First impressions: Sora 2 is full of deepfakes and SpongeBob

This week, OpenAI finally released Sora 2, the highly anticipated follow-up to its generative AI video model. OpenAI launched Sora 2 in a standalone iOS app (sorry, Android users), which is currently available for free on an invite-only basis. Because we didn’t have enough slop machines.

I managed to get a Sora 2 invite code, and I’ve been scrolling through the app and making videos ever since (and getting paid to do it).

My first impressions are a bit complicated. The technology is impressive, certainly. And I had some fun scrolling through the app, but just as often, I found content that left me feeling uneasy.

What is Sora 2?

Sora 2 is a new video generation model and app from OpenAI, the maker of ChatGPT. Sora 2 can make videos (with corresponding dialogue and audio) based on natural language prompts. It’s the first true rival to Google’s Veo 3 AI video maker, which has been in a league of its own since its launch earlier this year. I don’t even want to mention Meta’s lackluster Llama or Grok Imagine video tools in the same sentence as these apps, though Meta should get a boost now that it’s licensing Midjourney technology.

The invite-only screen…
Credit: Screenshot courtesy of Sora

…and what awaits you on the other side.
Credit: Screenshot courtesy of Sora

Using Veo 3 for the first time was one of those crossing-the-rubicon moments for me. The level of realism was equally impressive and horrifying. Sora 2 feels the same way. Like Veo 3, it’s mostly being used to make viral meme content and short-form videos like you’d see on TikTok. Videos of a Golden Retriever being arrested for shoplifting steaks at the grocery store, or an emotional support kangaroo being stopped at the airport, seem cute, not sinister.

But Sora 2’s potential for harm has a 1:1 relationship with its quality. The better and more realistic the videos are (and some of them are very good, and often realistic), the more I worry about deepfakes and misinformation.

Sora 2 has much better safeguarding than Grok

Sora will reject prompts for overtly sexual videos.
Credit: Screenshot courtesy of Sora

Sora makes it easy to report offensive content.
Credit: Screenshot courtesy of Sora

When xAI and Elon Musk launched Grok Imagine, a generative AI image and video generator, I was, frankly, horrified by the lack of safeguarding. Musk has pitched xAI and Grok as the politically incorrect alternative to artificial intelligence apps that are, he says, bogged down with liberal bias. Grok also has a much more laissez-faire approach to content moderation and safety, resulting in sexual deepfakes on Grok Imagine.

On the other hand, OpenAI has implemented much saner safeguards for Sora 2. If you upload an image to serve as the inspiration for a video, the app will reject your image if it detects a face — any face.

A screenshot of the Sora app showing a rejected photo, partially censored by Mashable.
Credit: Screenshot courtesy of Sora

If you want to create a video featuring a real person, you have to use the Cameos feature. This feature lets you create videos with the likeness of specific people — as long as they’ve agreed to participate in the feature.

When I tried to create videos of public figures like, say, Taylor Swift, using common jailbreaking techniques, the app refused to make the video. Needless to say, this is not the blonde showgirl I had in mind.

Cameos tool is cool and unnerving

Cameos are the most famous feature of the new Sora app so far. When you get access to the app, the first thing you have to do is opt in or out of the Cameos tool, which allows your likeness to be used in videos on the app. You can give yourself the ability to make videos of yourself, but also grant access to contacts, specific users, or the public.

The Cameos tool in Sora.
Credit: Screenshot courtesy of Sora

Cameos has resulted in a flood of videos featuring the likeness of OpenAI CEO Sam Altman. Cameos is a clever way to circumvent the deepfake problem by letting users essentially opt in to deepfakes.

This Tweet is currently unavailable. It might be loading or has been removed.

I created a video of myself, and it was weird. Sora didn’t get my voice right, but my face, body, hair, and general likeness were spot on. Seeing yourself say and do things you’ve never said or done is a very strange feeling.

Mashable Light Speed

Unfortunately, you may need to get used to that feeling — as AI enthusiasts are fond of saying, it’s a new era.

Sora 2 and Veo 3 are in a league of two

A lot of generative AI tools can animate photographs, but few can make realistic videos with corresponding dialogue and sound effects. Sora 2 does this easily, like Veo 3 before it. We’ll have a more in-depth comparison of Sora 2 and Veo 3 coming soon. In the meantime, I’ll just say Sora 2 lives up to the hype in ways that GPT-5 did not.

