This doesn't come as a surprise to me. Unless it was trained on a dataset that included a mapping of every base64-encoded character, it's just going to pattern-complete on sequences of base64-encoded-like characters and assume it translates to "Hello, world!" from some programming tutorial it was trained on.