Researchers at the French AI testing company Giskard have found that when language models like ChatGPT are asked to respond briefly, the likelihood of false or misleading information increases. Instructions such as “answer briefly” or “explain in two sentences” often cause models to produce inaccurate responses.
According to TechCrunch, Giskard’s study analyzed the behavior of major AI models, including OpenAI’s GPT-4o, Mistral Large, and Anthropic’s Claude 3.7 Sonnet. The team observed that when these models were prompted to give short answers—particularly on ambiguous or controversial topics—they were more prone to errors. “Our data shows that even simple changes in instructions significantly affect the tendency of models to hallucinate,” Giskard stated.
These so-called “hallucinations” are a persistent challenge in generative AI. Even the most advanced models, built on probabilistic algorithms, can fabricate information. Giskard’s research also suggests that newer reasoning-based models, such as OpenAI’s o3, may be even more prone to hallucinations than earlier versions. This raises further concerns about the reliability of AI-generated content.
Limited Space Means Limited Accuracy
One contributing factor to this problem is the lack of room for explanation in short responses. When models are instructed to be brief, they often can’t fully address complex queries or correct false assumptions embedded in the prompt. For developers working with code or technical data, this issue becomes even more significant, as brevity can compromise the quality and accuracy of the output.
Another finding of the study is that language models are less likely to challenge controversial statements when those statements are made with confidence. Additionally, AI systems that users rate as easy or pleasant to interact with aren’t necessarily the most accurate, adds NIX Solutions.
Guidance for More Reliable Use
Experts at Giskard recommend that users give AI models clear and well-considered instructions while avoiding strict limitations on answer length. They warn that even seemingly simple cues like “be concise” can subtly degrade the quality of information.
As the use of generative AI expands across industries, understanding these nuances is essential. We’ll keep you updated as more integrations and solutions addressing hallucination risks become available.