The Sora app can make videos in a variety of styles — fake police body-cam videos, ’90s TV commercials, music videos, sports broadcasts — that don’t immediately look like AI videos at all.

Sora 2 seems to play fast and loose with intellectual property

Mashable has written extensively about the fight between artists and IP holders and the AI industry. AI companies like Meta have won some early victories in these fights, and President Donald Trump’s executive orders and comments on the topic have favored the AI industry.

“You can’t be expected to have a successful AI program when every single article, book, or anything else that you’ve read or studied, you’re supposed to pay for,” Trump said when announcing The White House’s AI Action Plan this summer, per Politico. “We appreciate that, but just can’t do it — because it’s not doable.”

Pokemon in the style of ‘The Breath of the Wild’.
Credit: Screenshot courtesy of Sora

SpongeBob and Patrick in ‘Midsommar’ by A24.
Credit: Screenshot courtesy of Sora

Obviously, many artists and rights holders strongly disagree. Disney lawyers famously called Midjourney a “bottomless pit of plagiarism” in its lawsuit against the AI company. But, for now, the federal government seems to be clearing the way for companies like OpenAI to use IP with impunity, lest China gain the edge in the AI arms race.

So, unless Disney or Warner Bros. suddenly scores a surprise legal victory, I’d expect the SpongeBob Squarepants, Star Wars, and Rick & Morty AI videos to keep spreading like meme-fueled wildfire. Mashable asked OpenAI if the company had a licensing agreement with Warner Bros., which owns the rights to Rick & Morty, but the company declined to answer.

Deepfakes and memes as far as the eye can scroll

There are some recurring themes in the Sora app. Users seem to be turning Martin Luther King Jr.’s “I have a dream” and JFK’s “Ask not what your country” speeches into video memes. A representative example: “I have a dream that Xbox Game Pass will not raise prices.” Did I laugh the first time I heard Martin Luther King Jr. say, “I have a dream that I’m never going to give you up, never gonna let you down, never gonna run around and desert you”? I may have chuckled. But I wish I hadn’t.

I also saw a lot of SpongeBob memes in particular, and to a lesser extent, various riffs on Rick & Morty and Pokémon. That could be because a lot of millennials are using Sora, and millennials hold SpongeBob Squarepants and Pokémon in a special place in their hearts.

Screenshot of Sora app with user details blurred out.
Credit: Screenshot courtesy of Sora

Screenshot of Sora app with user details blurred out.
Credit: Screenshot courtesy of Sora

Public figures and beloved TV characters have always been fodder for memes. What’s novel here is the ability to easily create videos featuring these figures saying whatever you want. In my experience, Google Veo 3 is more sensitive to prompts involving IP.

How could this be problematic? With Elon Musk currently leading a Netflix boycott over transgender characters in children’s TV shows, now anyone can make a realistic video featuring characters saying whatever they want. It could be a powerful tool for outrage farmers.

So, it’s once again time to practice media literacy and hone your ability to identify viral AI videos.

I reached out to OpenAI for comment, and a company representative said that Sora was built to provide users with as much creative freedom as it could. The rep also said that IP holders can submit takedown requests through the company’s Copyright Disputes form; however, there is not a blanket opt out for IP holders.

“We’re under heavy load, please try again later.”

For early users, expect to see this message a lot. As when OpenAI first integrated image generation into ChatGPT, the company is facing extremely high demand for Sora. I don’t expect that to change any time soon. I repeatedly received this error message while testing the app. In fact, it was hard to use up all my credits due to this frequent error.

Sora app with error message: ‘We’re under heavy load, please try again later’
Credit: Screenshot courtesy of Sora

My favorite feature so far

In the Sora feed, you can scroll up or down to find new videos, which is typical. However, on some videos, you can also scroll sideways to see alternative versions of the users’ posts. This video album feature lets you see how the video turned out with slight tweaks to the prompt, which is pretty darn cool.

Body cam-style AI video of dogs shoplifting.
Credit: Screenshot courtesy of Sora

“Make it a pitbull cops treat him a lot worse because of his breed”
Credit: Screenshot courtesy of Sora

It’s a little addicting

Finally, I have to admit that Sora 2 is a little addicting, but only in the same way that all short-form video apps are addicting. Just like it’s easy to fall into a TikTok hole, it’s easy to fall into a Sora hole, and I suspect many early adopters are wasting a lot of time on the app.

As I said, the tech is certainly impressive. But, better slop is still slop, no matter how many likes it gets.

Disclosure: Ziff Davis, Mashable’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.