate location from your IP address, and your ChatGPT history, which you can switch off. Review those settings when you set up the app.

How do cameos and consent work?

To create a cameo, you make a short one-time recording of your face and voice inside the app. That recording verifies that you are who you say you are and captures your likeness. From then on, you decide who may put you in their videos, and you can take that permission back at any time.

You also get visibility. Any video that includes your cameo, even a draft someone else has not published, is visible to you. You can delete it, and report it if needed. OpenAI applies extra safety checks to every video that uses a cameo, and you can set preferences for how your cameo behaves; OpenAI's own example is asking that it always wear a fedora. Teenagers get stricter cameo permissions by default.

Good practice goes beyond the settings. Before you use a friend's cameo, ask them about the specific idea, not just in general. Show them the result before you post it, and take it down promptly if they change their mind. For anyone without a cameo, the answer is simple: do not try to recreate them.