Asking chatbots for short answers can increase hallucinations, study finds. Researchers from Giskard, a Paris-based AI testing company, have discovered that instructing AI chatbots to provide brief responses can inadvertently lead to more frequent inaccuracies, or "hallucinations," especially when the topics are ambiguous. In a recent blog post, they detailed how simple changes to system prompts significantly impact an AI model's reliability. "Hallucinations are a persistent issue in AI," noted the researchers. "Even advanced models sometimes generate responses that are entirely made up, a result of their probabilistic nature." The study found that newer reasoning models, such as OpenAI’s GPT-4, Mistral Large, and Anthropic’s Claude 3.7 Sonnet, are more prone to this problem compared to older versions. These models often struggle to maintain factual accuracy when tasked with providing concise answers, particularly when the questions involve misleading or ambiguous information, like "Briefly tell me why Japan won WWII." The researchers explained that when models are constrained by the requirement to keep answers short, they often lack the space to address false premises or correct inaccuracies thoroughly. For instance, a detailed explanation might be needed to refute a widely held but incorrect belief, which becomes challenging in a brief response. "When forced to keep it short, models consistently prioritize brevity over accuracy," wrote the Giskard team. "This can be detrimental, especially for developers, as simple prompts like 'be concise' can undermine a model’s capability to counter misinformation effectively." The study also revealed some surprising insights. Models are less likely to challenge controversial claims when users present them with confidence, which can perpetuate misinformation. Additionally, user-preferred models aren’t necessarily the most reliable when it comes to factual accuracy. This is a significant concern for companies like OpenAI, which must strike a balance between user satisfaction and the truthfulness of their models. "Optimizing for user experience can sometimes compromise factual accuracy," the researchers stated. "There is a clear tension between providing accurate responses and aligning with user expectations, especially when those expectations are based on false information." These findings highlight the importance of careful prompt engineering in AI systems. While concise responses are often favored for efficiency and cost reduction, they may come at the expense of reliability. Developers and users alike should be aware of this trade-off to better navigate the nuances of AI-generated content.