I’m guessing these are also specially trained image classifiers and not LLMs, so people’s intuitions about how LLMs work/fail may not apply.