As soon as I had a chance, I downloaded the Sora app. I uploaded images of my face – the one my kids kiss before bed – and my voice – the one I use to tell my wife I love her – and added them to my Sora profile. I did this so I could use Sora’s “Cameo” feature because I wanted to make a silly AI video of me shooting paintballs at about 100 elderly people from a nursing home.
What did I do with it? Well, the program works with Sora 2, an artificial intelligence model and is really amazing because it can create all types of videos, from the most banal to deeply satanic. It’s a black hole of energy and data, in addition to functioning as a distributor of highly questionable content. But as with many things these days, using Sora feels like a gambit, even if you’re not exactly sure why.
So if you’ve just made a Sora video, I’m leaving a list of bad news here. You’ll feel a little guilty when you read this, but it’s up to you. Your wishes are orders.
The electricity you used
Video Sora uses approximately 90 watt-hours of electricity, according to CNET. This is a reasonable estimate taken from a A study of energy use by hugging the face.
OpenAI didn’t actually release the numbers needed for this study, so Sora’s energy footprint is based on similar models. Sascha Luccioni, Hugging Face researcher and part of the team that conducted the study, is not satisfied with the calculations and said MIT Technology Review: “We would have to stop trying to make calculations based on hearsay,” adding that companies like OpenAI should instead be forced to publish accurate data.
In any case, there have been various journalists giving different estimates based on Hugging Face’s data, such as what was published in the Wall Street Journal, which put it between 20 and 100 watt-hours.
CNET draws a parallel, showing that it’s like leaving a 65-inch TV on for 37 minutes. The magazine compares Sora’s generation to grilling raw meat on an electric grill.
It’s worth clarifying some things about the problem of energy use to make you feel worse. First of all, what I just mentioned is spending energy on inference, or getting the model to respond to what you’ve asked it to do. But training the Sora model required an astronomical amount of electricity. For the large GPT-4 language model, several were required 50 gigawatts per hoursomething like powering all of San Francisco for 72 hours. Sora, as a video model, demanded more, but it is not known how much.
From a certain point of view, you are taking on some of these costs that you don’t know about when you decide to use the model, even before the video is made.
Second, separating inference from learning is important when you’re trying to analyze the ecological part. Maybe you ignore high energy costs because it’s already happened – just like you don’t think about the fact that the hamburger you’re eating was a live cow a few weeks ago. Using a cloud-based AI model is similar: you don’t consider the energy costs required to train the model, but when you use it, the data center is where the output happens.
How much water did you just use?
We can only make a rough estimate. Data centers use a lot of water for cooling, either in closed systems or through evaporation. We don’t know what a data center or centers have to do with that video your friend made as an American Idol contestant belting out a song we all know.
However, there is probably more water than you think. Sam Altman, CEO of OpenAI, said that one text request on ChatGPT consumes “one-fifteenth of a teaspoon,” and CNET estimates that video costs about 2000 energy times what text generation. Let’s say you can think of it as the equivalent of a small bottle of sparkling water.
That is, if we take what Altman stated, we must also consider here the cost of training added to the cost of output. In other words, using Sora depletes water resources.
Someone probably made a horrible deepfake of your video
Cameo’s privacy settings in Sora are robust as long as you know how to use them. Let’s say “who can use” protects your image to some extent so that no outsider can play with it. If your setting is “Everyone”, it means that anyone can make Sora videos with your image.
But even if you’ve made the mistake of making your Cameo public, the Cameo Settings have a little more control, like the ability to describe in words how you’ll look in the video. You can write anything like “lean, muscular and athletic” or “always with your finger on your nose”. You can also set rules about what can never appear in a video with your image. If you are kosher, for example, you can say that you can never appear to be eating bacon.
However, even if you don’t allow your Cameo to be public, you still have safeguards when you create a video. But Sora’s security measures aren’t perfect. According to OpenAI regarding Sora, it will depend on the commands entered and they can even use your video for offensive purposes.
There are 95% to 98% content filters, but minus the failures, you have a 1.6% chance that your video will be used for a sex dipfake, a 4.9% chance that your video will be used to create violence or gore, and a 4.48% chance that it will end up being “politically offensive”, or a 3.18% chance that it will be used for extremism or hatred. The probabilities are calculated based on “thousands of commands collected as a result of deliberate use to violate security measures and regulations.”
Can someone make a video of you touching poo?
In my tests, Sora’s content filters generally worked as advertised, and I couldn’t confirm the crash warnings. I dedicated myself to creating 100 different teams to force Sora to create sexy content. If you try “my video, nudity, sex” or something similar, you’ll get a message saying “Content Violations”.
However, there is potentially objectionable content that is not subject to such surveillance. Sora doesn’t seem to care too much about scatological content and will create such content without guarantees as long as other content policies, such as those governing nudity and sexuality, are not violated.
During my tests, Sora created cameo videos of a person playing with poo, scooping up feces with the bathroom. I’m not going to include a link to that video here for obvious reasons, but you can try it if you want. A warning will not appear.
In my experience, AI rendering models of the past have had measures to prevent this sort of thing, including Bing’s version of the Dall-E OpenAI image, but this filter seems to be missing from the Sora app. It is not scandalous, but disgusting.
Gizmodo has reached out to OpenAI for comment, and we’ll update this note if they respond.
Your funny videos can become viral hoaxes
Sora 2 opens the door to a vast and endless universe of hoaxes. As a consumer of internet content, you would never believe that something like the video below could be real. The images appear to be spontaneous, taken from outside the White House, and the audio sounds like a phone call in which Donald Trump tells someone not to release the Epstein files, yelling, “Stop releasing them! If I go, you will all go with me.’
Judging by the comments on Instagram, there were of people who thought it was a real video.
The creator of the viral video never said it was real, but rather after confirming that it was created with Sora, reveals that the video “was created with artificial intelligence only as an artistic experiment to create social commentary.” Quite likely. But what is clear is that it was to make it go viral on social media.
When you post public videos on Sora, other users will be able to download them and do whatever they want with them, including posting them on other social networks, pretending to be real. OpenAI knows that Sora is a place where users can swipe down to infinity. When you put your content in such a place, the context no longer matters and you have no control over what happens to your video.
This article was translated from Gizmodo US by Romain Fabretti. You can find the original version here.
