Adding 'Do not guess' reduces made-up fields from 71% to 20% in AI outputs

The introduction of the instruction 'Do not guess' led to a significant reduction in made-up fields in outputs from various AI models, dropping from 70.7% to 20.2%. This change demonstrates how a clear directive can lead to more accurate data extraction, which is essential for applications relying on precise information.

The results showed that all AI models performed worse when not given the instruction, with some models generating a high number of fabricated fields. This suggests that without appropriate guidance, AI models tend to fill gaps with inaccurate data, which can mislead users and undermine the utility of the AI system.

The cost of running these models varies, with some providing more accurate outputs at lower costs. For instance, GPT-6 Luna was identified as a strong performer at a low operational cost, indicating that efficiency in both accuracy and expense can be achieved.

However, the test was conducted under controlled conditions, meaning results may vary with real-world data. The impact of this finding is limited to the specific task of web extraction and may not generalize to other types of AI applications without similar testing.

Overall, this event underscores the importance of setting clear instructions for AI systems, which can lead to a marked improvement in output quality and user trust in AI-generated data.