Today I tried Sora 2. In a few seconds—literally just “looking straight, looking down, and saying 48 and 72”—I had a convincing digital version of me. It looks like me. It often sounds like me. From a pure technology standpoint, this is stunning. See also OpenAI Sora and “Sora 2 is here”.
Sora 2 Generated videos below starring me, Barrett Nash, CEO of InfiniteUp.
What blew me away about Sora 2
- Speed to “me.” The capture → avatar loop took seconds and produced something that feels uncomfortably close to my real face and voice.
- Directability. With a few prompts, I could place “me” in different scenes and get usable footage for edits and rough cuts.
“I was able to create a digital version of me by just looking straight, looking down and saying 48 and 72… took seconds.”
Browse more posts in our category: AI Ethics, Safety & Policy.
What’s still rough in Sora 2
- Narrative drift. In one prompt of me on a motorcycle, Sora cut to a random shot of a helmet dropping. Cool shot; wrong story.
- Audio placement. Sometimes the sync is great; sometimes the dialogue lands in the wrong moment or the ambience stomps the voice.
- Editor required. Directors will make magic here, but it still needs human timing, splicing, and taste.
“The audio sometimes is great, sometimes makes no sense where it is placed.”
The moment we just crossed with Sora 2
After today, I expect Sora-style avatars to show up everywhere: marketing, training, internal commas, indie film, social content. The “CEO: Sell my product” prompt is now a real button. (I’ve already started experimenting. 😄).
See all posts tagged Sora 2 here: Sora 2.
The part that worries me: deepfakes in the Sora 2 era
This is also the easiest deepfake kit most people will ever touch. Abuse is not hypothetical.
- Scams 2.0. Imagine the classic “Nigerian prince” email—now with video of the prince and his “vault.”
- Ransom hoaxes. Parents could get AI videos of their kids “in danger,” complete with believable voices.
- Low-friction identity theft. If a few seconds can build a strong avatar, then anyone who’s ever been on CCTV is potentially cloneable.
“Anyone who walks in front of a CCTV camera once in their life will be able to be reproduced billions of times.”
I uploaded my own image because I want to understand this from the inside. But here’s the paradox: I’d advise most people not to hand over face/voice data—and acknowledge that abstaining won’t prevent someone from cloning you using public footage.
Good overview: Artificial intelligence, deepfakes, and the uncertain future of truth.
How to use this power without torching trust
For creators & teams
- Label AI clearly. Burn in “AI-generated” slates and on-screen marks. Use content credentials (C2PA) when possible.
- Consent receipts. If you feature a person’s likeness or voice, store an auditable consent record.
- No-go prompts. Block prompts that impersonate private individuals or imply unlawful activity.
- Safety rails in editing. Before publishing, run a “harm pass”: could this be misused, misinterpreted, or weaponized?
- Human in the loop. Treat Sora-style output like B-roll and animatics. A human editor should own story, timing, and truth. Practical primer: Responsible AI (IBM).
For everyone
- Assume video can lie. Verify via a second channel (call back, ask for a safe word, use known numbers).
- Lock down your voice. Be cautious about long, clean voice samples online.
- Harden accounts. Enable MFA and set up bank/phone “no video auth” notes—video alone shouldn’t pass identity checks.
Where Sora 2 is great right now
- Concept reels and pre-viz for directors and product teams
- Internal explainers and training where speed beats polish
- Solo creators who need on-camera presence without lights, sets, and re-takes
Where it’s not ready (yet)
- Long, emotionally coherent scenes without human edit
- Unambiguous, high-stakes communication (health, legal, finance)
- Identity or source-of-truth verification
Sora 2 is a leap. It compresses the distance between an idea and a watchable clip to seconds. That’s thrilling—and destabilizing. If we pair this capability with consent, provenance, and a little humility, it can expand what small teams (and individuals) can make. If we don’t, it will dent trust hard.
I’m choosing to learn it, label it, and build with guardrails. Because this tech isn’t arriving—it’s here.