Do variations in prompt change the ultimate decision of the LLM? Using a series of prompt variations across a variety of text classification tasks, we show how even the smallest of perturbations, such as adding a space at the end of a prompt, can change an LLM's answer.
You're searching for your dream job, and AI promises the perfect match. But what if, behind the scenes, this matchmaking AI harbors prejudices? Our study uncovers how the subtle inclusion of demographic features in prompts can drastically alter the jobs recommended by ChatGPT.