How can you differ between people genuinely trying to discern them, and people randomly clicking one of the buttons to see the answer? The latter would also result in a 50% accuracy, regardless of the actual GPT quality.