Just spit balling, but maybe the program that does the transcription doesn’t just use the image, but instead scans the image, finds the Twitter account shown, and checks the tweet text in the image against the matching actual tweet.

And since it’s accessing the actual tweet, maybe that Walmart text is like a profile tag line or something that’s attached to the user?