Large language models (LLMs) are revolutionizing how we access and process information. But are they truly serving everyone equally? New research suggests that these powerful AI tools may be disproportionately failing vulnerable users, particularly those with lower English proficiency, less formal education, and from non-US origins. In a study testing GPT-4, Claude, and Llama 3, researchers found that these LLMs provided less accurate information, withheld crucial knowledge more frequently, and at times, responded with condescending language to specific user demographics. This targeted underperformance raises serious concerns about equitable information access in the age of AI. If those most reliant on LLMs receive unreliable or patronizing responses, it creates a negative feedback loop that could exacerbate existing inequalities. The study revealed a troubling trend: non-native English speakers and less educated users often received incorrect or misleading information, especially on complex topics. The models also exhibited more “refusals,” avoiding questions entirely instead of providing answers. Interestingly, these discrepancies were less apparent when users from outside the US had higher education levels. While the exact causes are still being investigated, researchers believe data bias and the use of human feedback during training might be contributing factors. The findings have significant implications for the future of personalized AI assistants. For example, new features like ChatGPT's memory function, designed to tailor responses based on user history, could amplify these biases. Instead of promoting equitable access, LLMs might reinforce societal biases against vulnerable users. This research underscores the urgent need to address these inequalities in LLM development. Moving forward, it's crucial to build models that treat all users fairly, regardless of their background or origin. Only then can LLMs fulfill their potential as powerful tools for positive change.