The paper confirms that even the biggest and most advanced models can still be misled by adding irrelevant information that can be easily filtered out by humans.