Photo by cottonbro studio on Pexels

Google’s Veo 3, its AI video generator released in May, is struggling with an unexpected problem: persistent, unwanted subtitles. Despite offering advanced features like sound and dialogue creation, users are reporting that Veo 3 frequently generates nonsensical or garbled subtitles, even when prompts explicitly request their exclusion.

Frustrated users are forced to expend resources by regenerating clips, utilizing external subtitle removal software, or cropping their videos to eliminate the unwanted text. Google has acknowledged the issue and claims to be developing a solution. However, reports of persistent subtitle errors underscore the challenges inherent in debugging large language models.

Experts suggest the problem may originate from the model’s training data. It’s likely that the dataset includes numerous videos with hardcoded subtitles from platforms like YouTube and TikTok, which are baked into the video frames. Consequently, the AI may misinterpret these as desired elements and generate them regardless of user prompts. Rectifying this would necessitate a labor-intensive cleansing of the training data. Some industry observers speculate that Google may have prioritized the rollout of its lip-synced audio tool, potentially at the expense of addressing the existing subtitle issue, hinting at a potential rush to market.