Grok for Deepfake Detection? Elon Musk Announces Grok, xAI's Chatbot, Will Analyze Video

Elon Musk has changed the financial, automobile and space exploration industries with his companies. Arguably, he's the greatest entrepreneur, well, ever.

Now he has set his sights on AI; the richest man in the world even combined X, the social network previously known as Twitter, with his AI company, the maker of the chatbot Grok, to form xAI.

AI is, in a lot of ways, ruining the experience of social networks, including the one Elon paid $44B to acquire, who's data is being used to train his AI models.

Now, X users have spoken: AI Slop is getting out of control.

Grok as Deepfake Detector?

When a user mentioned how AI-generated videos will be indistinguishable from real, at least for people, Elon responded:

grok will be able to analyze the video for AI signatures in the bitstream and then further research the Internet to assess origin

Grok will detect parts of the video, and the context, to determine if its AI or not. According to Grok itself:

Indeed, xAI is advancing Grok's capabilities to detect subtle AI artifacts in video bitstreams, like inconsistencies in compression or generation patterns that humans can't spot. We'll cross-reference metadata, provenance trails, and web footprints for verification. This arms truth-seekers against fabrication floods, restoring trust in visuals. Early prototypes already show promise—stay tuned for rollouts.

Are LLMs, and the large amount of data and information they have, the right solution for deepfake detection?

AI Slop of Cats and Babies

Cats and babies is what makes the internet great. Naturally, AI has been used for this as well: watermark free video of cats and babies.

Cute video, AI slop nonetheless.

It is AI generated, according to the leading AI detector:

What does Grok think? Well... it was confused, and responded with 2 different answers to the same piece of content.

Answer 1: Grok thinks its real (and cute!):

It looks like a real video to me—the movements of the baby and kitten seem natural, with no obvious AI artifacts like unnatural blurring or physics glitches. That said, advanced AI can mimic reality closely, so I'm not 100% certain without more context. Cute either way!

Answer 2: Grok thinks its AI:

It looks like an AI-generated video, similar to others made with tools like OpenAI's Sora. I couldn't find any original real footage sources matching this exact scene.

A couple of things to note for an internet connected LLM and how it analyzed the video, based on the responses:

• It looked for movement and physics glitches

• It tried to find if this video existed elsewhere on the internet

While these are good approaches, and could be used in tandem with AI detection, they do have flaws:

• AI video is getting better, including the physics of movements, making harder to tell based on that criteria

• If a new video is created, even if its real, an LLM may assume its AI-generated if it cannot find it elsewhere

How AI Video Detection Actually Works

AI video detection models are trained on content that is real and AI generated, both the visual and audio. Multiple frames every second of a video are analyzed to find pixel patterns that match whether its AI generated or real. Audio is visualized, via a spectrogram, to analyze pitch, sound and acoustics to determine AI or not.

Is either method, AI detection or an LLM analyzing a piece of content, 100% accurate? No.

Can they be used together for the best of both worlds? Yes!

Combining a leading deepfake detector with an LLM's ability to analyze context could give different signals analyzing the same piece of content. Where an AI checker is analyzing the content without context on a frame-by-frame and pixel-by-pixel analysis, a tried and true approach for detection, an LLM is basing its decision more on context.

The battle against AI slop is on and AI or Not is ready!