Google’s new anything-to-anything AI model is wild
Original reporting by The Verge

Last year, one reporter used generative AI to deepfake her son's stuffed animal on imaginary vacations, an experiment that revealed the surprising accessibility and quality of AI video tools. It was a whimsical foray into understanding where this nascent technology stood.
Now, Google introduces Omni Flash, its latest AI video generation model, promising enhanced consistency and real-world integration. Our reporter put Omni to the test, revisiting her plush deer, Buddy, for new adventures. The results were a baffling mix: some clips showed remarkable improvement over previous models, yet peculiar "AI jump scares"—like Buddy suddenly reorienting mid-skydive or an object morphing repeatedly—underscored persistent flaws. Even text-based edits, while improved, struggled with basic requests.
The personal test
The true revelation came when our reporter deepfaked herself. Starting with real footage, she prompted Omni to place her in scenarios like eating spaghetti or posing by the Eiffel Tower. The outcome was disquietingly real. Her husband, unaware of the AI elements, was convinced by a pasta-eating scene. While subtle "AI tells" remained, many clips were "convincing as hell," demonstrating a level of realism unsettlingly close to genuine footage. The ease with which such believable fakes can be generated pushes us firmly into the uncanny valley, forcing a re-evaluation of what we see and believe online.
While Google’s Omni Flash may not yet be the cinematic genie its marketing implies, my experiments with AI Buddy revealed its current limitations: it still struggles with nuanced artistic direction and maintaining perfect consistency in complex, fictional scenarios. However, the true revelation came with deepfaking myself, demonstrating Omni’s alarming capacity to generate convincing videos of real people in fabricated situations, capable of fooling even close family members. This capability marks a significant and unsettling evolution in generative AI.