# That Isn't Me Source: https://www.youtube.com/watch?v=-kDtt0QBNRU Recap page: https://rapidrecap.app/video/-kDtt0QBNRU Generated: 2025-11-22T18:33:40.709+00:00 --- ## Quick Overview The video demonstrates that despite advancements in AI video generation, subtle flaws like incorrect shadow convergence, lack of physical context in certain actions, and inconsistencies in generated features like hands or specific objects like corn, make deepfakes from models like Sora and Veo 3 still detectable upon close scrutiny, although the technology is rapidly improving and is already being used for scams. **Key Points:** - The presenter shows that AI-generated content, even from advanced models like Sora and Veo 3, exhibits detectable errors, such as incorrect shadow convergence in perspective drawings (09:06) and inconsistent lighting/shadows. - The demonstration used a model trained on approximately 2000 images of the presenter's face, yet still produced output that required scrutiny to identify as fake (02:28). - The presenter highlights that AI models often fail to correctly render physical laws, such as light casting shadows away from the light source, as shown in an image of a person meditating in an alley (09:04). - The video features a clip generated using Veo 3 that attempts to show the presenter lifting another person, but the context (like the subject's reaction) is poorly synchronized, illustrating issues with physical action coherence (00:14). - Bitdefender sponsored the video, providing statistics that global scam losses now exceed $1 trillion annually, emphasizing the real-world danger of sophisticated AI scams (03:31). - The presenter notes that newer models like Google's Veo 3 and Sora are generally more convincing than older ones, but specific flaws like inconsistent lighting or rendering of hands remain current issues (05:01, 09:52). - The presenter advises viewers to remain skeptical of unsolicited requests for information or money, especially from unfamiliar links, and to verify sources carefully (10:42). ![Screenshot at 00:11: The presenter holds up a bottle labeled "Linus Strength Pills" and sarcastically challenges David Linus to watch all of his 2019 cats videos, setting up the premise of testing AI video generation capabilities.](https://ss.rapidrecap.app/screens/-kDtt0QBNRU/00-00-11.png) **Context:** This video explores the current state of AI video generation technology, specifically testing the capabilities and limitations of models like Sora and Veo 3 in creating realistic footage of the presenter, Linus Sebastian. The video is sponsored by Bitdefender, a cybersecurity company, which uses the context of increasingly realistic AI-generated scams to promote their Scam Protection software. The core of the video involves generating clips based on specific prompts and analyzing the resulting visual artifacts where the AI fails to perfectly replicate real-world physics or context. ## Detailed Analysis The video investigates the realism of modern AI video generation tools, specifically testing models like Sora and Veo 3, using Linus Sebastian as the subject. The presenter initially shows a humorous, sponsored segment where he juggles computer cases while promoting Linus Strength Pills, followed by a clip of a colleague (Nick Plouffe) being carried, which was flagged by AI safety filters due to suggestive actions. The presenter explains that to bypass filters, they rewrote the prompt to frame the action as a helpful rescue, leading to a more successful (though still slightly awkward) AI generation (02:15). The presenter then demonstrates technical flaws in AI generation by showing an image where shadows cast by figures in an alley do not converge correctly to the light source, a failure to adhere to the laws of physics (09:04). He contrasts this with human-understood perspective drawing, which adheres to these laws. Furthermore, he shows examples where the AI generation failed to align audio and video synchronization perfectly (02:41) and where the generated subject's body proportions or features were clearly wrong, such as the distorted Will Smith eating spaghetti clip (09:40). The video is sponsored by Bitdefender, which highlights that global scam losses exceed $1 trillion annually, underscoring the danger of convincing deepfakes. The presenter concludes that while AI is rapidly improving—citing that it took years to reach the quality seen in the current test versus older attempts—subtle cues like lighting inconsistencies, poor physical context, and failure to maintain perspective mean that discerning real from fake is still possible, though increasingly difficult. ### Sponsorship & Context - Bitdefender sponsored the video to discuss AI-driven scams - Global scam losses exceed $1 trillion annually - The presenter tested various AI video models including Sora, Veo 3, and others - The goal was to see how easy it is to spot flaws (00:03, 03:25, 04:45) ### Prompt Engineering & Safety Filters - Initial prompt for carrying a colleague was flagged as non-consensual - Rewriting the prompt to specify a 'fireman's carry' or 'support' helped bypass filters (02:15, 05:38) ### Detecting Flaws - AI struggles with physics like accurate shadow convergence, as shown in the alley scene (09:04) - AI struggles with complex actions like carrying someone or maintaining consistent 3D space (06:36, 08:45) ### Model Comparison - Newer models like Veo 3 and Sora are more convincing than older ones, but still have flaws like inconsistent lighting and poor element rendering (05:01, 09:57) ### Call to Action & Conclusion - Viewers should be skeptical of suspicious links or requests for money/information, and verify sources, especially those using manufactured calls or deepfakes (10:42, 11:15) ![Screenshot at 00:04: The presenter juggling computer cases while introducing the sponsor, Bitdefender \(00:04\).](https://ss.rapidrecap.app/screens/-kDtt0QBNRU/00-00-04.png) ![Screenshot at 02:25: A screenshot from DeepFaceLab showing the process of training a model on thousands of images of the presenter's face \(02:25\).](https://ss.rapidrecap.app/screens/-kDtt0QBNRU/00-02-25.png) ![Screenshot at 09:04: An AI-generated image of a person meditating in an alley where the shadows cast by the light source do not converge correctly, illustrating a failure to obey physical laws \(09:04\).](https://ss.rapidrecap.app/screens/-kDtt0QBNRU/00-09-04.png) ![Screenshot at 09:40: A highly distorted deepfake of Will Smith eating spaghetti, demonstrating artifacts from an older, less constrained AI model \(09:40\).](https://ss.rapidrecap.app/screens/-kDtt0QBNRU/00-09-40.png) ![Screenshot at 10:56: The Bitdefender Scam Protection dashboard showing activity statistics, including 3,079 links verified and 87 links blocked \(10:56\).](https://ss.rapidrecap.app/screens/-kDtt0QBNRU/00-10-56.png)