"For months, YouTubers have been quietly griping that something looked off in their recent video uploads. Following a deeper analysis by a popular music channel, Google has now confirmed that it has been testing a feature that uses AI to artificially enhance videos. The company claims this is part of its effort to "provide the best video quality," but it's odd that it began doing so without notifying creators or offering any way to opt out of the experiment."
"Google's test raised eyebrows almost immediately after it began rolling out in YouTube Shorts earlier this year. Users reported strange artifacts, edge distortion, and distracting smoothness that gives the appearance of AI alteration. If you've ever zoomed in close after taking a photo with your smartphone only to notice things look oversharpened or like an oil painting, that's the effect of Google's video processing test."
"According to Rene Ritchie, YouTube's head of editorial, this isn't quite like the AI features Google has been cramming into every other product. In a post on X (formerly Twitter), Ritchie said the feature is not based on generative AI but instead uses "traditional machine learning" to reduce blur and noise while sharpening the image. Although, this is a distinction without a difference-it's still AI of a sort being used to modify videos."
YouTubers noticed odd visual changes in recent uploads and flagged artifacts, edge distortion, and an over-sharpened, 'oil painting' appearance when zoomed. Google confirmed a test of AI-based video enhancement deployed in YouTube Shorts that reduces blur and noise while sharpening images and claimed the goal was to provide the best video quality. The rollout occurred without notifying creators or offering an opt-out. YouTube editorial head Rene Ritchie described the system as using traditional machine learning rather than generative AI. Creator Rhett Shull investigated and characterized the effect as upscaling, though Google disputes that technical label.
Read at Ars Technica
Unable to calculate read time
Collection
[
|
...
]