Paper Tests ChatGPT and Rival Models on Misleading Visual Questions

Researchers describe a method for probing how vision language models handle questions that are not perfectly worded, the kind a real user actually types rather than a clean benchmark prompt. The team used the models themselves to generate question modifiers that insert information a questioner would not normally add: details that are beside the point, details that are vague enough to allow more than one reading, and details that are flatly false. The idea borrows from the philosopher Paul Grice’s maxims of cooperative conversation, which describe the unstated rules people follow when they ask and answer questions honestly and efficiently. A question that breaks one of those rules is harder to answer well even for a person, and the paper asks whether current models struggle in the same way or in a different way entirely.

The four models evaluated, ChatGPT, Claude, Gemini and Llava, all showed weaker performance once a question carried one of these violations, according to the paper. The authors also compared how the models handled violations a human had written into the question against violations the models themselves had generated, and found the two were not resolved the same way. A further experiment measured how long human participants took to work through the same confusing questions, using time on task as a stand in for cognitive effort, and found that people spent less time resolving model-generated violations than human-induced ones even though the models answered those cases less accurately.

That gap between human effort and model accuracy is the paper’s sharper finding. It suggests the kind of confusing question a model itself tends to produce is, in a narrow sense, easier for a person to brush past than it is for a model to answer correctly, which runs against the simple assumption that whatever is easy for a person to parse should also be easy for a model to parse.

The numbers

The paper does not publish a scoreboard of exact accuracy figures in its abstract, so this item reports the comparison in the terms the authors use: four named models tested, one human baseline measured by time on task, and two distinct violation sources, human written and model generated, compared against each other.

For anyone building a product on top of visual question answering, the practical read is narrower than the paper’s full scope: a prompt that looks innocuous to a reviewer can still carry an ambiguity or a false premise that the model will answer confidently and wrongly, and no amount of general benchmark performance tells you that in advance.