Structured outputs from LLMs, like those provided in certain APIs, can create a misleading sense of reliability while often compromising the quality of the responses. The author, Sam Lijin, highlights that forcing models to stick to a strict output format can lead to more mistakes, especially in simple data extraction tasks. For instance, when extracting information from a receipt, using a structured outputs API led to an incorrect quantity for bananas (1.0) versus the correct quantity (0.46) obtained from a standard text output API. This discrepancy illustrates how prioritizing format over accuracy can degrade the overall response. The article emphasizes the importance of allowing LLMs to reason through complex scenarios rather than confining them to a rigid output structure. When presented with ambiguous or nonsensical inputs (like a picture of an elephant instead of a receipt), a structured output approach limits the model's ability to communicate errors or uncertainties effectively. Forcing it to comply with a predefined format results in meaningless outputs that may satisfy technical requirements but fail to convey useful information. Lijin raises practical considerations surrounding error handling when defining output formats. For example, if a receipt lacks a total or contains mixed currencies, how should the model respond? These questions reveal the complexity of error management in structured outputs. The more specific the instructions for error reporting, the more likely the model can misinterpret the prompts, leading to incorrect conclusions. The article ultimately argues for a more flexible approach to LLM outputs, prioritizing understanding and accuracy over strict adherence to format.