AI confabulation

On Wednesday, a video from OpenAI’s newly launched Sora AI video generator went viral on social media, featuring a gymnast who sprouts extra limbs and briefly loses her head during what appears to be an Olympic-style floor routine.

As it turns out, the nonsensical synthesis errors in the video—what we like to call “jabberwockies”—hint at technical details about how AI video generators work and how they might get better in the future.

But before we dig into the details, let’s take a look at the video.

In the video, we see a view of what looks like a floor gymnastics routine. The subject of the video flips and flails as new legs and arms rapidly and fluidly emerge and morph out of her twirling and transforming body. At one point, about 9 seconds in, she loses her head, and it reattaches to her body spontaneously.

“As cool as the new Sora is, gymnastics is still very much the Turing test for AI video,” wrote venture capitalist Deedy Das when he originally shared the video on X. The video inspired plenty of reaction jokes, such as this reply to a similar post on Bluesky: “hi, gymnastics expert here! this is not funny, gymnasts only do this when they’re in extreme distress.”

We reached out to Das, and he confirmed that he generated the video using Sora. He also provided the prompt, which was very long and split into four parts, generated by Anthropic’s Claude, using complex instructions like “The gymnast initiates from the back right corner, taking position with her right foot pointed behind in B-plus stance.”

“I’ve known for the last 6 months having played with text to video models that they struggle with complex physics movements like gymnastics,” Das told us in a conversation. “I had to try it [in Sora] because the character consistency seemed improved. Overall, it was an improvement because previously… the gymnast would just teleport away or change their outfit mid flip, but overall it still looks downright horrifying. We hoped AI video would learn physics by default, but that hasn’t happened yet!”

So what went wrong?

When examining how the video fails, you must first consider how Sora “knows” how to create anything that resembles a gymnastics routine. During the training phase, when the Sora model was created, OpenAI fed example videos of gymnastics routines (among many other types of videos) into a specialized neural network that associates the progression of images with text-based descriptions of them.

That type of training is a distinct phase that happens once before the model’s release. Later, when the finished model is running and you give a video-synthesis model like Sora a written prompt, it draws upon statistical associations between words and images to produce a predictive output. It’s continuously making next-frame predictions based on the last frame of the video. But Sora has another trick for attempting to preserve coherency over time. “By giving the model foresight of many frames at a time,” reads OpenAI’s Sora System Card, we’ve solved a challenging problem of making sure a subject stays the same even when it goes out of view temporarily.”

Maybe not quite solved yet. In this case, rapidly moving limbs prove a particular challenge when attempting to predict the next frame properly. The result is an incoherent amalgam of gymnastics footage that shows the same gymnast performing running flips and spins, but Sora doesn’t know the correct order in which to assemble them because it’s pulling on statistical averages of wildly different body movements in its relatively limited training data of gymnastics videos, which also likely did not include limb-level precision in its descriptive metadata.

Sora doesn’t know anything about physics or how the human body should work, either. It’s drawing upon statistical associations between pixels in the videos in its training dataset to predict the next frame, with a little bit of look-ahead to keep things more consistent.

This problem is not unique to Sora. All AI video generators can produce wildly nonsensical results when your prompts reach too far past their training data, as we saw earlier this year when testing Runway’s Gen-3. In fact, we ran some gymnast prompts through the latest open source AI video model that may rival Sora in some ways, Hunyuan Video, and it produced similar twirling, morphing results, seen below. And we used a much simpler prompt than Das did with Sora.

AI models based on transformer technology are fundamentally imitative in nature. They’re great at transforming one type of data into another type or morphing one style into another. What they’re not great at (yet) is producing coherent generations that are truly original. So if you happen to provide a prompt that closely matches a training video, you might get a good result. Otherwise, you may get madness.

As we wrote about image-synthesis model Stable Diffusion 3’s body horror generations earlier this year, “Basically, any time a user prompt homes in on a concept that isn’t represented well in the AI model’s training dataset, the image-synthesis model will confabulate its best interpretation of what the user is asking for. And sometimes that can be completely terrifying.”

For the engineers who make these models, success in AI video generation quickly becomes a question of how many examples (and how much training) you need before the model can generalize enough to produce convincing and coherent results. It’s also a question of metadata quality—how accurately the videos are labeled. In this case, OpenAI used an AI vision model to describe its training videos, which helped improve quality, but apparently not enough—yet.

We’re looking at an AI jabberwocky in action

In a way, the type of generation failure in the gymnast video is a form of confabulation (or hallucination, as some call it), but it’s even worse because it’s not coherent. So instead of calling it a confabulation, which is a plausible-sounding fabrication, we’re going to lean on a new term, “jabberwocky,” which Dictionary.com defines as “a playful imitation of language consisting of invented, meaningless words; nonsense; gibberish,” taken from Lewis Carroll’s nonsense poem of the same name. Imitation and nonsense, you say? Check and check.

We’ve covered jabberwockies in AI video before with people mocking Chinese video-synthesis models, a monstrously weird AI beer commercial, and even Will Smith eating spaghetti. They’re a form of misconfabulation where an AI model completely fails to produce a plausible output. This will not be the last time we see them, either.

How could AI video models get better and avoid jabberwockies?

In our coverage of Gen-3 Alpha, we called the threshold where you get a level of useful generalization in an AI model the “illusion of understanding,” where training data and training time reach a critical mass that produces good enough results to generalize across enough novel prompts.

One of the key reasons language models like OpenAI’s GPT-4 impressed users was that they finally reached a size where they had absorbed enough information to give the appearance of genuinely understanding the world. With video synthesis, achieving this same apparent level of “understanding” will require not just massive amounts of well-labeled training data but also the computational power to process it effectively.

AI boosters hope that these current models represent one of the key steps on the way to something like truly general intelligence (often called AGI) in text, or in AI video, what OpenAI and Runway researchers call “world simulators” or “world models” that somehow encode enough physics rules about the world to produce any realistic result.

Judging by the morphing alien shoggoth gymnast, that may still be a ways off. Still, it’s early days in AI video generation, and judging by how quickly AI image-synthesis models like Midjourney progressed from crude abstract shapes into coherent imagery, it’s likely video synthesis will have a similar trajectory over time. Until then, enjoy the AI-generated jabberwocky madness.